diff --git "a/job.json" "b/job.json" new file mode 100644--- /dev/null +++ "b/job.json" @@ -0,0 +1,249802 @@ +{ + "in_dir": "C:\\Users\\user\\Downloads\\oobabooga_windows\\oobabooga_windows\\migtissera_Synthia-70B-v1.2", + "out_dir": "C:\\Users\\user\\Downloads\\oobabooga_windows\\oobabooga_windows\\migtissera_Synthia-70B-v1.2-2.55bpw", + "cal_dataset": "C:\\Users\\user\\Downloads\\oobabooga_windows\\oobabooga_windows\\0000.parquet", + "dataset_rows": 100, + "measurement_rows": 16, + "gpu_rows": 0, + "length": 2048, + "measurement_length": 2048, + "bits": 2.55, + "head_bits": 6, + "progress": "finished", + "shard_size": 8192, + "measurement": [ + { + "key": "model.layers.0.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.010615579783916473, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.010184699669480324, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.006386835128068924, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.006291421130299568, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.006199309602379799, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.004915353376418352, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.018612589687108994, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.009976892732083797, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.006272817961871624, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0061276438646018505, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.006276554893702269, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.006391346454620361, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.006099625024944544, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.005111755337566137, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0048576779663562775, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.005074269138276577, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.00483368057757616, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0047367531806230545, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.00482926657423377, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.004730968736112118, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.00472002848982811, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0048276823945343494, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004590169060975313, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004727803636342287, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.010615579783916473, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.01012005377560854, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.00950608029961586, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.0048700668849051, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.0047568827867507935, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.00465010479092598, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0021435790695250034, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.011269958689808846, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.009283497929573059, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.004798988811671734, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0045536234974861145, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0047899573110044, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.00497908191755414, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0045234872959554195, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0027062504086643457, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.0020367864053696394, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0026993846986442804, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.001981612527742982, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0017296219011768699, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0019712522625923157, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0017127281753346324, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0017585154855623841, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0019684534054249525, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0013546659611165524, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.001704330905340612, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.01012005377560854, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05398493632674217, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.04393395408987999, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.03633197769522667, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.03043084777891636, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.02334652468562126, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.017393771559000015, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03546474501490593, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03173137456178665, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.025118626654148102, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.01864815503358841, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.018230954185128212, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.018140261992812157, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01527587790042162, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01149455364793539, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01039813831448555, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009081721305847168, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.006349299103021622, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00587282981723547, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.005513349547982216, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.004860377870500088, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0047355773858726025, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004963907413184643, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.003320631803944707, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00357939419336617, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05398493632674217, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.030513916164636612, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.019822286441922188, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.014413585886359215, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.014391311444342136, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.012790540233254433, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.007444096729159355, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.01838909275829792, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.016516825184226036, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.01421528123319149, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.009065484628081322, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.009241492487490177, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.009398628026247025, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.00833461619913578, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.006809044163674116, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.006403118371963501, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.0049991062842309475, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0045204670168459415, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0044081322848796844, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.004038755316287279, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.003845011116936803, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.003214186755940318, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.003959836903959513, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0028279570396989584, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0035607817117124796, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.030513916164636612, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.027206771075725555, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.025291521102190018, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.024699442088603973, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.022197213023900986, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.012105917558073997, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.011461183428764343, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.013457946479320526, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.012517443858087063, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.01227525807917118, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.010822596959769726, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.010265548713505268, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0068728867918252945, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.006129442248493433, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.005951432976871729, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0059103332459926605, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0035192754585295916, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.0034056713338941336, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.0033932074438780546, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.0031883649062365294, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.0031770789064466953, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.002075646072626114, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.0024920955765992403, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.002025477821007371, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.0021074486430734396, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.027206771075725555, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.033175989985466, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.03084232099354267, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.030121462419629097, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.027071019634604454, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.014728720299899578, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.013933074660599232, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.016355356201529503, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.015220359899103642, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.014940721914172173, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.013128113001585007, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.012430805712938309, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.008254868909716606, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.007239314261823893, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.007020839489996433, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.006969672627747059, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.004138784483075142, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.0036904369480907917, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.0036743958480656147, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.0033842818811535835, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.00337000610306859, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.002218578476458788, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.0023447824642062187, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.0021489954087883234, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.0016892512794584036, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.033175989985466, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.02328990027308464, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.01937221549451351, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.01607653684914112, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.014150701463222504, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.010209633968770504, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.007834449410438538, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.015630090609192848, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.014041111804544926, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.010954812169075012, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.008694956079125404, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.008506504818797112, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.007970944046974182, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.007008582353591919, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.005372187122702599, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.004917123354971409, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.004146990366280079, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.0035207709297537804, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.0033534341491758823, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.0033136247657239437, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.003099318826571107, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.0025327764451503754, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.003107930300757289, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.0020848067943006754, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.0027289995923638344, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.02328990027308464, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.013434314168989658, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.01018445286899805, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.00583964167162776, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.005683921743184328, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.005361232906579971, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0025280537083745003, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.01088688150048256, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.009509683586657047, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.006113956216722727, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.004629918374121189, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.004958363715559244, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.005306848790496588, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.004521362483501434, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.0027844540309160948, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0021945959888398647, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.0027320955414325, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.001758083002641797, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0015293050091713667, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0016536248149350286, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0013625833671540022, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0016143475659191608, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.001639875816181302, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0011418116046115756, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.001298745279200375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.013434314168989658, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.012716341763734818, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.009396803565323353, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.005533321760594845, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.0052995840087533, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.004912695847451687, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.002361543709412217, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.009792476892471313, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.008618281222879887, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.005659270100295544, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.004148790612816811, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.004416611511260271, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0046881819143891335, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.004017972387373447, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.002496453933417797, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.0019717183895409107, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0023981682024896145, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0014939657412469387, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0012907496420666575, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.001378194778226316, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0011015398195013404, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.001406485796906054, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.001360093941912055, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0010045156814157963, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0010230401530861855, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.012716341763734818, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0662318617105484, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.05171496421098709, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.04462010785937309, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.03840954974293709, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.028525356203317642, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02160250023007393, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03890480101108551, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03510640934109688, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.030803726986050606, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.021756762638688087, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.021340111270546913, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.020136451348662376, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01691175065934658, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.013820203021168709, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01299845241010189, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.010022996924817562, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007255554664880037, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.006933275610208511, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.005935050547122955, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.005389168858528137, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005206095986068249, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005081487353891134, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004034972749650478, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0033418687526136637, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0662318617105484, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07327425479888916, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06016135588288307, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.054901979863643646, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.04056771844625473, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03325483947992325, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.02789173275232315, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.041255272924900055, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03712841123342514, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.034376345574855804, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.023590954020619392, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.021556712687015533, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.021251216530799866, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.018453029915690422, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.016781846061348915, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01636620983481407, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.011197981424629688, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.010069763287901878, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009925153106451035, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008586226962506771, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008412010967731476, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.007025238126516342, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.007932017557322979, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.006578127387911081, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006876960396766663, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07327425479888916, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.05328350141644478, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.050505489110946655, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.04970201104879379, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.04545290768146515, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.023658286780118942, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.022810213267803192, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.026016300544142723, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.02412659488618374, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.02388232946395874, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.021710839122533798, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.020886143669486046, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.013215752318501472, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.011595407500863075, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.011392807587981224, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.011347186751663685, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.006656828802078962, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.0062003182247281075, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.006187142338603735, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.00584819633513689, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.005842890590429306, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.00373643240891397, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.004193508997559547, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.003674958599731326, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.003319249488413334, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.05328350141644478, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.063599593937397, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.06039176508784294, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.059466272592544556, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.054536957293748856, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.02840491198003292, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.027382969856262207, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.031186357140541077, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.028944717720150948, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.028670739382505417, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.026088835671544075, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.025065520778298378, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.015715092420578003, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.013698550872504711, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.013463165611028671, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.01340731605887413, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.007839928381145, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.006910604890435934, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.006893789395689964, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.006451527122408152, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.006444851867854595, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.0041144415736198425, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.004139314405620098, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.004037539009004831, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.0027452351059764624, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.063599593937397, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.049515798687934875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.04476568475365639, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.03969215974211693, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.0359765850007534, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.023078734055161476, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.019588856026530266, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.03264560177922249, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.029705988243222237, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.023905567824840546, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.020654011517763138, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.02006528712809086, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.016700904816389084, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.014843424782156944, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.012017813511192799, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.011240809224545956, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.00877555925399065, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.0076224710792303085, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.007307341322302818, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.007261454127728939, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.006892282981425524, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.0053890603594481945, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.006495383568108082, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.004609460476785898, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.005668321158736944, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.049515798687934875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.014673582278192043, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.012657558545470238, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.009874317795038223, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.008589322678744793, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.006369961425662041, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.004551870282739401, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.010493169538676739, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.009631896391510963, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.006659789942204952, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0054494524374604225, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.005435348488390446, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0051854681223630905, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.004568049218505621, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.003173783188685775, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0027482975274324417, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.002629871480166912, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0018443071749061346, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0016471896087750793, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0016944839153438807, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.001464659464545548, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0014763767831027508, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0015607763780280948, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0010762496385723352, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0011836771154776216, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.014673582278192043, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.01227201335132122, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.010799620300531387, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.00807532574981451, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.007066160906106234, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.005323657765984535, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.003679591231048107, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.009257213212549686, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.008521493524312973, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.005538214463740587, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0046549574472010136, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.004677408374845982, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.004461691249161959, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.003995605744421482, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.002679026685655117, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.002266601426526904, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0023057812359184027, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0015905248001217842, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0013964649988338351, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0014857229543849826, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0012658037012442946, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.001372583326883614, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0013879017205908895, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0010440282057970762, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0010621384717524052, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.01227201335132122, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.09196989983320236, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.08208704739809036, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07791601121425629, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06855352967977524, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.042103707790374756, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03778506815433502, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.05176282301545143, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.046082694083452225, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.04328572377562523, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03563762456178665, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03396272659301758, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02656196989119053, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.022014405578374863, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02018250711262226, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.019735487177968025, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.013267249800264835, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.01031904760748148, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.010137846693396568, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.009008662775158882, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.008780090138316154, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.006808065343648195, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.006376439239829779, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005894260480999947, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003955882508307695, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.09196989983320236, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13655447959899902, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12377826869487762, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11985409259796143, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10197852551937103, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06309933215379715, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.058623310178518295, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07189344614744186, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06564344465732574, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0641510933637619, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.053044483065605164, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04895009472966194, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.036950115114450455, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.032038625329732895, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.030909840017557144, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03064197674393654, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01879849098622799, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.017124036327004433, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01703205704689026, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015302775427699089, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01525137573480606, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.010768468491733074, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.011984918266534805, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.010439588688313961, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.009686422534286976, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13655447959899902, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.08860991895198822, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.08412596583366394, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.08282820135354996, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.07612214982509613, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.03983459994196892, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.03842613846063614, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.04383113607764244, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.04056362807750702, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.04019266366958618, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.03667730093002319, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.035495828837156296, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.02232036180794239, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.019554175436496735, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.019239475950598717, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.019161364063620567, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.011249978095293045, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.010493423789739609, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.010475052520632744, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.009925908409059048, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.00991802103817463, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.006349359638988972, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.0071335118263959885, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.0062570576556026936, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.005678353365510702, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.08860991895198822, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.10680502653121948, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.10150900483131409, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.09997916221618652, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.09196339547634125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.04815320670604706, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.04644039273262024, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.05301227793097496, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.04902494698762894, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.048605725169181824, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.04430462792515755, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.04279288277029991, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0267640333622694, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.02326188236474991, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.022878019139170647, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.022787034511566162, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.013360388576984406, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.011796656996011734, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.011772152967751026, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.011041635647416115, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.011035756208002567, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.0070900036953389645, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.007148101460188627, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.006966180633753538, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.004845446906983852, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.10680502653121948, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.03673749789595604, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.03522762656211853, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.02039894089102745, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.018890641629695892, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.014923491515219212, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.00927356630563736, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.03492014482617378, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.03168430179357529, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.01516218762844801, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.014090068638324738, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.014028330333530903, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.013308520428836346, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.012596934102475643, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.007816767320036888, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.006120610050857067, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.00708960322663188, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.004899360705167055, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.003043845761567354, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.0047747730277478695, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.0028292308561503887, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.004465261474251747, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.004562824498862028, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.002247442491352558, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.002269558608531952, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.03673749789595604, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.02472532168030739, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.02243785746395588, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.020173825323581696, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.01788662001490593, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.011164248920977116, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0096213910728693, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.016843900084495544, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.014000137336552143, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.011436864733695984, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.00978226587176323, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.009894128888845444, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.008292179554700851, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.006714478600770235, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.005543949082493782, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.005230989772826433, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.004246972035616636, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0031979382038116455, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0030449614860117435, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.002964828861877322, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.002795588690787554, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0025267864111810923, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0024970739614218473, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.002131519140675664, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0019998380448669195, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.02472532168030739, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.025010373443365097, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.02246316336095333, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.020435091108083725, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.01802785135805607, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.011160844005644321, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.009661242365837097, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.014855056069791317, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.013608317822217941, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.011468270793557167, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.00965457409620285, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.009311527945101261, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.00738667044788599, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.006452571600675583, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0053879800252616405, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.005109773483127356, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0037224830593913794, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.002871622797101736, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.002734869485720992, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0025875966530293226, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0024278881028294563, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.002005623886361718, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.002023313893005252, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0016711960779502988, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.001408014795742929, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.025010373443365097, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11630116403102875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.10731445252895355, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.10426757484674454, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09368010610342026, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05376787483692169, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.050260789692401886, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.061521973460912704, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05616167560219765, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.054702311754226685, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04736258089542389, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.045135632157325745, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.031334541738033295, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.026813356205821037, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.025691654533147812, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.02541336789727211, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.015638461336493492, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.013009884394705296, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.012914344668388367, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.011679575778543949, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.011577097699046135, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00799236074090004, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.007645328063517809, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007480506785213947, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004599499981850386, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11630116403102875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.11777584254741669, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.10001963376998901, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07768239825963974, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06487210839986801, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.052990756928920746, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03777739778161049, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09165783226490021, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07862646877765656, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.05600325018167496, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04384443908929825, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04519560560584068, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04677173122763634, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.038399189710617065, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.026796584948897362, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.023267466574907303, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.023702280595898628, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.015849905088543892, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.014260332100093365, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.014358659274876118, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01245202962309122, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013386649079620838, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.013541484251618385, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009623156860470772, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010515647009015083, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.11777584254741669, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.12596124410629272, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.11888732761144638, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.11676052212715149, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.10627662390470505, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.05752889811992645, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.05511312931776047, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.06368908286094666, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.05878989398479462, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.058157749474048615, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.05220296233892441, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.049937065690755844, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.03228281810879707, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.028094327077269554, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.027534062042832375, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.027402246370911598, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.016180370002985, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.01446455530822277, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.014423731714487076, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.013436419889330864, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.013419999741017818, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.008767404593527317, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.009153581224381924, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.008592861704528332, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.006632274016737938, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.12596124410629272, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.13652177155017853, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.12891235947608948, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.12663255631923676, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.11527940630912781, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.06237738952040672, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.05975035950541496, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.06888654083013535, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.06373573839664459, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.06305385380983353, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.056585319340229034, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.05401954427361488, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0348435677587986, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.030279526486992836, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.02967616729438305, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.029537063091993332, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.017370227724313736, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.015193263068795204, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.0151521572843194, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.014026121236383915, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.014009607955813408, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.009083746932446957, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.009044582955539227, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.008890273049473763, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.005898888222873211, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.13652177155017853, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.14840815961360931, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.13841378688812256, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.13440163433551788, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.12220267206430435, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.0680711641907692, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.0644087940454483, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.0776323676109314, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.07158290594816208, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.06915298104286194, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.06159430742263794, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.05882922559976578, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.03938987851142883, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.03416401147842407, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.0326240137219429, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.032257068902254105, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.019736165180802345, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.017062446102499962, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.016918469220399857, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.015811000019311905, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.015607978217303753, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.010630577802658081, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01082670409232378, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.010148458182811737, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.007622633595019579, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.14840815961360931, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.03409404307603836, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.030023682862520218, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.027078120037913322, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.023735148832201958, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.01536154467612505, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.013027006760239601, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.02208864688873291, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.01877172663807869, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.015913721174001694, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.013067949563264847, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.012964637950062752, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.01109776645898819, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.00905599258840084, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.007522357162088156, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.007116405759006739, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.005659159272909164, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.004170449450612068, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.003984586335718632, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.003758201375603676, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0035337060689926147, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0032039557117968798, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.003106064395979047, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0026400804053992033, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0023569928016513586, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.03409404307603836, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.03482254594564438, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.030279647558927536, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.02722226269543171, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.023814033716917038, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.015433749184012413, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.013003958389163017, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.020617762580513954, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.01887442171573639, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.01606152392923832, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.012973965145647526, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.012466566637158394, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.010369454510509968, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.00897547323256731, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.007449023891240358, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.007051127962768078, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0051889363676309586, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.003942354582250118, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.003755607409402728, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.003472562413662672, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.003235449083149433, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.002731619169935584, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0027597646694630384, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.002220574300736189, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00187487225048244, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.03482254594564438, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11496822535991669, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1049628034234047, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.10151149332523346, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09063541889190674, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05319217965006828, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04923534765839577, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.060839492827653885, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05603396147489548, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05435599014163017, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04624562710523605, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0438038632273674, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03097742237150669, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.026787420734763145, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02546490728855133, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.025146640837192535, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.015475732274353504, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.012942804954946041, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.012830875813961029, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.011496484279632568, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.011360996402800083, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00793139636516571, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.007693557534366846, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00743882404640317, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004698657896369696, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11496822535991669, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.11190550029277802, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07559055835008621, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.052473269402980804, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05556851997971535, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04798922315239906, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.024966387078166008, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.0712571069598198, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06499889492988586, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.053029198199510574, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.034843187779188156, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.036397311836481094, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.036696746945381165, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03154197335243225, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.023970743641257286, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.02182825095951557, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01858735829591751, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01379905454814434, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.012991752475500107, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.011631123721599579, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.010318789631128311, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.010432648472487926, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.011140533722937107, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008145486004650593, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.008657682687044144, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.11190550029277802, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.14335867762565613, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.13542568683624268, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.13304182887077332, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.12134559452533722, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.06606537848711014, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.06327205896377563, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.0729290023446083, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.06749093532562256, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.06678523868322372, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.05995119363069534, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.057355694472789764, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.03708323463797569, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.03228890150785446, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.03166094422340393, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.031516097486019135, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.01855548843741417, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.016618553549051285, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.016578691080212593, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.01543559692800045, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.015413223765790462, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.010002411901950836, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.010496537201106548, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.009802911430597305, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.007588544860482216, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.14335867762565613, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1509719341993332, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.1426282823085785, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.14015331864356995, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.12785014510154724, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.06948723644018173, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.06654853373765945, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.07665751874446869, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.07099267840385437, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.07024969905614853, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.06304873526096344, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.06023125723004341, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.038809534162282944, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0337626114487648, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0330950953066349, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.032937441021203995, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.019366638734936714, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.016903283074498177, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.016856875270605087, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.015591444447636604, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.015571287833154202, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.010118680074810982, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.009996149688959122, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.009899492375552654, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.006430234294384718, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1509719341993332, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.14711228013038635, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.13671888411045074, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.13271868228912354, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.12008945643901825, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.06750379502773285, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.0637693703174591, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.07690785825252533, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.0709269568324089, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.06865068525075912, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.06071428582072258, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.05778305605053902, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.03904605284333229, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.03383579105138779, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.032339997589588165, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.03198425471782684, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.019536485895514488, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.01687280647456646, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.01673976518213749, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.015555493533611298, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.01535928063094616, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.010487829335033894, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.010647287592291832, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.010016405954957008, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.007434256374835968, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.14711228013038635, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.04148281365633011, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.03780278563499451, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.03551460802555084, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.031715601682662964, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.01893792487680912, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.017029521986842155, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.024786612018942833, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.02158069983124733, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.019371792674064636, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.016577305272221565, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.016244933009147644, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.012458747252821922, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01036469079554081, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.009183228947222233, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.008880943059921265, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.006312012206763029, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.004942575003951788, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.004798088222742081, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.004499698523432016, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.004333708435297012, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.003499451093375683, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0034324880689382553, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00304010440595448, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0025146768894046545, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.04148281365633011, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.040316034108400345, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.03648410364985466, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.034059710800647736, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.030344359576702118, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.01820068433880806, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.016230838373303413, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.022817067801952362, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.02094443142414093, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.01865921914577484, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.015812382102012634, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.01519406121224165, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.011469035409390926, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.00997895933687687, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.008758675307035446, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.008448613807559013, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.00576100405305624, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.004611360840499401, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0044617196545004845, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.004154326859861612, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.003977826796472073, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0030535655096173286, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.003105451352894306, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0026565713342279196, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0021320038940757513, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.040316034108400345, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1189316138625145, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.11030533164739609, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1074109897017479, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09697753936052322, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05498068034648895, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.051671892404556274, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06223616376519203, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05716381222009659, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.055881839245557785, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04874778166413307, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.046470507979393005, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0315716527402401, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.027264142408967018, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.026249315589666367, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.025996915996074677, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.015740152448415756, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.013279835693538189, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.01319793425500393, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.01197830494493246, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.011896093375980854, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00805116631090641, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00775457127019763, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007648174185305834, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004644361790269613, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1189316138625145, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14424832165241241, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1029314324259758, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0845031887292862, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06825526803731918, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06071218475699425, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04358937591314316, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08210833370685577, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07560564577579498, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06796064972877502, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04160672053694725, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04161876067519188, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04225429892539978, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.036520108580589294, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.029877694323658943, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.028113245964050293, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.021289672702550888, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01668917015194893, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.016199866309762, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01342580746859312, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.012310861609876156, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.011878984048962593, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.012649561278522015, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009900359436869621, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.009712553583085537, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14424832165241241, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.14789073169231415, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.13975457847118378, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1373363584280014, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.12562809884548187, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.06833852082490921, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.06548859924077988, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.07554470002651215, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.06979626417160034, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.06908462941646576, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.062110405415296555, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.05972009152173996, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.038462452590465546, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.03339935094118118, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.03275201469659805, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.03259385749697685, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.019261471927165985, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.017160756513476372, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.017119571566581726, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.015950998291373253, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.015929294750094414, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.010433878749608994, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.010795442387461662, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.010227097198367119, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.007761371321976185, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.14789073169231415, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.15595248341560364, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.14740781486034393, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1448870599269867, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.13252916932106018, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.07195975631475449, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.06897469609975815, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.0793912336230278, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.07349221408367157, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.07275880128145218, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.06540600955486298, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.06278010457754135, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04022224619984627, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.034972961992025375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.034284889698028564, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.03412279486656189, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.020062485709786415, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.017496971413493156, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.017452731728553772, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.016163013875484467, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.016141654923558235, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.010459105484187603, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.010330621153116226, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01023248489946127, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.006618383806198835, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.15595248341560364, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.1404290795326233, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.12969322502613068, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.12556539475917816, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.11302004754543304, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.06427937000989914, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.06039317697286606, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.07354117929935455, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.067759670317173, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.06544818729162216, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.057348255068063736, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.05438590794801712, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.03735249117016792, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.032340362668037415, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.030811700969934464, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.030449597164988518, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.01871974766254425, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.01611720584332943, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.015977799892425537, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.014777814038097858, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.014573582448065281, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.010120063088834286, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.010238759219646454, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.009645495563745499, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.007199713494628668, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.1404290795326233, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.048931729048490524, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.04488562420010567, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0423886813223362, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.03796572610735893, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.02234795317053795, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.020249130204319954, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.02841091901063919, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.025205343961715698, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.02278743125498295, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.019668947905302048, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.019121114164590836, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.014348892495036125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.012086980044841766, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.010801179334521294, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.010479082353413105, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.0072193145751953125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.005767996422946453, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.005613977089524269, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.00525684142485261, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.005087465047836304, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.003906479571014643, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.003936168272048235, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0034097579773515463, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0028527078684419394, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.048931729048490524, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.043506354093551636, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.039687078446149826, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.03710561990737915, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.03314035385847092, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.019621193408966064, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.017556969076395035, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.024563174694776535, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.02267395704984665, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.020042473450303078, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.01721540465950966, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.016519157215952873, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.012385405600070953, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01078604906797409, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.009425708092749119, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.009091049432754517, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.006200932431966066, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.004947955254465342, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0047721462324261665, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.004481510259211063, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.004282892215996981, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0032884913962334394, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0033133660908788443, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0028587665874511003, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.002234791172668338, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.043506354093551636, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12384695559740067, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.11520606279373169, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.11235328018665314, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1016039326786995, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05716918781399727, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.053878623992204666, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06463030725717545, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.0593084841966629, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.058031871914863586, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.050868842750787735, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04864733666181564, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03277462720870972, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.028237314894795418, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.027258578687906265, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.02700607106089592, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.016352681443095207, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.013777104206383228, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.013692324049770832, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.012470239773392677, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.012395935133099556, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008351141586899757, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0080155273899436, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007945440709590912, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004776643589138985, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12384695559740067, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16353937983512878, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12565450370311737, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10642576962709427, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08589065819978714, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07256453484296799, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05457543954253197, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.0965779572725296, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08893103152513504, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07772493362426758, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05034405738115311, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04970613494515419, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04968218132853508, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.042863812297582626, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03544703125953674, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03347901627421379, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024967731907963753, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.019286876544356346, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.018522322177886963, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015278558246791363, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.014026707969605923, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013719621114432812, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014067848213016987, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01144977007061243, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010258536785840988, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16353937983512878, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.14905335009098053, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.14093786478042603, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.13854598999023438, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.12682844698429108, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.06881311535835266, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.06598770618438721, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.07596633583307266, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.07028397917747498, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.06956007331609726, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.06264767050743103, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.060295768082141876, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.03860997408628464, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0335797555744648, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.03293294087052345, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.03277890756726265, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.019298868253827095, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.017162002623081207, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.017117436975240707, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.01595093496143818, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.015929123386740685, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.010310517624020576, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.010665761306881905, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.010103187523782253, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.007522341795265675, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.14905335009098053, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.15897013247013092, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.15040729939937592, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1478479951620102, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.13536466658115387, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.07340577989816666, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.0703706443309784, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.0809907391667366, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.07493415474891663, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.07417073845863342, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.0668034479022026, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.06422597169876099, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04106362909078598, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.03567376360297203, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.03498387709259987, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.03482048958539963, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.020486891269683838, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.017899224534630775, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.017850860953330994, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.016561666503548622, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.01653713919222355, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01072531845420599, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.010629692114889622, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.010498282499611378, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.006900162436068058, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.15897013247013092, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.14763303101062775, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.13497404754161835, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.1297193020582199, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1167149469256401, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.06732229888439178, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.0625302791595459, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.07831628620624542, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.07210257649421692, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.06884446740150452, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.059683531522750854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.05664820969104767, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.039758652448654175, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.034439269453287125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.03233233839273453, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.031824760138988495, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.0199190154671669, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.017002545297145844, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.01680750958621502, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.01553676463663578, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.015247121453285217, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.010757708922028542, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.011002589948475361, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.010096131823956966, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.007822079584002495, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.14763303101062775, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06561559438705444, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.061040330678224564, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.05891482159495354, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05312729999423027, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.030040765181183815, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.028083251789212227, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03538089990615845, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.032068636268377304, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03048819862306118, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.02677641436457634, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.02581731230020523, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.017876502126455307, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.015308069065213203, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.014414913952350616, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.014199767261743546, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.008958823047578335, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.007547429297119379, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.00745363300666213, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.006909849587827921, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.006810919847339392, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.00478060357272625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.004840011242777109, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004434051923453808, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0033972710371017456, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06561559438705444, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.058148909360170364, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.0538763552904129, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.051958855241537094, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.04672557860612869, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.026377765461802483, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.024582406505942345, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.030601361766457558, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.028289781883358955, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.026802530512213707, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02343614585697651, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.022389695048332214, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.015391532331705093, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01346110925078392, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.012618053704500198, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.012413358315825462, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.007703038863837719, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.006515375804156065, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.006423055659979582, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.005929103121161461, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.005831561051309109, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00404325919225812, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004075948148965836, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0037714859936386347, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0027099496219307184, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.058148909360170364, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1394500881433487, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1307758241891861, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.12800787389278412, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.11638842523097992, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.0643475279211998, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.06115151196718216, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.07195120304822922, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.0661395937204361, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.06512842327356339, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.05779561772942543, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.055332958698272705, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03642876073718071, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.03147716075181961, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.03062514029443264, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.030430646613240242, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.018169086426496506, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.015517444349825382, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.015453044325113297, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.014179370366036892, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.01412460021674633, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.009313947521150112, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0090226661413908, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.008969900198280811, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0055076684802770615, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1394500881433487, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1816468983888626, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12959258258342743, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.09866124391555786, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08817347884178162, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07781434059143066, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04974460229277611, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.11274900287389755, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.10393359512090683, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0869065448641777, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05585208535194397, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05654798820614815, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05799108371138573, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.050162434577941895, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.038389045745134354, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03510947525501251, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.029049718752503395, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02137247659265995, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.020180335268378258, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.017530532553792, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015408736653625965, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01584528014063835, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.016632776707410812, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012213397771120071, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01226914394646883, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1816468983888626, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.15683425962924957, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.14832265675067902, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.14572341740131378, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1334589719772339, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.07269790023565292, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.06967491656541824, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.08043801039457321, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.07430706918239594, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.07348810136318207, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.0661529153585434, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.06370730698108673, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.040896251797676086, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.03556404262781143, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.034857217222452164, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.034689974039793015, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.020458120852708817, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.018225226551294327, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.018177634105086327, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.016943631693720818, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.01692003197968006, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.0109761543571949, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.011420488357543945, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.010749255307018757, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.008155877701938152, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.15683425962924957, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.17086748778820038, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.16162829101085663, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.15884096920490265, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.14547745883464813, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.07914693653583527, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.07586570829153061, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.08742450177669525, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.08088267594575882, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08002355694770813, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07198089361190796, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.06921209394931793, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.044392235577106476, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.03852951154112816, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.03776397183537483, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0375845804810524, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.022159971296787262, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.019340844824910164, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.019288061186671257, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.01788904145359993, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.017864294350147247, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.011642001569271088, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.011518960818648338, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01138992141932249, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.007517124526202679, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.17086748778820038, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.1471477895975113, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.1335284411907196, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.12769901752471924, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.11460783332586288, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.06700689345598221, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.06174776330590248, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.07870977371931076, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.07245834916830063, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.06872617453336716, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.05902761593461037, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.05598047375679016, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.039950862526893616, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.034724775701761246, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.03230559825897217, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.03171496465802193, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02015635184943676, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.01717592403292656, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.01694713532924652, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.0156700536608696, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.015332430601119995, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.011099965311586857, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.011402629315853119, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.010360294952988625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.008333931677043438, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.1471477895975113, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06591829657554626, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.061388049274683, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.05934523046016693, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.053517237305641174, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.030286788940429688, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.028341129422187805, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.035542577505111694, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03230222687125206, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.030743256211280823, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.02700694091618061, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.026019398123025894, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.017990127205848694, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.015439517796039581, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.014531544409692287, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.014316064305603504, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009025349281728268, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.007604503538459539, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0075083584524691105, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.006962338928133249, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.006864331196993589, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.004800387658178806, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.004865214694291353, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0044441525824368, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0034048259258270264, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06591829657554626, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05866772681474686, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.05439567193388939, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.0524386465549469, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.04705421254038811, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.02663779817521572, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02477610670030117, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03096039965748787, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.02869805134832859, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.027051089331507683, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.023667067289352417, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0225826408714056, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01558434683829546, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.013659561052918434, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.012741653248667717, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.012511113658547401, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.00780765525996685, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0065850852988660336, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00648023234680295, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.005994399543851614, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.005882294848561287, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004108757711946964, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004139096941798925, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.003821589285507798, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.002751184394583106, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05866772681474686, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.13527895510196686, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.12681862711906433, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.12416238337755203, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.11256378889083862, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.06232424080371857, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.059251848608255386, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06983644515275955, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06409448385238647, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.06314469128847122, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.05593408644199371, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.05355517938733101, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03534437343478203, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.030508050695061684, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02968033403158188, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.029488762840628624, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.017627885565161705, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.015029734000563622, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.014967922121286392, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.013708395883440971, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.01366288773715496, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.009032391011714935, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008720144629478455, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.008685664273798466, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005271261557936668, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.13527895510196686, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.20928895473480225, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1393919736146927, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10808972269296646, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0977592021226883, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.087374746799469, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05533573403954506, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.12194870412349701, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.11009322851896286, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0999487116932869, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05997747555375099, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0612797774374485, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.06302153319120407, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.05402546375989914, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04394527152180672, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04126212000846863, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03268655762076378, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02573569118976593, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02509261667728424, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02134762704372406, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01976999081671238, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.019752178341150284, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.020523538812994957, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.017047416418790817, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01674622669816017, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.20928895473480225, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1503855586051941, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.14222434163093567, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1396075040102005, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.12791964411735535, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.07051554322242737, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.06750469654798508, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.07838345319032669, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.07229436188936234, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.07127880305051804, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.06420783698558807, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0619308277964592, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04040083289146423, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.03546910732984543, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.03466753289103508, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.03446445241570473, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.020548749715089798, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.019464384764432907, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.019407469779253006, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.018344495445489883, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.018300523981451988, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.011979387141764164, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013874560594558716, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01174754649400711, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01151498593389988, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1503855586051941, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1860155612230301, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.17583034932613373, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.17261098325252533, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.15801337361335754, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08688922971487045, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08318880200386047, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.09621745347976685, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.08899608254432678, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08784065395593643, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07900941371917725, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07610570639371872, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04924085736274719, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0430087149143219, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.042037975043058395, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04180886968970299, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.024707915261387825, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.022547051310539246, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.022471468895673752, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.021049529314041138, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.020999480038881302, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01362279150635004, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014892307110130787, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.013335132971405983, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011403840966522694, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1860155612230301, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.027310218662023544, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.025298165157437325, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.018780838698148727, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.01703638583421707, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.012726516462862492, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.009186895564198494, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.03604897856712341, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.020531801506876945, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.013044522143900394, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.01162511482834816, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.011689269915223122, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.014798035845160484, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.009836426004767418, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.007080339826643467, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.005867602303624153, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.007876858115196228, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.0044874935410916805, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.003783967811614275, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.004330835305154324, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.003582252189517021, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.00510379858314991, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.004092404153198004, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.003614787245169282, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.00313211465254426, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.027310218662023544, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07946176826953888, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07400485128164291, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07160743325948715, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06438387930393219, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03668566793203354, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.034334298223257065, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04208778589963913, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.038947511464357376, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03720041364431381, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03259315714240074, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.031079759821295738, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.021311765536665916, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.018615975975990295, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01759183779358864, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01734447106719017, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01068685669451952, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.009161947295069695, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009053494781255722, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008355596102774143, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008245673030614853, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005683157593011856, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005796747747808695, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005358800757676363, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003992717247456312, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07946176826953888, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06697653979063034, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06219470128417015, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06002524495124817, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05376267805695534, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.030693408101797104, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.028619440272450447, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.035534314811229706, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03280173987150192, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.031117483973503113, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.027195092290639877, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.025952234864234924, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.017941247671842575, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.015647463500499725, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014698443002998829, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014468222856521606, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.008973448537290096, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007603330537676811, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00749860517680645, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006909464485943317, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006804015953093767, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004715262912213802, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004762780852615833, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004412383306771517, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0031927076634019613, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06697653979063034, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.18603576719760895, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.17430780827999115, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1706211417913437, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.15457597374916077, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.08647230267524719, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.082137331366539, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.09666764736175537, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.08883179724216461, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.08762290328741074, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.07734742760658264, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.07397110760211945, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.04897993057966232, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04230101406574249, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04122338071465492, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04096554219722748, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.024429360404610634, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.020827990025281906, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.020755577832460403, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.018950102850794792, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.01888909377157688, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01250347588211298, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01202878076583147, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.012052169069647789, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.007223848719149828, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.18603576719760895, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18562471866607666, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12432875484228134, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.09475365281105042, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08710171282291412, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07885021716356277, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.048129767179489136, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10826065391302109, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09980795532464981, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08950565755367279, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05368828773498535, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.054714202880859375, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.056196179240942, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.048802655190229416, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03945506364107132, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03694511950016022, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.028463443741202354, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.022751448675990105, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.022073831409215927, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.018622932955622673, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01709030568599701, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.016210975125432014, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.017921550199389458, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013498433865606785, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.014313532039523125, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18562471866607666, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.18775252997875214, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.17715458571910858, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.17392361164093018, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.15913094580173492, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08775588870048523, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08393806219100952, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.09747924655675888, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.08983157575130463, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08878044784069061, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07957887649536133, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07658969610929489, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0497652068734169, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04314963147044182, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0422397255897522, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04202500358223915, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.024948399513959885, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.022235766053199768, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02217310480773449, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.020641757175326347, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.020606420934200287, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.013534469529986382, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014147184789180756, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.013244095258414745, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010313970036804676, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.18775252997875214, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.20760920643806458, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.1959717571735382, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19245049357414246, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17592093348503113, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09679359942674637, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09257204830646515, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10703813284635544, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09902232140302658, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09791078418493271, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08770779520273209, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08408020436763763, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05436524748802185, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.047206517308950424, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.046207889914512634, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.045973341912031174, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.027124609798192978, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.023579945787787437, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02350660040974617, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02171291410923004, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.021676797419786453, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014141098596155643, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013911111280322075, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.013814899139106274, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.008871132507920265, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.20760920643806458, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.1854202002286911, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.16931764781475067, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.1622915118932724, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.14559194445610046, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.0843871533870697, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.07828318327665329, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.09851185977458954, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.09082470089197159, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.08633464574813843, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.07473733276128769, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.07077110558748245, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.04988909512758255, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04337570071220398, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.040557656437158585, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.039861008524894714, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.025162342935800552, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02136966586112976, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02109411172568798, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.019528180360794067, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.019134391099214554, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.013739111833274364, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.013909452594816685, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.012859047390520573, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009920354932546616, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.1854202002286911, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06741170585155487, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.0628383457660675, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06052209809422493, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.054503340274095535, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03120247647166252, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.029069242998957634, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03649616986513138, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03357679024338722, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0316476970911026, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.027769120410084724, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.026489663869142532, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.018488634377717972, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.016060903668403625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.014995248056948185, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.014737308956682682, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009319991804659367, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.007850268855690956, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007727984804660082, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007175924256443977, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007052384316921234, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005052560940384865, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005048890132457018, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004707013722509146, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003516385331749916, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06741170585155487, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05865894630551338, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.054516684263944626, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.052430782467126846, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.047014642506837845, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.026935048401355743, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.024984022602438927, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03164130449295044, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.029159897938370705, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.027325225993990898, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.023901909589767456, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02282899059355259, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.015970876440405846, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01392183918505907, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.012929058633744717, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.012686599977314472, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.008011340163648129, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.006729443091899157, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00661216676235199, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006132518872618675, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006010296754539013, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004240215290337801, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004297318402677774, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.003925536293536425, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0029286397621035576, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05865894630551338, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1658099889755249, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.15501229465007782, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1516476422548294, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.13701477646827698, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.07685372978448868, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.07287847250699997, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.08610080182552338, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.07918812334537506, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.07794689387083054, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.06861040741205215, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.06542950123548508, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.04364212602376938, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.03767888620495796, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.03664623573422432, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.036401353776454926, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.021758267655968666, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.01849517971277237, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.018414773046970367, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.01677621528506279, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.016720198094844818, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.011111827567219734, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.010655326768755913, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.010670799762010574, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.006317203864455223, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1658099889755249, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18539580702781677, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12257201969623566, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.08731833100318909, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08114724606275558, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07498946040868759, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04594247043132782, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.11288220435380936, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.10433229058980942, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08927668631076813, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.054991599172353745, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05677662417292595, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.058533359318971634, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.051619358360767365, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.038723792880773544, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03506999462842941, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03008406050503254, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.023989757522940636, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.023016050457954407, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.021024296060204506, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0190443005412817, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.017376871779561043, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.02058263123035431, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013619670644402504, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.017237398773431778, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18539580702781677, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.20488335192203522, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19309136271476746, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.1895124465227127, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17300310730934143, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09589993208646774, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09161466360092163, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10636498779058456, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09814758598804474, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09702638536691666, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.0866844654083252, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0832502469420433, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05421622097492218, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04701890051364899, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04600738361477852, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04576782509684563, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02712906524538994, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.023943493142724037, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02387518249452114, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.022120404988527298, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.022081587463617325, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014483067207038403, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014839407987892628, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014150017872452736, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010384741239249706, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.20488335192203522, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2211194783449173, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2084072083234787, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.20456713438034058, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1866406351327896, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1033300831913948, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.0987258329987526, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11434093117713928, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10571788996458054, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10455265641212463, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09335166215896606, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08942895382642746, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05812443047761917, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05045075714588165, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04937189444899559, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.049121178686618805, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.029012108221650124, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.025215256959199905, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.025145037099719048, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.023170214146375656, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.023132894188165665, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01516683679074049, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014912174083292484, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014813737943768501, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009559153579175472, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2211194783449173, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.1950710564851761, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.17778272926807404, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.17023329436779022, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.15254122018814087, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.08865288645029068, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.08208136260509491, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10368393361568451, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.09549785405397415, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09070364385843277, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.07837188243865967, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.07416439801454544, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05244424194097519, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.045583710074424744, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04254503920674324, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.041799623519182205, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02646256797015667, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.022325724363327026, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.022021101787686348, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02035459317266941, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.01992158591747284, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.014481785707175732, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014431814663112164, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01354210078716278, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010142931714653969, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.1950710564851761, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08682696521282196, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.08059411495923996, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0780484601855278, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0700579434633255, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04009570926427841, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03750438988208771, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.045891012996435165, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04242052137851715, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0406908355653286, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.035485416650772095, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03373587131500244, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.023245183750987053, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.020254062488675117, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.019215142354369164, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.018966220319271088, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.011633053421974182, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.009971754625439644, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009865556843578815, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.00905870832502842, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008951089344918728, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006126937922090292, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006247327197343111, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005776289850473404, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004249706398695707, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08682696521282196, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07557014375925064, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06997478753328323, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06751763075590134, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06044158339500427, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.034641481935977936, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03226836398243904, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.040005750954151154, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03697226941585541, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03518431633710861, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.030615098774433136, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.029097184538841248, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.020177096128463745, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.017617471516132355, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.016558799892663956, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.016312047839164734, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.010101951658725739, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.008495232090353966, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008382522501051426, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007678356021642685, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007567306514829397, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005270062945783138, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0052093458361923695, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004932360257953405, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003350856015458703, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07557014375925064, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.17089208960533142, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.15946999192237854, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.15589939057826996, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.14056631922721863, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.07926423847675323, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.07502634078264236, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.08873522281646729, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.08170859515666962, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.08034893870353699, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.07056471705436707, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.06712175160646439, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.04494819790124893, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.038893841207027435, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.03776620328426361, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.03750169649720192, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.022408824414014816, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.019079718738794327, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.01899867318570614, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.01727329008281231, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.01720566488802433, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.011462208814918995, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.011027850210666656, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01101488433778286, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.006575681734830141, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.17089208960533142, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.19006621837615967, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14178504049777985, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1241440623998642, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09724263846874237, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08515901118516922, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06548752635717392, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10595040023326874, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09622499346733093, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.09100212156772614, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05619273707270622, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05386247858405113, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.054718952625989914, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04746771603822708, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.042572494596242905, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04134675860404968, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.028396114706993103, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.024623166769742966, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.024318907409906387, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.019939489662647247, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.019316604360938072, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.017196251079440117, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.018745770677924156, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.015893440693616867, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01561233215034008, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.19006621837615967, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1875474750995636, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.176650732755661, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.17329534888267517, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.15822485089302063, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08777204155921936, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08383738249540329, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.09749790281057358, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.08986356854438782, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08882930874824524, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07935548573732376, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07626301795244217, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.04977816343307495, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04311741515994072, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.042186833918094635, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.041967302560806274, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.024946289137005806, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.022071663290262222, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.022008221596479416, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.020412258803844452, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.0203707292675972, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.013441493734717369, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013837482780218124, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01313148345798254, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009859899058938026, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.1875474750995636, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2152981162071228, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.20283183455467224, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19907380640506744, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18151535093784332, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10065833479166031, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09609811007976532, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11137878894805908, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10298997908830643, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10184578597545624, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09090659022331238, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0870559811592102, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05666683241724968, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04916366562247276, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04810253903269768, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04785221070051193, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.028283268213272095, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02459144778549671, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02451871708035469, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.022596945986151695, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.022558260709047318, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014810836873948574, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014577864669263363, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01446626242250204, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009389063343405724, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2152981162071228, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.18534980714321136, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.16812029480934143, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.16113582253456116, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1441173106431961, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.08402711898088455, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.07777445018291473, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.09776493906974792, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.09010052680969238, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.08613884449005127, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.07393006235361099, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.06980746984481812, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.04939592629671097, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04292553290724754, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.040276266634464264, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.039634816348552704, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.024762287735939026, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02105369046330452, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.020819345489144325, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.01910822093486786, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.018740136176347733, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.013287828303873539, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.013465599156916142, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.012461090460419655, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.009377534501254559, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.18534980714321136, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06854096055030823, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06375055760145187, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06154898554086685, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05525048449635506, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03167921304702759, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.029577363282442093, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03651006147265434, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03388205170631409, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.032140206545591354, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.02811165153980255, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.026737457141280174, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.01852472499012947, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.016199683770537376, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.015214361250400543, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.014973797835409641, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.00928310863673687, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.00794792827218771, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007839389145374298, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007246847730129957, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0071379197761416435, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.004935016855597496, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0050796945579349995, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004625482950359583, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003529789624735713, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06854096055030823, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06736557185649872, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06263609230518341, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.0602835938334465, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05399060994386673, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.030954468995332718, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.028783241286873817, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.036068402230739594, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03345543146133423, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.0314057394862175, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.027438726276159286, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.026152215898036957, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.018222225829958916, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.015954993665218353, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014823258854448795, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014551340602338314, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009110014885663986, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007649010978639126, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007520910818129778, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006947848945856094, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006814233958721161, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004767922218888998, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0047914995811879635, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004418517462909222, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0031461443286389112, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06736557185649872, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.16019350290298462, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.14959076046943665, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.14617952704429626, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.13177643716335297, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.07409796118736267, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0701732486486435, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.08329203724861145, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.07649108022451401, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.07510805130004883, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.06601624935865402, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.06279978901147842, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.04220785200595856, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.03639522194862366, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.03531792387366295, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.03506892919540405, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02104448340833187, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.01784341037273407, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.017758987843990326, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.01616942323744297, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.016096362844109535, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.010772187262773514, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.010344373993575573, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.010316092520952225, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0061953128315508366, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.16019350290298462, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.22365982830524445, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1857871413230896, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.17373055219650269, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.13585898280143738, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.10233387351036072, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.08800403028726578, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1192125603556633, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.10951336473226547, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.10619654506444931, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07585775852203369, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06558596342802048, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0613228902220726, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.053078845143318176, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.049876876175403595, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.049117837101221085, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03091028891503811, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.027084294706583023, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02687568962574005, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02224637381732464, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.02195354737341404, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.017305251210927963, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.018565258011221886, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.016339872032403946, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.014317255467176437, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.22365982830524445, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.18574316799640656, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.17504703998565674, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.17177759110927582, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1566557139158249, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08688680827617645, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.08305526524782181, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.09618394821882248, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.08894559741020203, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08792194724082947, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07854938507080078, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.07535096257925034, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.049121957272291183, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04267711564898491, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04176822677254677, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0415504053235054, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.024562854319810867, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.021882984787225723, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02182464674115181, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.020250020548701286, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02021387219429016, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01320695225149393, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013781863264739513, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.012914035469293594, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.00989076029509306, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.18574316799640656, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.21464116871356964, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.20226165652275085, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19856953620910645, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18100880086421967, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10025522112846375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09577388316392899, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11091021448373795, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10257641226053238, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10145416855812073, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09055154770612717, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08671337366104126, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05640022084116936, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04893018305301666, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04788709059357643, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.047638241201639175, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02814382314682007, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.0244460329413414, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.024376753717660904, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02246161364018917, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02242421917617321, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.014713377691805363, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.014448893256485462, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014366427436470985, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.009246990084648132, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.21464116871356964, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.17601384222507477, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.1596171110868454, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.1520070731639862, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1359953135251999, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.07983963936567307, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.07340483367443085, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.09448187798261642, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.08716857433319092, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.08184276521205902, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.07029866427183151, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.0664544329047203, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.04764971509575844, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04158351570367813, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.03846748545765877, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.03767997771501541, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.024003956466913223, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.020398549735546112, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.020057380199432373, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.018600821495056152, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.018126297742128372, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.013003128580749035, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.013531772419810295, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.011986641213297844, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.00974353402853012, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.17601384222507477, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08676750957965851, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.0808553546667099, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07823601365089417, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07036406546831131, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.040023356676101685, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03749033808708191, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04677809029817581, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04253312945365906, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.040587037801742554, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03562511131167412, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03422323241829872, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.023725375533103943, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02034974843263626, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.019209234043955803, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.018928680568933487, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.011899641714990139, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01002155989408493, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009893739596009254, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.009157275781035423, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.009032903239130974, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006340385414659977, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006365959532558918, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00588812492787838, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004406373482197523, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08676750957965851, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07304362207651138, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06802824139595032, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06532236188650131, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05863482505083084, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03353428095579147, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.031139109283685684, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03934547305107117, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.0364665761590004, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03401960805058479, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.029854092746973038, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.028469543904066086, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.019811805337667465, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01739300601184368, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.016080528497695923, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.015751434490084648, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009925649501383305, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.008309434168040752, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00814912747591734, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007572341710329056, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007405662443488836, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005232230760157108, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005239751189947128, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004822395276278257, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003440593369305134, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07304362207651138, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.20531928539276123, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.19219335913658142, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.18810248374938965, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1700991839170456, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09561397135257721, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09078846871852875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10717146098613739, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.0982871800661087, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09687086939811707, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08544876426458359, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08144848793745041, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05432269349694252, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04683225601911545, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04562290757894516, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04531087726354599, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.027139978483319283, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02305719628930092, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.022973140701651573, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02092607505619526, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.020866025239229202, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.013876896351575851, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013319318182766438, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013334771618247032, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.007974429987370968, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.20531928539276123, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.27608394622802734, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.21735991537570953, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.19650447368621826, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.14549773931503296, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.12464779615402222, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.10084514319896698, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.15152424573898315, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.13880394399166107, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.13221612572669983, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.08395352959632874, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.07697393000125885, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0780620127916336, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06724870949983597, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.060835469514131546, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.05925308167934418, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.039568908512592316, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03294844925403595, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.03252075985074043, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02552664279937744, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.024708928540349007, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.02227410301566124, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.022854682058095932, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.020326822996139526, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01722012832760811, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.21735991537570953, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2062767893075943, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19404251873493195, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19033661484718323, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.17309874296188354, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09673994779586792, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09229006618261337, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.10718633234500885, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.09905661642551422, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.09791908413171768, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08718574792146683, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08345682919025421, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05478733032941818, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.047661587595939636, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04662198945879936, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04637919366359711, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.027493532747030258, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.024621857330203056, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.024554774165153503, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.022766973823308945, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.022727150470018387, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.015000675804913044, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.015758616849780083, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.014681634493172169, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01158601138740778, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2062767893075943, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.232026144862175, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.21824908256530762, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.21410974860191345, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.19465288519859314, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10847216844558716, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10345545411109924, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11996569484472275, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11105451732873917, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10980860888957977, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09765755385160446, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09323440492153168, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06099648028612137, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05298355221748352, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0518234521150589, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05155382305383682, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0304307471960783, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02646181546151638, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.026383474469184875, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02424868382513523, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.024205755442380905, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01586500182747841, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01564152166247368, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.015484758652746677, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010000312700867653, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.21824908256530762, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2097468227148056, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19202157855033875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18522977828979492, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.16568055748939514, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09534269571304321, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.08918263763189316, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10942604392766953, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10069567710161209, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09727882593870163, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08442779630422592, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.07964517176151276, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05537349358201027, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04796293377876282, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04563043266534805, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.045081861317157745, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.027837326750159264, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.023830126971006393, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02363189123570919, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02174071967601776, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02142808772623539, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015097495168447495, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01510531548410654, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014357442036271095, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010564721189439297, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2097468227148056, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.10652513056993484, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.09916491061449051, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.09629462659358978, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08653074502944946, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04929085820913315, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.046338364481925964, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.05609605088829994, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.05174313858151436, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.049984369426965714, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.043773483484983444, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.041629064828157425, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02844608947634697, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.024733930826187134, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.023631049320101738, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.023366576060652733, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.014246216043829918, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.012272711843252182, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.012164855375885963, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.011181749403476715, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.011074931360781193, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.007555933203548193, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.007674308959394693, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.007178450934588909, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.005262915976345539, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.10652513056993484, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08995437622070312, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.08368458598852158, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.0808882862329483, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.07253938168287277, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.041376348584890366, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.038709379732608795, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04749893397092819, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04406564310193062, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.041969358921051025, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03671471029520035, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.034869901835918427, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.023992864415049553, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.021016323938965797, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.019810041412711143, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01951514557003975, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.011998264119029045, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.01019041147083044, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.010063459165394306, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.009256838820874691, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.009122677147388458, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.006277203559875488, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.006290406920015812, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005901062861084938, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0041094678454101086, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08995437622070312, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.20816034078598022, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.19425052404403687, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.18993589282035828, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1711285561323166, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09666405618190765, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09159901738166809, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10811766982078552, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.09941384941339493, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09806141257286072, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0860520675778389, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08182719349861145, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05479370430111885, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04734009876847267, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04610206186771393, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04579516500234604, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02731418050825596, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02329465001821518, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.023207930848002434, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.021086223423480988, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.021033119410276413, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.013969678431749344, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01344240177422762, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013442604802548885, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008038761094212532, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.20816034078598022, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2773757576942444, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.2144296020269394, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1888626217842102, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.15459686517715454, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.12285338342189789, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.09586676955223083, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1560983806848526, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.14355120062828064, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.13298065960407257, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.08968541771173477, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.08236946910619736, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0803372785449028, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06943254917860031, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.06008390337228775, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.05768342688679695, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.04063396155834198, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03274198994040489, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.03206414356827736, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.026715625077486038, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.02532157115638256, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.022845979779958725, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.02343478985130787, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.020106296986341476, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.017554273828864098, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.2144296020269394, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.22208668291568756, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.208785280585289, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.20471076667308807, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1861240118741989, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1042691022157669, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09941626340150833, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11550906300544739, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10679860413074493, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1055595651268959, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09384440630674362, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08979663252830505, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05899476632475853, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05130161717534065, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.050176508724689484, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.04991157725453377, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02954847365617752, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02633947692811489, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.026262249797582626, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02428850717842579, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02425275184214115, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.015915129333734512, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016634970903396606, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.015553304925560951, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011988767422735691, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.22208668291568756, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24665899574756622, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2319130003452301, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22741663455963135, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2065924108028412, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11554043740034103, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11014582961797714, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12787318229675293, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11832926422357559, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11697269976139069, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10390157252550125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09912874549627304, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06500231474637985, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.056486405432224274, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05523465573787689, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05494185537099838, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03245484456419945, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.028203293681144714, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.028120556846261024, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.025816993787884712, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.025773685425519943, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016925649717450142, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.016678337007761, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016513964161276817, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01066082064062357, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22741663455963135, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21513377130031586, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.1959633231163025, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18871870636940002, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1684034913778305, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09755684435367584, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.0909804180264473, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11203490942716599, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10327938199043274, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09968630969524384, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.0860263779759407, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08095534890890121, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05661098659038544, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04915459454059601, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04666978120803833, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04607608914375305, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.028323141857981682, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024308549240231514, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.024097774177789688, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.022091010585427284, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.021758412942290306, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015135669149458408, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.015341781079769135, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014355385676026344, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010630465112626553, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21513377130031586, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1095091924071312, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.10168197751045227, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.09852440655231476, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08829262107610703, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.05063536390662193, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04740989953279495, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.05845480039715767, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.05347292870283127, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.05138600245118141, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04481162875890732, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04275355115532875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.029630735516548157, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02557586133480072, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.024293681606650352, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.02398589439690113, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01485743559896946, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01263368595391512, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.012502760626375675, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.011485052295029163, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.011353121139109135, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.00788347702473402, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00794987566769123, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.007398281712085009, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0054628197103738785, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1095091924071312, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.09245563298463821, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.08578339964151382, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.08252058178186417, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.07388642430305481, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.04256230220198631, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.039566103368997574, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04945698752999306, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04578869417309761, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.04319275543093681, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.037616368383169174, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03577229008078575, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.024952491745352745, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02186872996389866, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.020411096513271332, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.020057912915945053, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01250490639358759, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.010585227981209755, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.010422750376164913, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0095985671505332, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.00943627581000328, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.006597003433853388, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.006673030089586973, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.006144429091364145, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004486306104809046, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.09245563298463821, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.24502550065517426, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.22867530584335327, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22373566031455994, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2017821967601776, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11462873220443726, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10861319303512573, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12811943888664246, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.1177654042840004, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11617855727672577, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.1020306721329689, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09698595851659775, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.064971923828125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05618467181921005, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05470680072903633, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05436565354466438, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03241221234202385, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.027653206139802933, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.027550294995307922, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.025024639442563057, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.024956682696938515, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016583994030952454, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015965061262249947, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01598460040986538, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009540271013975143, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.22867530584335327, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.28624051809310913, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1974589228630066, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.15918207168579102, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1336953490972519, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.12388158589601517, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0848505049943924, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1663547158241272, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.14996199309825897, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.13783174753189087, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07996876537799835, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0822928324341774, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.08604838699102402, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0734037533402443, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.06173663213849068, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.05872207507491112, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.04451717436313629, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.035289786756038666, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.034531764686107635, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.028051340952515602, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.026225361973047256, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.026749761775135994, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.02711481600999832, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.02355123683810234, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.021756242960691452, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1974589228630066, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.23295967280864716, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.21874792873859406, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.21442027390003204, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.19468340277671814, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10960064828395844, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10438650846481323, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12155868113040924, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11232639849185944, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1109800934791565, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09844844788312912, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09401482343673706, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.062164951115846634, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.053971290588378906, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.052752863615751266, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0524645633995533, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03112713061273098, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.027648139744997025, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02757180482149124, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.025454070419073105, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02540300227701664, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.016770072281360626, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01742091029882431, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016385836526751518, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012496011331677437, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.21874792873859406, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.25674131512641907, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.24107538163661957, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23632961511611938, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21444323658943176, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12047095596790314, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11467508226633072, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.133345827460289, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12342419475317001, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12199173122644424, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10807999223470688, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10293315351009369, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06784456223249435, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05895489826798439, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.057615917176008224, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05729682743549347, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03385741636157036, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.029433535411953926, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.029343195259571075, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02689373306930065, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02684405818581581, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017634162679314613, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017422107979655266, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017193090170621872, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011157490313053131, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21444323658943176, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2199653536081314, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20032617449760437, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19278699159622192, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17210376262664795, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09968909621238708, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09290139377117157, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11466860771179199, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10573119670152664, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10193374007940292, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08792085200548172, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08280825614929199, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05789680778980255, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05025887116789818, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04764002561569214, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04702029004693031, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.028947120532393456, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024733465164899826, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02450869046151638, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.022455992177128792, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.022102661430835724, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015425151214003563, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.015502827242016792, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014604792930185795, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010581741109490395, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2199653536081314, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.10829851776361465, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.10056496411561966, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.09725279361009598, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08729856461286545, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.05021921917796135, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.046962372958660126, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.05761043727397919, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.05328759923577309, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.05096739903092384, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04448547586798668, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04223664104938507, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.029212500900030136, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.025515243411064148, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.024138104170560837, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.023804612457752228, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01464549545198679, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.012612566351890564, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.012467577122151852, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.011485639028251171, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.011340229772031307, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.007787984795868397, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.008032292127609253, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0073406882584095, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0056053767912089825, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.10829851776361465, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.09682582318782806, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.08975744247436523, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.08631732314825058, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.07734182476997375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.04463598504662514, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04143412038683891, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.052017692476511, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04810291901230812, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.0453161858022213, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0395079031586647, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03752781078219414, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02627166174352169, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02296326868236065, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.021395361050963402, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.02101972885429859, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.013171222060918808, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.011079911142587662, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.010904453694820404, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.010051420889794827, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.009867481887340546, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.006945091765373945, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.006966735236346722, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.006460293661803007, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004642980173230171, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.09682582318782806, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.22751030325889587, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.21225419640541077, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.20754875242710114, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18691973388195038, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10609745979309082, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10051140934228897, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11880605667829514, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10914109647274017, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10760324448347092, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09441177546977997, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0897214263677597, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06023566052317619, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.052042629569768906, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.050663091242313385, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05031486973166466, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.030057568103075027, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.025577396154403687, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.025482572615146637, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0231251809746027, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.023052236065268517, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015353373251855373, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01474007498472929, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014771966263651848, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008746471256017685, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.22751030325889587, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2808287441730499, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.20453409850597382, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1735498607158661, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.13526348769664764, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.1253393143415451, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.09175962954759598, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1587604284286499, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.1453859508037567, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.1348501294851303, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07851564139127731, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.07796226441860199, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0816514864563942, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.07025794684886932, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.06128257140517235, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.05898603796958923, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.04119447246193886, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03328036144375801, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.03260239213705063, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02490388974547386, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.02337372861802578, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.022873535752296448, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.023608706891536713, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.020161893218755722, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01759531907737255, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.20453409850597382, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24023190140724182, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22541938722133636, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22095829248428345, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2004852145910263, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11316816508769989, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10770733654499054, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12542809545993805, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1159721240401268, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1145952045917511, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10149867832660675, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09688390046358109, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06418051570653915, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05579661950469017, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05453558638691902, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.054241281002759933, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.032167527824640274, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.028710749000310898, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02862866222858429, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02643604390323162, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026384351775050163, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017409371212124825, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01826271042227745, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017001252621412277, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013285422697663307, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22541938722133636, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.25763314962387085, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2417692095041275, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2369488924741745, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21485885977745056, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12097016721963882, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11512217670679092, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1338910311460495, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12397807091474533, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1225191280245781, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10844887048006058, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10323330014944077, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06810697913169861, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.059225138276815414, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05786620453000069, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.057551272213459015, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03399849683046341, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.029566224664449692, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02947445772588253, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026995766907930374, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026943495497107506, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01770428754389286, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01751171238720417, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017255809158086777, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01121513731777668, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21485885977745056, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21795660257339478, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19765272736549377, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.1896616369485855, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.16873405873775482, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09880007058382034, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09165641665458679, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11433679610490799, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10524296760559082, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10107030719518661, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.0865950658917427, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08146160840988159, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05782531201839447, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05009974539279938, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04732194170355797, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04664812982082367, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02899574674665928, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024737462401390076, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.024489471688866615, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.022402634844183922, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.022023463621735573, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015686651691794395, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01577777788043022, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01481888722628355, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011040977202355862, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.21795660257339478, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12088937312364578, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11205478012561798, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1084250807762146, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0970708578824997, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.05609159171581268, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.052417151629924774, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06469418108463287, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.059345245361328125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.05695308372378349, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04949163272976875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0471220500767231, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03285563737154007, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.028397969901561737, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.02693108096718788, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.02657397836446762, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.016459936276078224, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.014003480784595013, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.013855023309588432, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01269612368196249, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01254549901932478, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.00872047059237957, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.008809614926576614, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00818889681249857, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0060401917435228825, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12088937312364578, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10652455687522888, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.09853429347276688, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.09486200660467148, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.08467563986778259, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.049114227294921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04566144943237305, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.05684254318475723, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05264986306428909, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.04989955574274063, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04325252026319504, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04097926244139671, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02870839275419712, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.025098126381635666, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.023519255220890045, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.02314244955778122, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.014357308857142925, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.012116026133298874, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.011939777061343193, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.010937632992863655, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.01075693964958191, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.007505424786359072, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.007517100777477026, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00699961744248867, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0049027325585484505, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10652455687522888, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.23201881349086761, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.21600165963172913, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.21092589199543, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18945053219795227, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10818955302238464, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10217330604791641, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12131670117378235, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11158500611782074, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10974878072738647, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09584441035985947, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09103188663721085, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.061578020453453064, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0531870499253273, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0516597256064415, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05129305273294449, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.030727244913578033, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.026171375066041946, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.026057712733745575, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02360820583999157, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.023513920605182648, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01574673503637314, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015211739577353, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015127725899219513, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009221501648426056, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.21600165963172913, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2726522386074066, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.2168881893157959, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.19443073868751526, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.15702687203884125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.12555867433547974, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.10132087022066116, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1535291224718094, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.14088274538516998, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.13083918392658234, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0889924094080925, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.08194447308778763, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07936106622219086, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0682678148150444, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.061345282942056656, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.05964570865035057, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.04025774076581001, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.033408693969249725, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.032788995653390884, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.026766015216708183, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.025795292109251022, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.022815795615315437, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.023448163643479347, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.02073783427476883, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.017841363325715065, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.2168881893157959, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2390364110469818, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2242049276828766, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2196677178144455, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.19913646578788757, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11253233999013901, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10703825205564499, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12478984147310257, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11536753922700882, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1139756366610527, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10079360008239746, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09616842865943909, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06376807391643524, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05541041120886803, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05412815511226654, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.053824249655008316, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03193502128124237, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.028316043317317963, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02823394350707531, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02600616030395031, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.025957418605685234, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017164837568998337, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017769845202565193, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01675504632294178, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012655668891966343, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2242049276828766, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.25743627548217773, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.24138236045837402, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23651278018951416, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21427693963050842, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12090203911066055, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11498475074768066, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13387899100780487, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12394121289253235, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12247584760189056, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10824008285999298, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10295268893241882, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06811866164207458, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0592147558927536, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.057833172380924225, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05750589445233345, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.034003350883722305, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.029523484408855438, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.029434071853756905, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026922134682536125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.0268690325319767, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01768963411450386, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017453858628869057, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017231740057468414, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01113746128976345, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21427693963050842, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2215278148651123, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20120124518871307, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19284828007221222, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1721271425485611, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10038545727729797, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09308724850416183, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.1169968992471695, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.1074308380484581, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.1027444526553154, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08838231861591339, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08334411680698395, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.059078238904476166, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05119713023304939, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04815787076950073, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04743035510182381, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.029723413288593292, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.025327159091830254, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.025056779384613037, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.023047521710395813, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02263161540031433, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01623569056391716, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.016374973580241203, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.015287934802472591, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011666071601212025, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2215278148651123, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13196510076522827, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12240204960107803, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11863186955451965, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10634127259254456, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06130640208721161, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0574016198515892, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06976800411939621, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06458530575037003, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06225689873099327, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0541931614279747, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.051388438791036606, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0354141965508461, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03091398999094963, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.029431357979774475, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.029080096632242203, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01772923953831196, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.015310744754970074, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.015163528732955456, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.013898828066885471, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013756303116679192, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.00935710221529007, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009611646644771099, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008874617516994476, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006612653844058514, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13196510076522827, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1147712841629982, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.10643713176250458, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.10248836129903793, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09168700873851776, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.053161006420850754, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.049438055604696274, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06166337430477142, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05701570585370064, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05398673564195633, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04695281758904457, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.044586118310689926, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03117378056049347, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02720733731985092, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02548389881849289, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.025065556168556213, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.015600061044096947, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.013114568777382374, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.01292031817138195, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.011869116686284542, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.011660962365567684, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008174215443432331, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008127512410283089, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007627129554748535, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005286889616400003, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1147712841629982, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.24136348068714142, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.22498852014541626, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22001108527183533, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19810925424098969, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11284670978784561, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10666956752538681, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1261799931526184, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11608030647039413, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11445823311805725, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10015837848186493, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09517577290534973, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06407102197408676, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05534672364592552, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05386372283101082, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05351344123482704, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.031960975378751755, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.027232129126787186, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.027135008946061134, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02459089644253254, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02451123483479023, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016345705837011337, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015737351030111313, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015733471140265465, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009428231976926327, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.22498852014541626, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2818790674209595, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.22113411128520966, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1943829357624054, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.15455572307109833, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.12667316198349, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.09954697638750076, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.163938969373703, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.14909382164478302, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.13553285598754883, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.09188326448202133, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.08527354896068573, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.08434813469648361, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.07259956002235413, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.06251812726259232, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.059934720396995544, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.043372977524995804, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03491261228919029, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.034060731530189514, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.028843944892287254, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.02738283760845661, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0256053414195776, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.02586817368865013, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.022858038544654846, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.020237896591424942, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.22113411128520966, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2431914210319519, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2277982383966446, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22310124337673187, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2018118053674698, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11471901088953018, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10901054739952087, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1274416595697403, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11768756806850433, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11623457819223404, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.1025342270731926, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09768585860729218, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06526587158441544, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.056601256132125854, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05526609346270561, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05495245382189751, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.032698750495910645, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.029013287276029587, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.028925679624080658, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026623373851180077, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026569616049528122, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017682965844869614, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.018354082480072975, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017260346561670303, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013225299306213856, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2277982383966446, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2686273157596588, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2516349256038666, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2463979721069336, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22289538383483887, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12638098001480103, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12003277987241745, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14027346670627594, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1296433061361313, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12803953886032104, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11289689689874649, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10734286904335022, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07150444388389587, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06199052184820175, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06050928682088852, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06015888229012489, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03572510555386543, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.030964892357587814, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03086554817855358, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.028206074610352516, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.028142770752310753, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.018769847229123116, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.018425732851028442, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01828586496412754, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011902114376425743, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22289538383483887, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.23142653703689575, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.21138721704483032, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.20350810885429382, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18171589076519012, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10511597990989685, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09813804179430008, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12139934301376343, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11144164204597473, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10732338577508926, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09293964505195618, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08753201365470886, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06126837804913521, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05312119424343109, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.050388775765895844, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.049738720059394836, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.030879320576786995, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.026455221697688103, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.026216687634587288, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02413727343082428, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.023773612454533577, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01696779951453209, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.016976390033960342, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.016140641644597054, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.01207374781370163, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.21138721704483032, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14111866056919098, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13096050918102264, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1270238608121872, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11399528384208679, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06587709486484528, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06168396770954132, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07484094798564911, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06916837394237518, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06687211245298386, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.058159153908491135, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.055153101682662964, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03800906240940094, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.033164843916893005, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03167833015322685, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.031320970505476, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.019045807421207428, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.016496771946549416, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0163559727370739, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.014967397786676884, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.014829367399215698, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.010082433931529522, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.010372830554842949, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009611197747290134, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.007183460984379053, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14111866056919098, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12481066584587097, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.11572080105543137, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.11157675087451935, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09996307641267776, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.058084696531295776, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.054041966795921326, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.0672663226723671, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06201865151524544, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05897122994065285, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.05127749964594841, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0486413836479187, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03404572233557701, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.029679764062166214, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02789592370390892, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.02746405266225338, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01704631745815277, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.01438064407557249, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.014191441237926483, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.01300036907196045, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.012801949866116047, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008930591866374016, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008917153812944889, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.008378030732274055, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005865633022040129, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12481066584587097, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.27448171377182007, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2558850347995758, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2503257095813751, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.22537927329540253, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1290009468793869, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12200057506561279, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.14414538443088531, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.132720947265625, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.13081981241703033, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11440765112638474, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10872934013605118, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07326853275299072, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06332357227802277, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.061602991074323654, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.061220843344926834, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03655194118618965, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.03113238699734211, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.031010258942842484, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.028094351291656494, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02801370434463024, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.018683914095163345, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.017942139878869057, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01799483224749565, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010663625784218311, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.22537927329540253, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.27997738122940063, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.20841243863105774, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.17691688239574432, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.14431171119213104, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.12372076511383057, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.09279800206422806, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1626003384590149, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.14856836199760437, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.1345885992050171, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.08653083443641663, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.08342576026916504, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.08509720116853714, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.07332290709018707, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.06231573596596718, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.05954071506857872, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.04472553730010986, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03666134551167488, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.03581107035279274, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.03067227639257908, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.029096504673361778, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.028150109574198723, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.028859589248895645, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.025354323908686638, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.02399696782231331, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.20841243863105774, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2420964539051056, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22660723328590393, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22185255587100983, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20044264197349548, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11433860659599304, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10855910927057266, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12701818346977234, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11732429265975952, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11587084084749222, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10203298926353455, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09711890667676926, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06498190015554428, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05642174929380417, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05508588254451752, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05476687476038933, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0326097272336483, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.028896497562527657, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.028807170689105988, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026472702622413635, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026417089626193047, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01765800267457962, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.018238969147205353, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017235910519957542, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013095777481794357, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22660723328590393, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2759968042373657, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.25824031233787537, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.25287285447120667, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22847004234790802, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1299903392791748, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12337364256381989, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14414584636688232, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13337311148643494, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13171672821044922, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11593306064605713, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11015821993350983, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07348056137561798, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06377775967121124, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06224546208977699, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06187921389937401, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03668729588389397, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03182796388864517, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.031724121421575546, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.028941456228494644, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.028877535834908485, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.019189737737178802, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01890077069401741, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.018685143440961838, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012156138196587563, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22847004234790802, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.23286336660385132, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.21491047739982605, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.2080591320991516, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1864788681268692, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10614365339279175, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.1000438705086708, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12087635695934296, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11142893135547638, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10795338451862335, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09473171830177307, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08933820575475693, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06094571575522423, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05296488478779793, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05069791525602341, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.05015471577644348, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03053966723382473, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02626682259142399, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02606121264398098, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02406085655093193, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.023761847987771034, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01635264791548252, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.016311779618263245, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.015657763928174973, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011103082448244095, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.21491047739982605, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12916181981563568, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12015850841999054, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11641830950975418, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10454784333705902, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.060506422072649, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.056709688156843185, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06909361481666565, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06387060135602951, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.061401382088661194, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0535886287689209, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05085732042789459, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03509185463190079, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.030605319887399673, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.029107268899679184, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.028747018426656723, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.017605703324079514, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.015144570730626583, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.014991841278970242, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.013760175555944443, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013612253591418266, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009343664161860943, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00950952060520649, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008868739008903503, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0065301284193992615, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12916181981563568, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11714786291122437, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.10892486572265625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1050800010561943, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09421928972005844, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05477851256728172, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.05110712721943855, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06331957131624222, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.058448947966098785, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.0555972121655941, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.048451848328113556, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.045984476804733276, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.032016098499298096, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02795417606830597, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.026294289156794548, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.025900375097990036, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01604470983147621, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.013525007292628288, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.013343549333512783, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.012239823117852211, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.012056947685778141, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008410811424255371, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008339360356330872, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007902503944933414, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0054208762012422085, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11714786291122437, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2724071443080902, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2543392777442932, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.24855685234069824, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.22378447651863098, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1284080296754837, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12138102203607559, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.14375582337379456, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.13224995136260986, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.13014985620975494, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11395980417728424, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.1082366481423378, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07309487462043762, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06318771094083786, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06140385940670967, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06098105013370514, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03646396845579147, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.03107941523194313, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.030948098748922348, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.028048155829310417, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02796073630452156, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.018668457865715027, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.018013987690210342, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01794811151921749, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010847222059965134, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.22378447651863098, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.25038936734199524, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.18986037373542786, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1577678918838501, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1259741187095642, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.11312904208898544, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.08184259384870529, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.159303680062294, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.13953489065170288, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.12081591784954071, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07973892986774445, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0782485231757164, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.08364017307758331, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06987323611974716, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.05816367641091347, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.05502777546644211, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.048056162893772125, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03580938279628754, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.03472232446074486, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.030866378918290138, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.029295993968844414, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.03451357036828995, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.029603004455566406, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.032050371170043945, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.025499720126390457, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.18986037373542786, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2658253610134125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.24876552820205688, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.24350851774215698, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22016578912734985, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12586915493011475, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11945709586143494, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1398143321275711, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1291845440864563, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1275641918182373, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11223327368497849, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10678782314062119, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.071588896214962, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.062122587114572525, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.060636211186647415, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.060281798243522644, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03585464134812355, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03174406662583351, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.031641844660043716, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.029042363166809082, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.028986437246203423, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01931118220090866, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019948044791817665, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.018839305266737938, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.014222677797079086, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22016578912734985, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2873380482196808, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.26879721879959106, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.26316574215888977, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23783724009990692, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13565696775913239, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12868694961071014, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15052852034568787, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13925862312316895, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13749563694000244, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12088610976934433, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11481380462646484, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07678935676813126, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06660958379507065, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06498624384403229, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06460437178611755, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0383588932454586, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03322305902838707, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.033113863319158554, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03018949367105961, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03012351132929325, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020082350820302963, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019712474197149277, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01955481432378292, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01264700386673212, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15052852034568787, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.24752108752727509, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.22747474908828735, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.21982993185520172, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.19675017893314362, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.11277562379837036, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.10596499592065811, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12868645787239075, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.1186925396323204, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.11485353112220764, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.10025844722986221, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.09449980407953262, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06503269821405411, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05649449676275253, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05393960326910019, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.05332699045538902, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.032592542469501495, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.028096742928028107, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02787032723426819, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.025693396106362343, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.025356663390994072, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01754232496023178, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01767982542514801, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.016754694283008575, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.012274988926947117, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.22747474908828735, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12406925857067108, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11512130498886108, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11104792356491089, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09958409518003464, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.058108288794755936, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05417453497648239, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06707807630300522, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.062043022364377975, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.059013932943344116, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.051302678883075714, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04870350658893585, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.034090667963027954, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.029772980138659477, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.028014004230499268, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.027585018426179886, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01710387133061886, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.014651820994913578, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.014465246349573135, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01330258697271347, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01311218086630106, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009101017378270626, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00936852302402258, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008547850884497166, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006536533124744892, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12406925857067108, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1086699366569519, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.10072296112775803, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.09636890888214111, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.08618070930242538, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05071834474802017, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04683243855834007, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.059865858405828476, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05530142784118652, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05150140821933746, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04461713507771492, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.042405951768159866, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.030255848541855812, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.026455692946910858, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02439870499074459, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.023882415145635605, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.015163931995630264, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.01263373252004385, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.01239245105534792, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.011411950923502445, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.01116324495524168, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00797807052731514, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.007995049469172955, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007343432866036892, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0052955858409404755, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1086699366569519, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2668606638908386, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2488446682691574, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.24297833442687988, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2187858521938324, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.12574706971645355, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.11883559077978134, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.14112357795238495, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.12957416474819183, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.12756961584091187, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.111408531665802, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10570128262042999, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07178930938243866, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.061891525983810425, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.060125771909952164, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.059700798243284225, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03580712899565697, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.03041943348944187, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.030285824090242386, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.027412347495555878, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02731594257056713, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.018314750865101814, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.017599839717149734, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.017549466341733932, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010518976487219334, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2187858521938324, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.29086780548095703, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.22711068391799927, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.20339670777320862, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.15289419889450073, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.13330413401126862, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.10604573041200638, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.16182953119277954, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.14833243191242218, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.14005333185195923, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0901409313082695, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0809672549366951, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.08378495275974274, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.07205691188573837, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.06526223570108414, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.06358689069747925, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.04255406931042671, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.035706229507923126, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0351850762963295, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02798810787498951, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.02711445279419422, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.024262692779302597, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.02515689842402935, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0222332663834095, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.019364064559340477, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.22711068391799927, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2721402645111084, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.25453028082847595, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.24914415180683136, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22510011494159698, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12886355817317963, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12225018441677094, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14307671785354614, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13229671120643616, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13062506914138794, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11482635140419006, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10916057974100113, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07320833206176758, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06356199830770493, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.062022726982831955, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.061659082770347595, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03663663938641548, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03235950320959091, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03225881978869438, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02956058643758297, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.029498931020498276, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.019611140713095665, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020177597180008888, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019115641713142395, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.014208241365849972, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22510011494159698, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2897801399230957, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.27097469568252563, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2652578055858612, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23964127898216248, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13687343895435333, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12980633974075317, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15178586542606354, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14050190150737762, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1387358456850052, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12188661098480225, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11570005118846893, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0773773267865181, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06720400601625443, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06555040180683136, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06515812873840332, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03862691670656204, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03347430378198624, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.033361922949552536, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030386464670300484, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.030314486473798752, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020123276859521866, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01980070397257805, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01958467625081539, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012618163600564003, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15178586542606354, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2533106505870819, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.23282793164253235, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.22503341734409332, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.20131760835647583, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.11540848016738892, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.10843675583600998, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.13160055875778198, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.12129954993724823, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.11752776056528091, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.10249379277229309, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.09659573435783386, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06633134186267853, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05761902406811714, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05509236827492714, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.05449008196592331, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03320183977484703, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02848542109131813, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02826884388923645, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.025978248566389084, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.025646505877375603, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.017694629728794098, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.017622951418161392, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.016912510618567467, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011897776275873184, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.22503341734409332, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12374360859394073, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11496201157569885, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11076484620571136, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09935051202774048, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.058094095438718796, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05402087792754173, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06720677763223648, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.062291085720062256, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.05894157662987709, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0513264462351799, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0487716943025589, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03419312834739685, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.029933663085103035, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.02803993597626686, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0275783259421587, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.017160138115286827, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.014734268188476562, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.014528341591358185, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.013403383083641529, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013197142630815506, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009172042831778526, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009532160125672817, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008595435880124569, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.00673729507252574, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12374360859394073, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10959254205226898, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.10179716348648071, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.09752293676137924, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.08717088401317596, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05130545422434807, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04744905233383179, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.060561809688806534, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.056030742824077606, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05209219455718994, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04531833902001381, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.043202802538871765, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03068755380809307, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02681593783199787, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02470802143216133, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.024190349504351616, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.015373926609754562, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.012799983844161034, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.012555425055325031, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.011601378209888935, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.01134332176297903, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008112993091344833, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00812562182545662, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007477085571736097, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005395026411861181, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10959254205226898, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2688758969306946, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2505967319011688, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2447778731584549, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2205650508403778, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.12674789130687714, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.1197739765048027, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.142051562666893, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.13062502443790436, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.12857799232006073, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11229151487350464, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10667987167835236, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07231919467449188, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06241482123732567, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06061931326985359, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06019533798098564, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03609327971935272, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.03070494346320629, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.030560573562979698, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.027683651074767113, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02759574167430401, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01849447190761566, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.017846645787358284, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.017738480120897293, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010794347152113914, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2205650508403778, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2750093340873718, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.20109082758426666, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1695646047592163, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.13777419924736023, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.11962888389825821, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.08850549906492233, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.156636580824852, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.144308403134346, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.13211512565612793, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.08134201914072037, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0804925486445427, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.08107449114322662, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06983478367328644, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.05870729312300682, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.055810753256082535, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.041147567331790924, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03222158923745155, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.03141319751739502, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.025384223088622093, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.02356347069144249, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.02313396893441677, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.02356799878180027, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01984637975692749, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01765420101583004, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.20109082758426666, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2738807499408722, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.25615835189819336, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.25071683526039124, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22638961672782898, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12976574897766113, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12304364144802094, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14422693848609924, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13319844007492065, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1315094530582428, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11553384363651276, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10986903309822083, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07378353178501129, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06404313445091248, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.062486547976732254, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0621182955801487, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.036970507353544235, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.032685041427612305, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03258129954338074, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.029873527586460114, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02980537712574005, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.019894611090421677, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020505238324403763, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019403398036956787, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.014558643102645874, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22638961672782898, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2905341386795044, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2717033624649048, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.265951007604599, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2402665764093399, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13730759918689728, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13020360469818115, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15239113569259644, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14099039137363434, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13921719789505005, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12226832658052444, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11606697738170624, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07772865891456604, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0674690529704094, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06581299006938934, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06542149931192398, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.038831137120723724, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.033661674708127975, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.033549606800079346, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030565522611141205, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.030493367463350296, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.02032376267015934, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019994407892227173, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01978600211441517, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012849453836679459, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15239113569259644, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.251954585313797, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.231466606259346, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.2236737608909607, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.2002323716878891, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.1147448867559433, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.10778742283582687, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.13073328137397766, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.1206524446606636, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.11685777455568314, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.10188092291355133, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.09604628384113312, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.0658268854022026, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05723375827074051, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05471498519182205, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.05411133915185928, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.032856978476047516, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.028150761500000954, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.027925999835133553, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02563626505434513, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.0252976156771183, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01726657897233963, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.017222827300429344, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.016472646966576576, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011361049488186836, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.2236737608909607, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13155212998390198, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12208256870508194, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11777276545763016, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10555680096149445, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.061717357486486435, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05745602026581764, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07121849060058594, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06593766808509827, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06267430633306503, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05448834225535393, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.051692500710487366, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0361725352704525, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.031635526567697525, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.02973663993179798, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.029278120025992393, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.018181880936026573, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.015527980402112007, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.015322517603635788, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.014089343138039112, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013879338279366493, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009698829613626003, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009891331195831299, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009111067280173302, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006838651839643717, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13155212998390198, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11851658672094345, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.10997997224330902, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.10554677993059158, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09455978870391846, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05553055182099342, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.051388222724199295, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06495574116706848, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.0601579025387764, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.056363970041275024, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04893313720822334, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04657456651329994, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03287442773580551, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02875431999564171, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.026665719226002693, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.02615930512547493, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.016469702124595642, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.013766398653388023, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.013529643416404724, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.012437102384865284, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.012197337113320827, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008637328632175922, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008617601357400417, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.008017631247639656, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005651948973536491, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11851658672094345, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.27607056498527527, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.25714340806007385, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2512260675430298, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2259075790643692, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.13014572858810425, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12288792431354523, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.14607660472393036, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.13411560654640198, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.13203418254852295, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11518280953168869, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.1093376874923706, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07427477091550827, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06405194848775864, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06225638836622238, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06184476241469383, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03707798942923546, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.031506866216659546, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.031380828469991684, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.028384000062942505, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.028282294049859047, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01898048259317875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01827680878341198, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01820344477891922, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010993048548698425, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2259075790643692, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2991289794445038, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.232665553689003, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.20808333158493042, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.16309115290641785, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.1364201158285141, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.10796692967414856, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.16631026566028595, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.152173712849617, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.14417356252670288, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.09368661791086197, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.08553922921419144, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.08569713681936264, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.07446940243244171, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.06739544123411179, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0656387209892273, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.04404839128255844, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03758768364787102, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.03709007799625397, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.03015187196433544, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.029214203357696533, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.025700189173221588, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.027235999703407288, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.023693012073636055, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.021674400195479393, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.20808333158493042, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.27063554525375366, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.25294941663742065, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.24754463136196136, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22359967231750488, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1282699704170227, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12159594148397446, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14263564348220825, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13171853125095367, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13003692030906677, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11417863517999649, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10856816172599792, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07309509068727493, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06335430592298508, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.061798058450222015, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06143150478601456, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03665949031710625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.032325033098459244, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.032222893089056015, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02952791564166546, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02946333773434162, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.019822414964437485, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.02026752009987831, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01933193765580654, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.014389864169061184, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22359967231750488, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.28961634635925293, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2707146406173706, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2649418115615845, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23921693861484528, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13692472875118256, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12973321974277496, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15213431417942047, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14057248830795288, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13879220187664032, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12182208895683289, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.1156868115067482, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07765095680952072, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06729812920093536, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0656311959028244, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06523190438747406, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03885948285460472, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.033612001687288284, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.0335022509098053, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030516285449266434, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.030440010130405426, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020511262118816376, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020035704597830772, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.0199724230915308, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012969665229320526, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15213431417942047, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.24732527136802673, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.22704863548278809, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.21945440769195557, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.19637957215309143, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.11263608187437057, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.105797678232193, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12822827696800232, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11836843192577362, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.11475276201963425, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09995448589324951, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.0941920056939125, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.0647512823343277, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.056211765855550766, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05375278368592262, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.0531613826751709, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03236599266529083, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02775764837861061, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.0275459885597229, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.025292472913861275, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.024965642020106316, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01719924621284008, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.017123553901910782, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.016423041000962257, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.01149529218673706, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.22704863548278809, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.11573166400194168, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1073208823800087, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10328898578882217, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09243592619895935, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.05421437695622444, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05038636922836304, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06291869282722473, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.05831315368413925, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.055091775953769684, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04781636223196983, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04531398043036461, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03191201016306877, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02790945954620838, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.02609642595052719, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.025661680847406387, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01599002443253994, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.013589807786047459, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0133895855396986, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01231340691447258, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.012105097994208336, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.008452465757727623, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.008633464574813843, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.007889954373240471, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0058901640586555, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.11573166400194168, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11172860860824585, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.10352800041437149, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.09952403604984283, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.08888158202171326, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05218927189707756, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04836392402648926, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.060906462371349335, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.056315090507268906, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05300868675112724, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04598630592226982, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.043643768876791, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.030850602313876152, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.026938119903206825, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.025049610063433647, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.024602225050330162, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.015455933287739754, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.01290384866297245, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.012695017270743847, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.011647844687104225, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.011425137519836426, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008093634620308876, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00802802573889494, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0075058466754853725, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005207541864365339, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11172860860824585, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.24735330045223236, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.23003850877285004, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22448213398456573, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2016458362340927, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11612838506698608, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10954871773719788, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1306246668100357, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.12009693682193756, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11783009767532349, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10274263471364975, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09741126000881195, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0663713663816452, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05730869621038437, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.055516213178634644, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.055083807557821274, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.033132705837488174, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.028101306408643723, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02795623615384102, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02529997006058693, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.025173677131533623, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016960280016064644, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.016310669481754303, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.016231972724199295, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00978100672364235, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.23003850877285004, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2701915502548218, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.21158070862293243, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.19077813625335693, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.15175458788871765, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.12287802994251251, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.09943739324808121, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1487104743719101, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.1357630491256714, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.12917359173297882, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.08671896159648895, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.07896405458450317, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07648075371980667, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06589531898498535, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.060136184096336365, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.05870572105050087, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03873594477772713, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03286805376410484, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.032466232776641846, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02636178396642208, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.02565085142850876, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.021926037967205048, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.02306770533323288, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.020190879702568054, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.017817042768001556, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.21158070862293243, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2651827335357666, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2478652000427246, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.24256901443004608, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2190244197845459, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12567222118377686, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11914181709289551, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.139725461602211, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12907648086547852, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12741981446743011, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.1119285449385643, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10636954754590988, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07152733951807022, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06209712475538254, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06055786833167076, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06019854545593262, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03585618734359741, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.031722575426101685, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03162354975938797, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.0290012639015913, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02893165312707424, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.019350042566657066, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019964052364230156, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.018862850964069366, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.014248348772525787, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2190244197845459, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.28160274028778076, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.26314106583595276, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.257479727268219, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2325877547264099, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13308630883693695, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1260942667722702, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14765605330467224, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13666865229606628, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13491922616958618, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11841418594121933, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11235591024160385, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07528621703386307, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06537234038114548, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06373907625675201, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06335117667913437, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03759187459945679, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.032550837844610214, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03244117647409439, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.029528222978115082, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02945622242987156, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01959150657057762, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019263412803411484, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01904626004397869, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012274520471692085, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14765605330467224, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2155810296535492, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.19695094227790833, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.18960057199001312, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.16939932107925415, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09819116443395615, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09167075902223587, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11309538781642914, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10408476740121841, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10019389539957047, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08677246421575546, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08182711899280548, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05722656100988388, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04965599998831749, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.047124799340963364, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04651957377791405, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.028808822855353355, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024834975600242615, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.024611350148916245, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02267422154545784, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.022342555224895477, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015807097777724266, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01605542004108429, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.015042584389448166, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011565360240638256, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2155810296535492, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13357087969779968, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12386555969715118, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11971393972635269, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10716373473405838, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0625709667801857, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.058347322046756744, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07268495857715607, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06641885638237, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0635237842798233, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05512666329741478, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.052236493676900864, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0370430052280426, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0318865031003952, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.030133886262774467, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.029705775901675224, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.018634522333741188, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.015733961015939713, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.015548689290881157, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.014258969575166702, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.014074747450649738, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.00994906760752201, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.010000167414546013, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009286511689424515, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006945202127099037, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13357087969779968, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12086253613233566, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.11193335056304932, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.10754804313182831, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09605419635772705, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.056390877813100815, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.05224349722266197, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.0658704861998558, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06088586151599884, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05731562152504921, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04960256442427635, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04703451320528984, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.033358342945575714, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0291497390717268, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.027124058455228806, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.026638757437467575, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01670749858021736, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.014049948193132877, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.013819843530654907, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.012683318927884102, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.012446973472833633, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008800385519862175, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008853990584611893, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.008158596232533455, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005897709168493748, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12086253613233566, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2596774995326996, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.24159039556980133, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2359640896320343, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2120542824268341, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.12213671207427979, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.1152493953704834, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13698264956474304, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.1259315460920334, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.12390425056219101, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10806635022163391, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.1024676039814949, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0696089044213295, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06013462692499161, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05837589129805565, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.057981833815574646, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03473299741744995, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.029535779729485512, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02939891256392002, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.026566270738840103, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02646695449948311, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.017762675881385803, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.017073893919587135, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.017031298950314522, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010211079381406307, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2120542824268341, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.23997528851032257, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.2028958797454834, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.19151684641838074, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.14405132830142975, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.11164684593677521, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.09788381308317184, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.13005581498146057, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.1177380383014679, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.11535590887069702, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.08170993626117706, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.07061830163002014, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.06737211346626282, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.05777786672115326, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.05508621037006378, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.05441959947347641, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03495538607239723, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03078470565378666, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.030640168115496635, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02551797404885292, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.025365745648741722, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.021154921501874924, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.021907702088356018, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.020436635240912437, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.017848843708634377, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.2028958797454834, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.23873840272426605, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22326526045799255, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2185031771659851, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.197478249669075, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1129915788769722, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10718114674091339, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12537077069282532, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11599165201187134, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11452358216047287, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10070143640041351, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09580334275960922, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0641949251294136, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0558350570499897, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05448150634765625, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05416048690676689, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03214772418141365, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.028622878715395927, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.0285316351801157, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026202965527772903, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026147812604904175, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017305700108408928, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01812799647450447, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016874581575393677, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01307554543018341, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22326526045799255, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.26594433188438416, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.24863898754119873, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.24339893460273743, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21994896233081818, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1254911571741104, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11900170892477036, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13913625478744507, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12881281971931458, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12720099091529846, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11179192364215851, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10618608444929123, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07092288881540298, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06161438673734665, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06009513884782791, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05973770096898079, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.035417959094047546, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.030702149495482445, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03059965930879116, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.027881108224391937, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02781694009900093, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01847892254590988, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.018182892352342606, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017979159951210022, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011613317765295506, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21994896233081818, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.22525307536125183, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20678338408470154, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19964240491390228, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17892137169837952, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10251925140619278, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09616183489561081, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.1174757108092308, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10814913362264633, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10449105501174927, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09113572537899017, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.0860350951552391, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05935751274228096, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05145275592803955, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.049046047031879425, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04847265034914017, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02973671443760395, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02557511068880558, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.025366276502609253, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.023392707109451294, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.023074734956026077, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015990599989891052, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.016145547851920128, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.015232621692121029, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011250169947743416, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.22525307536125183, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13093437254428864, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12131449580192566, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1173991784453392, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10506924241781235, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.061273787170648575, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05720321089029312, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07021517306566238, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06476398557424545, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06223882734775543, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05396605283021927, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05108213424682617, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.035694755613803864, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.031040530651807785, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.029485033825039864, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.029108023270964622, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.017890220507979393, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01534774899482727, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.015192537568509579, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.013890091329813004, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013733968138694763, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009474978782236576, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00966739747673273, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008939802646636963, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006658035330474377, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13093437254428864, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11785659193992615, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.10910987108945847, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.10508417338132858, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09382904320955276, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05489695817232132, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.05094614252448082, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06360428780317307, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05893945321440697, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05576028302311897, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04836155101656914, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04577568918466568, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03216881677508354, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.028156740590929985, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0263720341026783, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.025919359177350998, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01610007882118225, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.01357981562614441, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.013370816595852375, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.012240096926689148, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.012032870203256607, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008409546688199043, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008416551165282726, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00783564429730177, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005481654778122902, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11785659193992615, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2478419989347458, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.23049499094486237, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2251405119895935, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20230445265769958, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11636897176504135, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10975013673305511, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13038600981235504, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11994194984436035, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1180817186832428, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10289520770311356, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0975601077079773, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06615174561738968, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05726727098226547, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05560975521802902, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05521282181143761, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03302159160375595, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.028098076581954956, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02798294462263584, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02528076060116291, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.025188108906149864, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01687418483197689, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01622699573636055, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.016202181577682495, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009641273878514767, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.23049499094486237, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2969628572463989, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.23375461995601654, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.211170494556427, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.16724400222301483, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.13582219183444977, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.10952623933553696, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1620619148015976, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.14946052432060242, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.14244377613067627, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.09392546862363815, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.08631078153848648, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.08320389688014984, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.07197790592908859, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.06581976264715195, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.06430886685848236, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.04180644452571869, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.034897804260253906, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.03446486219763756, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0269586481153965, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.02614590711891651, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.02273758128285408, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.023244399577379227, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.020830348134040833, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01667618751525879, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.211170494556427, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24082176387310028, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2251254916191101, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22028417885303497, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.19901980459690094, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11404421180486679, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10813441872596741, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12651947140693665, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1171155646443367, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11561624705791473, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10158832371234894, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09652012586593628, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06479138880968094, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05634624511003494, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05496583878993988, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05463573709130287, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03241313248872757, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.0288070198148489, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.028718814253807068, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02634703367948532, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026289457455277443, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01736408658325672, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.0181594118475914, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016928136348724365, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013001763261854649, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2251254916191101, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.27475446462631226, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2568272650241852, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.25137633085250854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22708289325237274, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12971054017543793, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12298211455345154, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14374187588691711, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13318482041358948, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1315014809370041, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11548791080713272, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10957314074039459, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07323423027992249, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0636945590376854, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.062119342386722565, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06174693629145622, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03655055910348892, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.031713664531707764, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03160723298788071, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.028777368366718292, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.028711292892694473, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.018970822915434837, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.018751919269561768, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.018451688811182976, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011929306201636791, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22708289325237274, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.228810653090477, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20987163484096527, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.20248790085315704, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18159592151641846, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10398485511541367, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09741182625293732, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11960039287805557, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10988946259021759, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10601459443569183, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09243649989366531, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08738246560096741, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06021546199917793, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05225580558180809, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04970340058207512, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04908597469329834, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.030143139883875847, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.025814885273575783, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02558586746454239, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02358764410018921, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.0232426468282938, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.016056332737207413, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01617126166820526, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01522742211818695, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011076641269028187, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.228810653090477, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.11178077012300491, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.10367187857627869, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.09973442554473877, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08930803835391998, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.05235655605792999, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04859267547726631, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.060921911150217056, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.05641038343310356, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.053166840225458145, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04619278013706207, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04386939853429794, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03096773661673069, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02709592878818512, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.0253008846193552, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.024860888719558716, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.015623639337718487, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.013352373614907265, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01315480750054121, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.012149633839726448, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01194689329713583, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.008473029360175133, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.008722608909010887, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.007937241345643997, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006237707566469908, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.11178077012300491, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10425307601690292, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.0967206284403801, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.09265588223934174, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.08295107632875443, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.048786357045173645, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04505667835474014, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.0575091578066349, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05310559272766113, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.049511246383190155, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04303246736526489, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.040870118886232376, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.029127297922968864, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0254229586571455, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.023469455540180206, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.022962447255849838, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01461401954293251, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.012117382138967514, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.011880860663950443, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.010958231054246426, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.010707611218094826, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.007702268194407225, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.007638872601091862, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00710499519482255, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004999073687940836, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10425307601690292, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2656323313713074, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.24726785719394684, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2414553016424179, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.21688275039196014, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.125079944729805, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.11802157014608383, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.14036759734153748, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.12888874113559723, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.12692034244537354, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11057297140359879, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10483909398317337, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07138271629810333, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0615575835108757, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05980585142970085, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05940965563058853, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0356285497546196, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.030215663835406303, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.030084798112511635, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02716098353266716, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.027069129049777985, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01818620227277279, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.017434263601899147, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.017444651573896408, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010330136865377426, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.21688275039196014, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2659868896007538, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.19422249495983124, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.16027885675430298, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.13208173215389252, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.11696569621562958, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.08397930860519409, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.15983061492443085, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.14318837225437164, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.12787382304668427, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.08040305972099304, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.07941560447216034, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0816534161567688, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.07015980780124664, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.05856656655669212, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.05548875778913498, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.04246271029114723, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03388041630387306, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.03291827067732811, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.027930377051234245, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.026172151789069176, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.025314606726169586, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.026415836066007614, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.021986959502100945, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.021392466500401497, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.19422249495983124, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2554471492767334, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23879782855510712, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23369868099689484, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2112339437007904, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12109006196260452, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11479821801185608, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13447314500808716, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12435928732156754, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12276236712932587, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.107839435338974, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10250934958457947, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06885913014411926, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0598403662443161, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.058363839983940125, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.058020204305648804, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0344952754676342, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.030564261600375175, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.030466729775071144, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02794220671057701, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02787657454609871, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01853204518556595, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01922778971493244, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.018059251829981804, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01371094211935997, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2112339437007904, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2821650803089142, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2638068199157715, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2582050859928131, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2332649528980255, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1334121823310852, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12645016610622406, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1480896919965744, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13698016107082367, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13525453209877014, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11875354498624802, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11273493617773056, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07554744929075241, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06557566672563553, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06394969671964645, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06356798112392426, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03776616230607033, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03274086117744446, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.032630085945129395, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.029731130227446556, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02966204658150673, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.019821876659989357, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019501319155097008, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019290650263428688, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012595162726938725, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1480896919965744, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.23952125012874603, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.21932601928710938, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.21132580935955048, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1894892305135727, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10885012149810791, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.10178772360086441, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12515170872211456, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.1153060719370842, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.11105497926473618, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09659302979707718, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.09125132858753204, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06314998865127563, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05474916845560074, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05196557939052582, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.05129551142454147, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.031684111803770065, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.026866205036640167, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.026612089946866035, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.024482565000653267, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02410057559609413, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.017031408846378326, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01668006367981434, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01616554893553257, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011196963489055634, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.21932601928710938, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1340569108724594, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12442409247159958, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12039308249950409, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10794553905725479, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06291808187961578, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.058774206787347794, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07200057804584503, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06662485003471375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0638727992773056, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.055529478937387466, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05266139656305313, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.036590319126844406, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03195696696639061, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.030298708006739616, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.029899273067712784, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.018339546397328377, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.015812745317816734, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.015642451122403145, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.014344246126711369, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.014177250675857067, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009723340161144733, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.010014034807682037, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009203687310218811, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006943693850189447, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1340569108724594, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12152020633220673, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.11276325583457947, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.10855972766876221, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09726987779140472, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05688753351569176, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.05282868072390556, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.0662280023097992, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06126542389392853, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05777600035071373, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.05025118589401245, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04778993874788284, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.033551737666130066, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.029318097978830338, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.027371928095817566, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.026898980140686035, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01681024208664894, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.014166727662086487, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.013938849791884422, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.012829618528485298, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.012594960629940033, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008862455375492573, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008881027810275555, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.008254429325461388, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005879545584321022, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12152020633220673, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.27554380893707275, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.256550669670105, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.25062987208366394, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2253330498933792, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.12987278401851654, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12264930456876755, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1454959660768509, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.13372591137886047, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.13171665370464325, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11489308625459671, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10894573479890823, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07395066320896149, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06385640054941177, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06206950172781944, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06166321784257889, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03690263628959656, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0313691571354866, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.031231682747602463, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.028220975771546364, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.028134271502494812, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.018834931775927544, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01806788705289364, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.018088623881340027, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.01071413978934288, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2253330498933792, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.30611056089401245, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.22303439676761627, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1906818449497223, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1502312272787094, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.13829322159290314, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.10281113535165787, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.17076188325881958, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.1566726416349411, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.14678950607776642, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.08509799838066101, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.08454323559999466, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.08753591030836105, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.07554181665182114, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.06725958734750748, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0651845633983612, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.04412645846605301, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03604019433259964, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.035403959453105927, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.026294859126210213, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.024906538426876068, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.024283602833747864, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.024761473760008812, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.021735291928052902, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.018024561926722527, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.22303439676761627, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2529388964176178, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23640818893909454, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23129591345787048, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20911964774131775, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.120034359395504, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11377380043268204, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13337980210781097, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12324900925159454, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12167814373970032, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10686423629522324, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10169702023267746, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06843505054712296, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05941855162382126, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.057966358959674835, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05761881545186043, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03434646874666214, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.030550653114914894, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.030457638204097748, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.027972491458058357, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.0279132928699255, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.018702905625104904, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01948951743543148, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.018252160400152206, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.014208485372364521, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20911964774131775, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2844710946083069, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.26595911383628845, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.26026782393455505, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23516392707824707, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1345256119966507, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12750741839408875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.149313822388649, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13814637064933777, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13637107610702515, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11971961706876755, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.1136772632598877, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07620196789503098, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0661090835928917, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06446663290262222, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06407763808965683, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03807897865772247, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03296360373497009, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.032853592187166214, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.029921550303697586, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02984628826379776, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01996823400259018, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01957327499985695, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01942933164536953, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01256227120757103, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.149313822388649, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.24286909401416779, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.22312413156032562, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.21534352004528046, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.19351379573345184, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.11051347851753235, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.10369085520505905, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12703584134578705, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11669175326824188, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.11263479292392731, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09846238046884537, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.09316097944974899, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06395663321018219, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05546827241778374, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.052818555384874344, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.052184563130140305, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03203637897968292, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.027424577623605728, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02718134969472885, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.025098029524087906, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.024736875668168068, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.017093930393457413, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.017156334593892097, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.016217928379774094, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011739727109670639, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.22312413156032562, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14107882976531982, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13104049861431122, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12688496708869934, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11384560912847519, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06626302003860474, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06196056306362152, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07558755576610565, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07004748284816742, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06723647564649582, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.058541301637887955, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05551297217607498, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03843022882938385, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03364194557070732, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03194326162338257, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.031535886228084564, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01927005872130394, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.016726721078157425, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.016553714871406555, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015198704786598682, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01503072865307331, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.010246576741337776, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01066307257860899, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009720196016132832, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.007499402854591608, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14107882976531982, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12614046037197113, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.11722465604543686, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.11301723122596741, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.10133300721645355, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05912699177861214, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0550711527466774, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06852778047323227, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06334145367145538, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.06002980098128319, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.05229621380567551, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04963938146829605, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.034709200263023376, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.030336735770106316, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0284501314163208, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.02797607146203518, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.017408035695552826, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.014678810723125935, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.014465322718024254, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.013288338668644428, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.01307175774127245, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.009147466160356998, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.009136918932199478, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.008551384322345257, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00600883224979043, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12614046037197113, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.28180503845214844, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.26276424527168274, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2567899823188782, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2313893884420395, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1329750269651413, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12569475173950195, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.14881519973278046, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.13686810433864594, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1348746418952942, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11788452416658401, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.11189824342727661, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07563142478466034, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06538660824298859, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06361047923564911, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06318913400173187, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03774060681462288, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.03214627504348755, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0320221446454525, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.028965044766664505, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02887570858001709, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.019285470247268677, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.018523847684264183, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.018537787720561028, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011012186296284199, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.14881519973278046, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2553505003452301, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1705743670463562, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1354922205209732, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1236933246254921, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.11309196054935455, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07475133240222931, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.14706526696681976, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.13223852217197418, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.12328384816646576, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07309263944625854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0723871961236, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07588134706020355, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06519556045532227, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.056645821779966354, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.054477304220199585, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.0394633449614048, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.032658785581588745, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.03213048726320267, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02613494172692299, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.024831021204590797, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.02382015623152256, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.02503788098692894, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0214137714356184, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.020557565614581108, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1705743670463562, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2575703263282776, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.24073252081871033, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23558303713798523, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21297380328178406, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12205357104539871, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.115725077688694, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13566911220550537, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12534165382385254, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12373678386211395, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10871057212352753, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10337254405021667, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06945638358592987, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06025565788149834, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05878088250756264, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0584293007850647, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03481723740696907, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.030701084062457085, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.030605372041463852, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.028048718348145485, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.027982531115412712, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01875969208776951, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019199272617697716, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.018294576555490494, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01356114074587822, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21297380328178406, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2869459390640259, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2682797610759735, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2625686526298523, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23729315400123596, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13568472862243652, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1286298632621765, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15059995651245117, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13929054141044617, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13754841685295105, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12080028653144836, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11468464881181717, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07684053480625153, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0666736289858818, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0650256797671318, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06463462859392166, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03838443383574486, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.033240485936403275, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03313063085079193, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.0301799513399601, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03010670654475689, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.0200949814170599, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019728941842913628, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01955489069223404, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.0126497196033597, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15059995651245117, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2452159970998764, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2252727746963501, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.21740995347499847, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.19551235437393188, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.11144351214170456, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.1045718789100647, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12748688459396362, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11776283383369446, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.11360833048820496, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.0994039848446846, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.09410571306943893, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06435032188892365, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05588264390826225, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05318111926317215, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.052532948553562164, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03212971240282059, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02747545950114727, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.027230989187955856, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02513613924384117, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.024772537872195244, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.016983991488814354, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.017011364921927452, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.016122903674840927, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011408185586333275, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2252727746963501, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.11962264031171799, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1109829917550087, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10702686011791229, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09599184989929199, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.05615751072764397, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05227900668978691, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06498333811759949, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.060058582574129105, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.05704500153660774, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0495820976793766, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.047099530696868896, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.033041033893823624, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02884768135845661, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.027120940387248993, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.026699889451265335, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01658996194601059, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.014230623841285706, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01404361892491579, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.012929562479257584, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.012735440395772457, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.008860426023602486, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009164059534668922, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008320881053805351, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006458536721765995, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.11962264031171799, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10860784351825714, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.10059644281864166, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.09623869508504868, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.08612743020057678, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05084303393959999, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04688944295048714, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.060488712042570114, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.055593304336071014, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05166461691260338, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.044860441237688065, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04281216487288475, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03070569597184658, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0267199520021677, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02459695190191269, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.02406560443341732, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.015425562858581543, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.01286841370165348, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.012610659003257751, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.011678512208163738, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.011410359293222427, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00822740700095892, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008343921974301338, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007565898820757866, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005759766325354576, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10860784351825714, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.28307366371154785, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.26346802711486816, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.25732558965682983, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.23124614357948303, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.13357658684253693, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12606528401374817, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.14947253465652466, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.13750304281711578, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.13558204472064972, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11805564910173416, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.11182694882154465, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07595299184322357, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06569357961416245, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06388013809919357, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06343664228916168, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.037893377244472504, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.03227008134126663, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.032139960676431656, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.029014049097895622, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.028900068253278732, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01936899684369564, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.018611382693052292, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.018629876896739006, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011052806861698627, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.14947253465652466, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2774761915206909, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.18475426733493805, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13696597516536713, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.13167178630828857, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.12079291045665741, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06749743223190308, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.16457456350326538, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.15164552628993988, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.13337929546833038, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.08039410412311554, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.08152227848768234, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.08442565053701401, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.07329593598842621, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.059352658689022064, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.05559411272406578, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.042561572045087814, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03264642879366875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.031408052891492844, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.025519810616970062, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.022995172068476677, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.02346123941242695, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.02440381608903408, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.019257057458162308, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01805141195654869, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.18475426733493805, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.26361992955207825, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2464163452386856, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.24115127325057983, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21786415576934814, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12509475648403168, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11858387291431427, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1388884037733078, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12847502529621124, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12682272493839264, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11137345433235168, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10590781271457672, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07115074247121811, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06185402348637581, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06032629311084747, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05996217578649521, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.035614389926195145, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03163915127515793, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03154297545552254, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.028933800756931305, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.028870444744825363, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.019143536686897278, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019981009885668755, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01866232044994831, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.014340304769575596, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21786415576934814, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2908748388290405, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.27194684743881226, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.26615455746650696, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.24039915204048157, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13757044076919556, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1303858906030655, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1528625786304474, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14127516746520996, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13948805630207062, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.122442826628685, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11624761670827866, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07793862372636795, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06762073189020157, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06593768298625946, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06553852558135986, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.039012305438518524, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03371772915124893, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03360431641340256, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03060179017484188, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03052535466849804, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.02051941119134426, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.02001616731286049, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019969241693615913, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012841520830988884, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1528625786304474, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2485259771347046, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2286168485879898, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.22073189914226532, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.19848580658435822, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.11304959654808044, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.10614676773548126, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12935784459114075, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11933376640081406, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.1151675432920456, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.10085458308458328, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.0954800695180893, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06520775705575943, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05659019574522972, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.053888436406850815, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.05323919653892517, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03252335265278816, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02773086540400982, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02747950702905655, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.025347664952278137, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.0249831210821867, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.017124127596616745, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.017004020512104034, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.016263753175735474, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011197810061275959, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2286168485879898, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13945090770721436, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1294739991426468, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12527398765087128, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11228960752487183, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06551901251077652, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.061207760125398636, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07514061778783798, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06937895715236664, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06652788072824478, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05783582106232643, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0548265166580677, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.038142576813697815, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.033367421478033066, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03161831945180893, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03118797205388546, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.019238395616412163, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.016585668548941612, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01640206389129162, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015071442350745201, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.014884710311889648, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.010351772420108318, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.010625091381371021, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009798443876206875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.007498880382627249, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13945090770721436, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12443304061889648, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1154666393995285, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.11107981204986572, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09944143891334534, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.058211177587509155, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0540955476462841, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06766099482774734, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06245967000722885, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.0590992346405983, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.051347363740205765, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.048820991069078445, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03428163751959801, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.029884593561291695, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.027970969676971436, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.02749437838792801, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.017183594405651093, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.014398741535842419, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.014179401099681854, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.013009187765419483, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.012788754887878895, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.009006018750369549, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00892904307693243, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.008401932194828987, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005807703360915184, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12443304061889648, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2758592665195465, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.25672462582588196, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2507592737674713, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2253512442111969, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.13010023534297943, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12274738401174545, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1457255482673645, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.13419951498508453, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.13205714523792267, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11500657349824905, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10909406840801239, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07403717190027237, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06406527012586594, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.062201857566833496, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06176234781742096, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.036964427679777145, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.031471800059080124, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03132452070713043, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02829301357269287, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.028199531137943268, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01891234517097473, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.018208486959338188, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0181793961673975, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.01086917333304882, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2253512442111969, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.30579864978790283, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.21675466001033783, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.18417087197303772, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1541445553302765, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.13806352019309998, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.10253098607063293, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.16940633952617645, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.1538388878107071, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.14649339020252228, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.08798587322235107, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.086763396859169, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.08753838390111923, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.07720664143562317, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.07009109854698181, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.06845080852508545, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.04643021523952484, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.04151470959186554, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.041154757142066956, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.033624161034822464, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.03261477127671242, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.028961181640625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.03224349766969681, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.027173129841685295, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.027580896392464638, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.21675466001033783, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.26502278447151184, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.24773766100406647, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2423686683177948, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2190198302268982, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12581583857536316, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11925297975540161, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13972711563110352, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12919123470783234, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1275617927312851, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11197993159294128, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10647975653409958, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07163514196872711, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.062230031937360764, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06070021912455559, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06033579632639885, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.035866741091012955, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03187908977270126, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.031777024269104004, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.029152343049645424, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02908368594944477, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.019331613555550575, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020175807178020477, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.018852876499295235, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.014527779072523117, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2190198302268982, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.29187461733818054, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2727918326854706, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.26695483922958374, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.24119381606578827, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13810908794403076, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.130900576710701, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15334142744541168, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1418146938085556, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14000898599624634, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12287942320108414, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11670314520597458, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07825257629156113, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0678957849740982, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06620477885007858, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06580496579408646, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03910962864756584, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03386871889233589, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03375466540455818, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030734507367014885, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.030658241361379623, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020502716302871704, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020132331177592278, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01995246671140194, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01294654794037342, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15334142744541168, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.247827410697937, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2273954302072525, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.21925456821918488, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.19730313122272491, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.11265178769826889, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.10551909357309341, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12898153066635132, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11925597488880157, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.11489010602235794, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.10041658580303192, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.09506997466087341, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06506814807653427, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.0566512793302536, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.0537545382976532, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.0530676543712616, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03253653272986412, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.027753610163927078, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.027496596798300743, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02537759765982628, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.024990202859044075, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01723557710647583, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01718248426914215, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.016344130039215088, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011477475985884666, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2273954302072525, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13831062614917755, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12832611799240112, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12395374476909637, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11124911159276962, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06498207896947861, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06060791015625, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07491924613714218, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06907489895820618, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06597159802913666, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05734013020992279, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05442451685667038, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03806985542178154, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03309813141822815, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.031268294900655746, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03082296811044216, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.019081195816397667, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.016224782913923264, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01602991484105587, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01468203216791153, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.014486417174339294, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01007378101348877, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01016521267592907, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009483933448791504, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0068788970820605755, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13831062614917755, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12665581703186035, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.11757539957761765, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1131669282913208, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.10138321667909622, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05942104756832123, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.05517183244228363, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06906282156705856, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06391678005456924, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.06034793332219124, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.05242789164185524, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04971090331673622, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.035032302141189575, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.03061862289905548, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.028590813279151917, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.028104085475206375, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.017556799575686455, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.014762083999812603, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.014533900655806065, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.013336772099137306, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.01310691051185131, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.009240829385817051, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.009232315234839916, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.008608099073171616, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.006097679026424885, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12665581703186035, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2677268981933594, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.249052494764328, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.24325914680957794, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2187049239873886, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.12612800300121307, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.11898089200258255, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1412910670042038, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.13007858395576477, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.12792930006980896, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11153297126293182, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10568650811910629, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0717906504869461, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.062144603580236435, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06029503792524338, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05985645949840546, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03583898767828941, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.03048226796090603, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.030350446701049805, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.027434471994638443, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.027321668341755867, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.018323654308915138, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01763901673257351, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.017576823011040688, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010519999079406261, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2187049239873886, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2936614751815796, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.2177141159772873, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.18723313510417938, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.14588730037212372, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.12986786663532257, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.09685725718736649, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1652350276708603, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.15153725445270538, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.14100764691829681, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0875757485628128, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.08294782042503357, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.08492077887058258, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.07339796423912048, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.06361321359872818, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.061115290969610214, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.04281684756278992, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.034850794821977615, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.03416726365685463, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.027319658547639847, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.025810230523347855, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0237610824406147, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0250997357070446, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.020800301805138588, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01900768280029297, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.2177141159772873, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2672123610973358, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.24965757131576538, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.24428986012935638, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22065402567386627, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12679582834243774, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12014872580766678, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14097492396831512, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13022418320178986, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12856945395469666, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11280659586191177, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10724622756242752, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07217288762331009, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06261692941188812, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06105982884764671, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.060694172978401184, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0361475870013237, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.031867045909166336, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.031764715909957886, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02908460795879364, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.029012925922870636, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01940252259373665, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.0198929812759161, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01891092024743557, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.014012830331921577, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22065402567386627, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2925242781639099, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2733314037322998, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.26748329401016235, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.24160577356815338, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13846825063228607, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1312023550271988, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1537272334098816, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14223341643810272, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1404036283493042, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12317220866680145, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.1168980747461319, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07844168692827225, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06808771938085556, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06638367474079132, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06598120927810669, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.039200883358716965, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03396637737751007, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03385128825902939, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030816344544291496, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.030739914625883102, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020534656941890717, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020199274644255638, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019976045936346054, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013006127439439297, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1537272334098816, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.24317768216133118, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2230856865644455, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.2150769680738449, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.19342990219593048, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.1105867326259613, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.1035800352692604, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12689168751239777, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11708278954029083, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.11276500672101974, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09848577529191971, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.09323444962501526, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06404069066047668, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05556819215416908, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.052792955189943314, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.05212267115712166, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03203423693776131, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.027283338829874992, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.027029521763324738, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.024932848289608955, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02455485425889492, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01704126037657261, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.016921021044254303, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.016159962862730026, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011349450796842575, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2230856865644455, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14788633584976196, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13719572126865387, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13285629451274872, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11906485259532928, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06941895931959152, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06491546332836151, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07915639877319336, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07328212261199951, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07050170749425888, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.061207693070173264, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05796170234680176, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.040226757526397705, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03512775897979736, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03339104354381561, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.032973382622003555, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.020145848393440247, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0173513051122427, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01717708632349968, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015706224367022514, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015536244958639145, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.010626603849232197, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01087261363863945, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.010066370479762554, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.007425976451486349, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14788633584976196, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1374424695968628, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.12738509476184845, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.12299930304288864, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.11012087017297745, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.06431834399700165, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.059944894164800644, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.07419756799936295, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06855440884828568, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.06537427008152008, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.05668738856911659, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.05375910922884941, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.037588126957416534, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.03282807767391205, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.030965061858296394, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.030512889847159386, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.018856417387723923, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.01604378968477249, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.015844501554965973, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.014501281082630157, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.014304202049970627, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.009918775409460068, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.010032891295850277, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.009312685579061508, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0067406315356493, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1374424695968628, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2675391435623169, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.24900783598423004, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2430053949356079, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.21826010942459106, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1259700059890747, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.11880539357662201, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1411939263343811, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.12999534606933594, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.12784472107887268, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11129657179117203, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.1054440438747406, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07170658558607101, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.062071770429611206, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06026048958301544, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05982370302081108, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.035799093544483185, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.030514584854245186, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03038143552839756, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0274569820612669, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.027345698326826096, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.018322795629501343, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.017711561173200607, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01760363020002842, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010658808052539825, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.21826010942459106, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.27725887298583984, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.21332654356956482, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.19021663069725037, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.15481533110141754, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.12565070390701294, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.09923103451728821, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.15396158397197723, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.14001117646694183, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.133036807179451, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.08719588816165924, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.08152997493743896, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0793895497918129, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06806232035160065, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.06162275746464729, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0600120946764946, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.04044834151864052, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03381667286157608, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.033382389694452286, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.026933662593364716, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.026047848165035248, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.02312924899160862, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.023943329229950905, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.021248353645205498, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01853298768401146, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.21332654356956482, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.25632286071777344, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2395755499601364, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23445162177085876, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21187257766723633, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12161879986524582, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11526893079280853, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13510987162590027, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12486980855464935, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12328137457370758, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10826514661312103, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10297597944736481, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0692049115896225, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.060103438794612885, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.058619290590286255, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05827052891254425, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.034669533371925354, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.030734430998563766, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.030634934082627296, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.028095779940485954, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.028031200170516968, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.018683262169361115, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019371381029486656, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.018214277923107147, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01386470627039671, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21187257766723633, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.28323519229888916, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.26475200057029724, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.25912684202194214, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23409026861190796, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13391561806201935, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12692157924175262, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14850449562072754, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13750821352005005, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1357753574848175, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11915349960327148, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11315841227769852, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07574596256017685, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06581300497055054, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06417526304721832, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06378694623708725, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03782093524932861, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03282271698117256, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03271133825182915, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02979211136698723, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.029720239341259003, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.019743384793400764, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019497260451316833, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019204674288630486, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012527373619377613, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14850449562072754, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2280980497598648, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2086441069841385, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.20054090023040771, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18040630221366882, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10370409488677979, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09667523205280304, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11991757154464722, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11055529862642288, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10589198768138885, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09213243424892426, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08728190511465073, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.060542408376932144, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.052566707134246826, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04960228502750397, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04888509213924408, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.0302948709577322, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.025796743109822273, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.025520065799355507, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.023566048592329025, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02315812185406685, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01616002805531025, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.0162822175770998, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.015231105498969555, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011165806092321873, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2280980497598648, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.11568798124790192, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.10732030868530273, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10310257971286774, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09242037683725357, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.054205745458602905, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.050212014466524124, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06328360736370087, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.05863196775317192, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.05507786199450493, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.047831714153289795, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04546744003891945, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03214561939239502, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02809835970401764, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.026112686842679977, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.025629449635744095, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.016109827905893326, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.013614466413855553, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.013389742001891136, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.012347998097538948, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01211164053529501, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.008522283285856247, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.008704722858965397, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.007896147668361664, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.005957149434834719, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.11568798124790192, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0987205058336258, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.09146259725093842, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.08713667094707489, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.07805074751377106, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.046018391847610474, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.042222339659929276, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.05500894784927368, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05096889287233353, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.04677238687872887, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.040593698620796204, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.038626041263341904, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02782534807920456, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.024376878514885902, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02214774116873741, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.021603424102067947, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.013957589864730835, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.011485019698739052, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.011200797744095325, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.010390429757535458, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.01009239349514246, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.007375078275799751, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.007353407796472311, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.006684960797429085, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004834748804569244, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0987205058336258, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.24252849817276, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.22582879662513733, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22047604620456696, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1981307864189148, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11398115754127502, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10748152434825897, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12844781577587128, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.1179104670882225, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11567587405443192, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10086436569690704, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09579383581876755, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06526556611061096, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05629651993513107, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.054486941546201706, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05405554175376892, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0325598269701004, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02757185697555542, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0274291280657053, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02483557164669037, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02471221052110195, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016640614718198776, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.016008997336030006, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01589701697230339, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00956620741635561, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.22582879662513733, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.25947853922843933, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.19517076015472412, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1725207418203354, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.14305438101291656, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.11620049178600311, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0884573757648468, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.14181387424468994, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.1299063116312027, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.12364647537469864, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.08034428209066391, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.07500993460416794, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07289226353168488, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06284771859645844, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.05672108381986618, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.05517863854765892, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03684245049953461, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03086898848414421, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0304806437343359, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02440236508846283, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.02353052794933319, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.020659342408180237, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.02161705121397972, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.018831051886081696, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.016441570594906807, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.19517076015472412, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24700897932052612, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23089608550071716, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22593729197978973, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20423950254917145, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1171722486615181, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1110786572098732, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13008461892604828, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12033793330192566, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11879128962755203, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10432695597410202, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09919919073581696, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0666479617357254, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.057941582053899765, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.056509826332330704, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05617509409785271, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03339480608701706, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.029657255858182907, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.02956303581595421, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.027118561789393425, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.027059612795710564, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017985956743359566, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01873650774359703, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.0175352580845356, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013457964174449444, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23089608550071716, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2802686393260956, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.26198071241378784, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2564009428024292, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23172253370285034, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13251884281635284, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12560397386550903, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14694945514202118, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13605095446109772, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13433751463890076, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11796031892299652, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11199040710926056, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07487457245588303, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0651126578450203, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0634985938668251, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06311448663473129, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.037405673414468765, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03246142342686653, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03235144540667534, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02946109138429165, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.029390383511781693, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01948383077979088, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01925845630466938, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.018956929445266724, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012338778004050255, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14694945514202118, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.23200096189975739, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2123870998620987, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.20440791547298431, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18404971063137054, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10535749793052673, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09836411476135254, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12135075777769089, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11206945776939392, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10761859267950058, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09375923871994019, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08890493214130402, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.061290666460990906, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.0532241128385067, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.050299517810344696, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04960007593035698, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.030667679384350777, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.026003314182162285, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.025738544762134552, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.023748477920889854, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.023351583629846573, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.016339298337697983, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01618184521794319, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.015440456569194794, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010847571305930614, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2123870998620987, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12426842749118805, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11540208011865616, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11145596206188202, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0999058187007904, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.058297477662563324, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05435476079583168, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06704983860254288, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06210211664438248, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0591820552945137, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05146344006061554, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.048806462436914444, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03408237546682358, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02979031205177307, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.02808683179318905, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.02767256833612919, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.017077308148145676, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.014657808467745781, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.014477277174592018, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.013299447484314442, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013115744106471539, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009049242362380028, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009316797368228436, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008516165427863598, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006451418623328209, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12426842749118805, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11622928828001022, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.10787627100944519, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1037670448422432, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.0928458571434021, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05437593162059784, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0504327118396759, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06350494921207428, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05862606316804886, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.0552411749958992, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04795987904071808, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04555966332554817, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03217598795890808, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.028114106506109238, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02617746964097023, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.025709183886647224, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.016148265451192856, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.013570653274655342, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.013352970592677593, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.012271133251488209, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.012042423710227013, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008535563014447689, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008551063016057014, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00794242974370718, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005706605967134237, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11622928828001022, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2536470592021942, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.23610356450080872, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.23048336803913116, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20731118321418762, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11938505619764328, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.1126294806599617, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13390766084194183, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.1232461929321289, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.12116515636444092, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10557235777378082, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10009904950857162, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0680592954158783, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.058869581669569016, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.057094015181064606, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05667097866535187, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03395261615514755, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.028898710384964943, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02875870279967785, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.026026474311947823, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02590746246278286, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01738208532333374, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.016778435558080673, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.016665590927004814, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010098263621330261, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.23048336803913116, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2835422158241272, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.2078813910484314, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.18008342385292053, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.14193882048130035, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.12748998403549194, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.09339433163404465, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.15662774443626404, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.14367590844631195, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.13624702394008636, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0836782157421112, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.07887528091669083, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.08078300952911377, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0698305070400238, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.06259147822856903, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.06079985573887825, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.04096570983529091, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03437354415655136, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.033904753625392914, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.026598535478115082, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.025555575266480446, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.022966820746660233, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.024460729211568832, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.02082720957696438, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.018881279975175858, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.2078813910484314, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24654927849769592, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2304156869649887, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22549284994602203, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2037825882434845, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11685892194509506, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11076372116804123, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12975110113620758, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12002500146627426, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11848951131105423, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.1040334552526474, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0989072322845459, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06634103506803513, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0576561763882637, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.056225452572107315, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05588671192526817, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03319450095295906, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02926645614206791, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.029173409566283226, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026701679453253746, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026637185364961624, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017678504809737206, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.018160028383135796, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017218196764588356, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01266408059746027, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2304156869649887, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2832378149032593, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2647838294506073, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.25914430618286133, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23415550589561462, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13397608697414398, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12698781490325928, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14864735305309296, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1375909447669983, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1358192414045334, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11923887580633163, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11319082975387573, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0757560059428215, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06584269553422928, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06420093029737473, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0638161152601242, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03788476437330246, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.032830338925123215, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03271833062171936, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.029795778915286064, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.029720772057771683, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01982446201145649, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019487783312797546, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01928667724132538, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01249539665877819, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14864735305309296, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2320651262998581, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2127905786037445, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.20460551977157593, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18444660305976868, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10536551475524902, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09840016812086105, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12194602936506271, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11244024336338043, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.1076052114367485, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09400280565023422, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08921106904745102, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06141867861151695, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05334227904677391, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.050295375287532806, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04956283047795296, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.030750149860978127, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.025960521772503853, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02567191608250141, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02374110370874405, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02331480383872986, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.016370680183172226, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.016120733693242073, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01541826594620943, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.01070405449718237, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2127905786037445, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13013866543769836, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12068702280521393, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11654236167669296, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10450854897499084, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06097430735826492, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05685138702392578, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06999528408050537, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06486444175243378, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06192079186439514, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.053768280893564224, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05098605155944824, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03559023514389992, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.031116340309381485, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.029354717582464218, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.028927050530910492, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01784696988761425, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.015305568464100361, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01511620543897152, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.013872439973056316, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013685942627489567, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009456238709390163, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009705194272100925, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008902417495846748, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0066938335075974464, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13013866543769836, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10734237730503082, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.09943441301584244, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.09548183530569077, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.08551248908042908, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.0500924289226532, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04643293470144272, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.05847129598259926, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05416448414325714, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.0508517324924469, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.044121138751506805, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.041864532977342606, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.029590873047709465, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.025889798998832703, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02407599613070488, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.02363520860671997, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.014841265045106411, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.012428886257112026, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.01222185604274273, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.011221589520573616, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.011004392988979816, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.007820270024240017, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.007786741480231285, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007261981256306171, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005129451397806406, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10734237730503082, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.24312838912010193, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.22621816396713257, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2208913266658783, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19841766357421875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11414588242769241, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10764910280704498, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12817424535751343, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11774799227714539, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1157844215631485, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10094526410102844, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0956987664103508, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06505414098501205, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.056219086050987244, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05454922094941139, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.054139237850904465, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03245547041296959, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02760106325149536, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.027469472959637642, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.024846581742167473, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.024738265201449394, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0166203323751688, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.016008196398615837, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015939051285386086, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00959713663905859, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.22621816396713257, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.27628323435783386, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.21616393327713013, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.19554615020751953, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.146502286195755, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.12564267218112946, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.1017644926905632, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.15033304691314697, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.13724380731582642, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.1318623125553131, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.08565302938222885, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.07553927600383759, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07705651223659515, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06642267107963562, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.061174020171165466, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.059885554015636444, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.039245620369911194, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.033076681196689606, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.032749757170677185, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.025783265009522438, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.025163287296891212, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.02232663705945015, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.02270588092505932, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.020845726132392883, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01720030978322029, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.21616393327713013, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24238117039203644, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22667229175567627, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22186075150966644, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20058932900428772, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11496250331401825, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1090143695473671, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1275954395532608, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11806325614452362, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11652930080890656, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10242567211389542, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0974050760269165, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06530175358057022, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05682126432657242, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05542166531085968, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.055088791996240616, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.032730717211961746, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.029053736478090286, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.0289612989872694, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026575729250907898, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.026510490104556084, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017598528414964676, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.018310369923710823, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017156394198536873, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013100466690957546, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.22667229175567627, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.28081709146499634, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2626214325428009, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2570517361164093, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23247212171554565, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1327953189611435, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12590985000133514, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1472049504518509, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13632863759994507, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.134606271982193, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11830146610736847, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11228321492671967, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0750487819314003, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0652347132563591, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06361857056617737, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06324198842048645, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03747246414422989, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03250676766037941, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.032398298382759094, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02951962687075138, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02944828011095524, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.019502444192767143, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019255319610238075, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.018969204276800156, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01230243407189846, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1472049504518509, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2282731831073761, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20936034619808197, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.20118893682956696, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1812945455312729, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10379814356565475, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09689691662788391, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12082905322313309, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11096462607383728, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.1059848964214325, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09261497110128403, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08789454400539398, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06074113771319389, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05282839015126228, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.049742020666599274, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04899225011467934, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.030343521386384964, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.0260306429117918, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.025729404762387276, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.0238900575786829, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.023461002856492996, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.016081981360912323, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.016655156388878822, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.015073515474796295, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011654604226350784, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2282731831073761, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14299306273460388, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13283638656139374, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12888206541538239, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11564742028713226, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06708802282810211, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06287147849798203, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07618505507707596, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07052276283502579, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06809726357460022, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.059246938675642014, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05614869296550751, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.038665253669023514, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03374588489532471, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03221328929066658, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.031840939074754715, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.019341958686709404, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.016631951555609703, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.016483616083860397, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01504773460328579, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0149071691557765, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.010131017304956913, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.010254236869513988, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009634814225137234, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006843231618404388, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14299306273460388, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12730054557323456, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.11815251410007477, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.11411894112825394, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.10234132409095764, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05959421768784523, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.05557948350906372, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06856440752744675, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.0634765774011612, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.06048323214054108, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.052584223449230194, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04988910257816315, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.034735843539237976, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.030324753373861313, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.028596144169569016, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.028189420700073242, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01740558259189129, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.014712601900100708, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.014530348591506481, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.013294628821313381, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.013105012476444244, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00913542415946722, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.009062652476131916, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00859024003148079, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005896611604839563, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12730054557323456, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.25030267238616943, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.23302388191223145, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22761964797973633, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20471514761447906, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11748657375574112, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.11093059182167053, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13145852088928223, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.12111362814903259, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1191847175359726, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10392124950885773, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09858737140893936, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06674183905124664, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0578145869076252, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.056134916841983795, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05573597922921181, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03327071666717529, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.028407247737050056, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.028278129175305367, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02558269537985325, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02548033557832241, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.017038319259881973, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.016449764370918274, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.016399314627051353, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009856735356152058, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22761964797973633, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2710350453853607, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.20790497958660126, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.18508417904376984, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.14145365357398987, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.12215667217969894, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0972161814570427, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1489601582288742, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.13664594292640686, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.1294213980436325, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.08001968264579773, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.07606444507837296, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07645872980356216, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06593827903270721, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.05942456051707268, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.057799555361270905, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03849580138921738, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03196223825216293, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.03150694817304611, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0242628026753664, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.023380886763334274, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.02121865376830101, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.02199840173125267, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.019236598163843155, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.016306592151522636, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.20790497958660126, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.24486473202705383, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2291456013917923, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.224327951669693, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20296277105808258, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11610792577266693, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11018591374158859, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1288606971502304, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.11919306963682175, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11769550293684006, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10356538742780685, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.0984947606921196, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06598028540611267, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05735992640256882, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.055966317653656006, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05564068257808685, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.033039629459381104, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.029306134209036827, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.029214465990662575, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.026822810992598534, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02676679939031601, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017733093351125717, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01843811571598053, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01728622056543827, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013157116249203682, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2291456013917923, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.28030043840408325, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.26225993037223816, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2567852735519409, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23224809765815735, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13250410556793213, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12571458518505096, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14694508910179138, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13600659370422363, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1343243271112442, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.1181061714887619, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11221926659345627, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07489342242479324, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06507191061973572, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06347404420375824, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06309661269187927, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03740653768181801, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.032415952533483505, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03230803832411766, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.029448948800563812, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.029377158731222153, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01949027180671692, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019181087613105774, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.018963919952511787, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012216879054903984, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14694508910179138, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2255559265613556, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20653875172138214, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19819162786006927, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17861883342266083, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10249326378107071, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09547100961208344, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11921118944883347, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10990240424871445, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10475090146064758, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09131801873445511, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08667753636837006, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06021161749958992, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.052212394773960114, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.049006663262844086, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.048235807567834854, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.030126437544822693, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02544815093278885, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.025140343233942986, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.023283129557967186, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02283615991473198, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.016104502603411674, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01604778692126274, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01510118879377842, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010898330248892307, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2255559265613556, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14248791337013245, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13223299384117126, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12818007171154022, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11498220264911652, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06667035818099976, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.062405459582805634, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07614725083112717, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07019123435020447, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06766257435083389, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05885719880461693, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05588396266102791, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03875608369708061, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03369631618261337, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.032104916870594025, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03172444552183151, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01943480409681797, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.016785288229584694, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.016630079597234726, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015246315859258175, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015093931928277016, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.010353379882872105, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.010654478333890438, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009808079339563847, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.007475011050701141, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14248791337013245, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12433380633592606, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.11530546844005585, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.11123766005039215, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09960940480232239, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.057925350964069366, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.05394970253109932, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06679553538560867, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06187351793050766, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05880560353398323, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.051120683550834656, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04845542460680008, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.033816006034612656, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.029550977051258087, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.027779292315244675, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.027341501787304878, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.016919588670134544, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.014258528128266335, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.014063498936593533, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.012865444645285606, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.012667469680309296, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008824371732771397, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008749542757868767, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.008257742039859295, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0056284223683178425, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12433380633592606, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.23471014201641083, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.218333899974823, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.21295486390590668, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19154873490333557, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11014126241207123, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10390235483646393, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12339422106742859, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11373475193977356, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1117348000407219, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09743442386388779, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0923595055937767, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0625864565372467, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05428826063871384, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05261692404747009, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.052222032099962234, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03123614937067032, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.026614774018526077, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.026486629620194435, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02396852895617485, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02385602332651615, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01596432738006115, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015433424152433872, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015333433635532856, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0092361094430089, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.218333899974823, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2919869124889374, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.24052096903324127, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.22220230102539062, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1712617427110672, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.13275961577892303, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.11285990476608276, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.161896213889122, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.14674372971057892, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.13985534012317657, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0972733125090599, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.08841083198785782, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.08360940217971802, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.07128554582595825, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.06506051868200302, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.06353361904621124, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.04251469671726227, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03570381924510002, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.03528302162885666, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.029290981590747833, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.02858247235417366, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.024336621165275574, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.02517612837255001, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.02244543470442295, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01956157013773918, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.22220230102539062, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.250793993473053, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23476900160312653, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2298612892627716, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20805475115776062, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11889483034610748, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11282661557197571, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13192711770534515, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12203438580036163, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12047871947288513, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10609864443540573, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10101723670959473, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0675131231546402, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05872787535190582, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05729833245277405, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05696391314268112, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03381172940135002, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.030016619712114334, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.029925575479865074, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02748396061360836, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.027423491701483727, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.018133290112018585, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.018896115943789482, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017683006823062897, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013501743786036968, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2298612892627716, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2784423530101776, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.26056644320487976, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2551288604736328, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23085005581378937, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.131581649184227, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12483824044466019, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1458498239517212, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13504701852798462, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1333533674478531, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11732707172632217, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11152273416519165, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07432755082845688, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06461215764284134, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06302408874034882, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06264785677194595, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03713329881429672, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03221861273050308, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03210846334695816, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.029282843694090843, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02921253629028797, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.019377801567316055, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01910359598696232, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.018865875899791718, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012233900837600231, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23085005581378937, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.22776876389980316, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2076021432876587, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.1985751986503601, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17849324643611908, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10336470603942871, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09577035158872604, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12129203230142593, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.1114736795425415, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10574975609779358, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09158694744110107, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08690065890550613, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.061161451041698456, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.052996501326560974, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04947897419333458, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.048623230308294296, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.030716191977262497, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.025777542963624, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02543417364358902, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.023505298420786858, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.023007787764072418, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.016603145748376846, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.016421224921941757, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.015517544001340866, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011269609443843365, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.22776876389980316, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14751708507537842, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13700158894062042, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1330306977033615, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11938732862472534, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06887733936309814, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06460747867822647, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07795092463493347, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07219759374856949, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.069938525557518, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.060880109667778015, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05771083012223244, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03958653286099434, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.034542277455329895, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03304344415664673, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03268523886799812, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01979188621044159, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01708119548857212, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.016937386244535446, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015474250540137291, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015337381511926651, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.010369348339736462, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.010544152930378914, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00987656507641077, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.007072786800563335, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14751708507537842, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.13328935205936432, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.12367378175258636, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.11966650933027267, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.10727014392614365, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.06203104555606842, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.057970549911260605, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.07102852314710617, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06577034294605255, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.0629890114068985, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.05479142442345619, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.05189800262451172, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.035926416516304016, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.03137190267443657, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.029727386310696602, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.029316818341612816, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01796729490160942, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.015240260399878025, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.015062793157994747, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.013766946271061897, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.01358978170901537, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.009360719472169876, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.009289290755987167, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00883588008582592, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005960908252745867, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.13328935205936432, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.22681035101413727, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.21113350987434387, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.20607011020183563, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18538019061088562, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10601415485143661, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10009023547172546, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11867373436689377, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10940322279930115, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10762261599302292, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09392411261796951, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08919602632522583, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06022283434867859, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05220633000135422, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.050639908760786057, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.050264354795217514, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.030060505494475365, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.025642450898885727, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.025516878813505173, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02311851643025875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02302374318242073, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015378640033304691, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014891831204295158, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014792034402489662, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008985258638858795, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.22681035101413727, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2644463777542114, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.21898162364959717, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.20361626148223877, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.16042447090148926, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.11984303593635559, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.10251228511333466, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.14372330904006958, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.13203880190849304, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.12611724436283112, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.09091566503047943, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.07901734113693237, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07394395023584366, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06375298649072647, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.058300379663705826, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0569729283452034, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03719508647918701, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03133390471339226, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.030974900349974632, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02593228593468666, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.025303689762949944, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.020531484857201576, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.02140180394053459, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.018904970958828926, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.015937676653265953, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.21898162364959717, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.25153177976608276, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23559267818927765, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23067441582679749, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20898418128490448, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11907979100942612, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11306894570589066, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1319901943206787, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1222120001912117, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12066798657178879, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10641257464885712, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10135526955127716, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06751463562250137, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.058767061680555344, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05734837055206299, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05701465159654617, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.033759620040655136, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02999318391084671, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.029898768290877342, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02747991308569908, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.027417151257395744, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.018003731966018677, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01881292089819908, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017543569207191467, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013372737914323807, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23067441582679749, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2705840468406677, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.25338488817214966, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.24815547466278076, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22469216585159302, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1277078539133072, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12125962972640991, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1415279358625412, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13108676671981812, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12942427396774292, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11406246572732925, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10849560052156448, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07212407141923904, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06267166137695312, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.061141274869441986, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.060781680047512054, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03599518910050392, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03122323751449585, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03111681342124939, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.028409883379936218, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.028342172503471375, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.018705984577536583, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.018468627706170082, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.018197594210505486, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011764667928218842, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22469216585159302, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.20926612615585327, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.18934273719787598, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.17995236814022064, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.16112402081489563, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.09495080262422562, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.08713138848543167, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11299262940883636, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.1037101149559021, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.097389817237854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08345489948987961, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.07920593023300171, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.05704191327095032, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04948554188013077, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04568004235625267, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.044763144105672836, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.028625497594475746, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024219315499067307, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.023840581998229027, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.022053807973861694, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.021517880260944366, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.015569763258099556, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.016052372753620148, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.014384660869836807, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011596428230404854, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.20926612615585327, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.40.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14580495655536652, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13534626364707947, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13150842487812042, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11810854077339172, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06795400381088257, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06383189558982849, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07735476642847061, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07121379673480988, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06904710829257965, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06020579859614372, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05725983530282974, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.039337869733572006, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03414541482925415, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03267164155840874, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.032317645847797394, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.019709698855876923, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.017006084322929382, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.016871027648448944, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015462489798665047, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015330987982451916, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01043569203466177, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01066800206899643, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009921558201313019, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.007379280868917704, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14580495655536652, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.40.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12222354114055634, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.11344251036643982, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1094638779759407, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09822418540716171, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.056794773787260056, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.05297382175922394, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06547004729509354, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06070466712117195, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05769546329975128, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.05027797818183899, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04777975752949715, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03316644951701164, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.028997056186199188, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.027240606024861336, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.026802675798535347, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.016614602878689766, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.014031696133315563, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.013835227116942406, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.01271106954663992, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.01251787319779396, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008728155866265297, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008673167787492275, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.008182484656572342, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0056653995998203754, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12222354114055634, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.40.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.24892280995845795, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.23203377425670624, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22670766711235046, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20431219041347504, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11679470539093018, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.11044111847877502, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13073855638504028, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.12030360847711563, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11848297715187073, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10364509373903275, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09844326972961426, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06636529415845871, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05742332339286804, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.055804766714572906, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05542076751589775, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03312407433986664, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02825193665921688, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.028130922466516495, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02550404891371727, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.025412706658244133, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.016945693641901016, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01636497490108013, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.016303105279803276, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009839273057878017, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22670766711235046, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.40.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.28027215600013733, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.228052020072937, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.20960435271263123, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1695196032524109, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.1270705610513687, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.10596445202827454, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1526547223329544, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.1406603753566742, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.13403566181659698, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.09408992528915405, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.08612775802612305, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07831238955259323, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06770134717226028, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.06155407056212425, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.06003095954656601, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03927742689847946, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03266192600131035, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.03223396837711334, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02638060785830021, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.025556497275829315, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.021166305989027023, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.021851472556591034, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.019246136769652367, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.015676768496632576, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.228052020072937, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.40.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2531546354293823, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23703183233737946, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23204763233661652, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21026714146137238, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11981599032878876, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11377349495887756, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13291849195957184, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12295283377170563, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12141868472099304, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.1070338562130928, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10200145095586777, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06796209514141083, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.059104107320308685, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05769281089305878, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.057359255850315094, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.034037962555885315, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.030171487480401993, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.030074071139097214, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.0276382714509964, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.027576331049203873, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.018255647271871567, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01891319639980793, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017804039642214775, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013430768623948097, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21026714146137238, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.40.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.27241796255111694, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.25507673621177673, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.24982643127441406, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22624754905700684, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12859541177749634, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1220913976430893, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14265531301498413, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13196538388729095, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1303357034921646, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11485353857278824, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10929612815380096, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07273120433092117, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06313362717628479, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06161653622984886, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06125664710998535, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03635155409574509, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.031536757946014404, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03143646568059921, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02871079184114933, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02864968031644821, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.019068488851189613, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.018770970404148102, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.018576057627797127, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012127104215323925, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22624754905700684, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.40.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.23020939528942108, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20940451323986053, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.20052184164524078, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1796354204416275, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10472462326288223, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09702413529157639, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.1219937652349472, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11227194964885712, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10709817707538605, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09232914447784424, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08738583326339722, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.061754997819662094, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05360452085733414, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05030246824026108, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04949484393000603, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03109915927052498, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.026537684723734856, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.026225456967949867, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02419162355363369, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.023737536743283272, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01697690784931183, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.017294466495513916, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.015975764021277428, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.012430491857230663, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.23020939528942108, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.41.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14378643035888672, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13331829011440277, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12920953333377838, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11596052348613739, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06695093214511871, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06271205097436905, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07678704708814621, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07051457464694977, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06804089993238449, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.059196747839450836, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05634251609444618, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03899703174829483, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.033846545964479446, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.032235607504844666, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.031843151897192, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.019601749256253242, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.016866905614733696, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01671150140464306, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015340209938585758, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015183459967374802, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.010485298000276089, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.010719742625951767, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00994456373155117, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0075266011990606785, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14378643035888672, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.41.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12997731566429138, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.12029165029525757, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.11580205708742142, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.10362321883440018, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.060307469218969345, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.056044258177280426, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.07002367079257965, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.0647725909948349, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.06133313104510307, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.05318534001708031, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.05046709626913071, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03542511537671089, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.03097117319703102, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.028982194140553474, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.028484052047133446, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01774686761200428, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.014974549412727356, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.014754882082343102, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.013546666130423546, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.013315560296177864, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0093199722468853, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00937048438936472, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.008683283813297749, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.006207508035004139, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12997731566429138, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.41.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2596971392631531, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.24199000000953674, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.23644179105758667, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.21303874254226685, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.12220146507024765, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.11542096734046936, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13676808774471283, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.1259632110595703, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1239711344242096, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10829708725214005, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10280506312847137, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06939584761857986, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06015399843454361, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05840589851140976, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.057988934218883514, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.034630946815013885, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.029544701799750328, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.029405713081359863, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.026637054979801178, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.026541244238615036, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.017715107649564743, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.017114564776420593, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.017041519284248352, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010235368274152279, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.21303874254226685, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.41.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2909308969974518, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.23792672157287598, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.22005099058151245, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.17453192174434662, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.1315358430147171, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.11057571321725845, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.16222333908081055, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.14420905709266663, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.13844089210033417, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.09792663156986237, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.08849333971738815, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.08279335498809814, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.07180193811655045, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.06627202033996582, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.06488846987485886, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.04412064701318741, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.038805752992630005, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.038501862436532974, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.033267125487327576, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.03273507207632065, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.027749858796596527, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.029600104317069054, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.02632913365960121, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.02513807639479637, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.22005099058151245, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.41.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.25093308091163635, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23488062620162964, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2299738973379135, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20816881954669952, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1187322586774826, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11270222812891006, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13168203830718994, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12184148281812668, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12033073604106903, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10598243772983551, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10085605084896088, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06731250137090683, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0585554838180542, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05714232847094536, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05680955946445465, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.033685408532619476, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.029822295531630516, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.029728885740041733, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.027287602424621582, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.027226483449339867, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.017978738993406296, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01861427165567875, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017531393095850945, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01312687061727047, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2299738973379135, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.41.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.27874597907066345, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.26090455055236816, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.25546351075172424, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23120245337486267, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13159888982772827, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12489287555217743, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14595161378383636, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13505354523658752, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13339835405349731, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11744509637355804, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.111670583486557, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07438577711582184, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06462974846363068, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0630572959780693, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06268177181482315, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0371701680123806, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03227977454662323, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03217272832989693, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.029368070885539055, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.029302004724740982, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01944948546588421, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019214238971471786, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.018941234797239304, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012409034185111523, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14595161378383636, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.41.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2320076823234558, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.21197375655174255, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.2031974196434021, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18268319964408875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10549513250589371, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.0980466827750206, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12277720868587494, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11316012591123581, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10779203474521637, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09355834126472473, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08876510709524155, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06206747889518738, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.053800057619810104, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05049540475010872, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.049691345542669296, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.031039616093039513, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.026276294142007828, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02595822513103485, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.023969540372490883, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02350897341966629, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.016546180471777916, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.016651364043354988, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.015495487488806248, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011413829401135445, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.21197375655174255, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.42.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14904163777828217, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13823851943016052, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13420931994915009, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.12060768902301788, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06951865553855896, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06517728418111801, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07908125221729279, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0728452280163765, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0706062987446785, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06147666275501251, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05841021239757538, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.040148936212062836, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03489861637353897, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.033393729478120804, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03303707391023636, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.020113639533519745, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.017314888536930084, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.017174219712615013, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015701105818152428, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015564842149615288, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01059875637292862, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.010766969993710518, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01010524295270443, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.007328456733375788, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14904163777828217, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.42.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12911677360534668, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.11964654922485352, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1155364140868187, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.10350938141345978, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.06000478193163872, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.05589614436030388, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06937192380428314, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.0640108734369278, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.06102639436721802, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.05300164595246315, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.05034656822681427, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0351661778986454, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.030636049807071686, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02882790006697178, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.02839195914566517, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.017635013908147812, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.01488744467496872, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.014691940508782864, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.013479427434504032, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.013273519463837147, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.009279696270823479, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.009261636063456535, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00870294589549303, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.006126256659626961, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12911677360534668, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.42.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2536456882953644, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.23651377856731415, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.23098446428775787, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20820602774620056, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11936045438051224, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.11275198310613632, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1332690566778183, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.12300557643175125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.12111327052116394, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10583540052175522, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10044156759977341, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06767801195383072, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05872708931565285, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05706150457262993, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05665924400091171, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03376353904604912, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.028891347348690033, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02876981720328331, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0260635856539011, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.025967245921492577, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.017284397035837173, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01677239127457142, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.016658740118145943, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.01011227909475565, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20820602774620056, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.42.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.28212395310401917, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.21521736681461334, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.19211390614509583, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.15460804104804993, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.1267917901277542, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.10001769661903381, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.15911974012851715, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.1411319077014923, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.13435472548007965, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.08790717273950577, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.08234783262014389, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.08137699961662292, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06920468062162399, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.06289313733577728, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.06132284551858902, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.04309644177556038, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.035624027252197266, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.03522573038935661, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.029015840962529182, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.02823226898908615, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.026965651661157608, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.026286838576197624, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.02532055415213108, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.021407265216112137, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.21521736681461334, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.42.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2558478116989136, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2395438700914383, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23449458181858063, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21226517856121063, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12103672325611115, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11487549543380737, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13427022099494934, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12421033531427383, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12264952063560486, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.1079920083284378, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10272187739610672, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06860720366239548, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05961868539452553, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.058181799948215485, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05784031003713608, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.034308429807424545, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03024553321301937, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.030147859826683998, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.027633504942059517, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.027574952691793442, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01825706474483013, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01870882324874401, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01779787801206112, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01298825815320015, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21226517856121063, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.42.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.28166210651397705, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.26364418864250183, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2581317722797394, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2335999608039856, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13297314941883087, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12619104981422424, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1473863571882248, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13645023107528687, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13474740087985992, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11861544847488403, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11273694783449173, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07510172575712204, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.065251424908638, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06366892904043198, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06329210102558136, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03750196844339371, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03251522779464722, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03240923583507538, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.029559699818491936, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02949288859963417, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01953285187482834, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01923663541674614, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019017457962036133, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012264447286725044, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1473863571882248, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.42.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.22856034338474274, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20874221622943878, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19968156516551971, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17960916459560394, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10387029498815536, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09632598608732224, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12155631184577942, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11208906769752502, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10623069852590561, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09214891493320465, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08748050779104233, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06138699874281883, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.053235094994306564, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04967587813735008, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04881918802857399, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.030685270205140114, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.025775808840990067, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.0254233255982399, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.023494139313697815, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.022992629557847977, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01626327820122242, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.016276463866233826, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.015119893476366997, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010981859639286995, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.22856034338474274, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.43.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13035757839679718, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12096105515956879, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11699054390192032, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10497356206178665, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.060707177966833115, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05671004578471184, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06973327696323395, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06439806520938873, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06164729595184326, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05365321785211563, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05104605108499527, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.035400472581386566, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.030876293778419495, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.029221534729003906, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.028826355934143066, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.017790766432881355, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.015283629298210144, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.015113629400730133, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.013894480653107166, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01372134406119585, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009514540433883667, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009738652966916561, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008996295742690563, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006807760335505009, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13035757839679718, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.43.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11342453956604004, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.10520387440919876, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.10094793885946274, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09043686091899872, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.0526815727353096, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04876343160867691, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06182555854320526, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05712844431400299, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.053529042750597, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04650842770934105, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04426867514848709, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03126382455229759, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02732154168188572, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.025292791426181793, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.02480509877204895, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.015664823353290558, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.01306484080851078, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.012833544984459877, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.011828924529254436, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.011581876315176487, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008286818861961365, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008215310983359814, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007663003169000149, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00540461391210556, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11342453956604004, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.43.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2279490977525711, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.21209295094013214, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.20696213841438293, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1862134039402008, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10663729161024094, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.1006087064743042, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11961746960878372, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11018374562263489, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10822124779224396, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09443694353103638, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08961542695760727, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06065467745065689, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05253595858812332, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0509149432182312, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.050561368465423584, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03025261126458645, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02578321471810341, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.025649825111031532, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.023233048617839813, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.023125246167182922, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015489363111555576, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014967603608965874, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014876706525683403, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00899460632354021, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2279490977525711, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.43.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.27639129757881165, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.2251071035861969, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.20593670010566711, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.15804769098758698, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.1265724152326584, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.10538583248853683, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.15283788740634918, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.1404043734073639, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.1324204057455063, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.09149843454360962, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.08013350516557693, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07850496470928192, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06766590476036072, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.06137131154537201, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.05981317535042763, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.039427269250154495, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03271031379699707, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.03219309449195862, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02609451673924923, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.025290390476584435, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.021395614370703697, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.022068742662668228, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.019421253353357315, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01599111594259739, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.2251071035861969, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.43.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2577454745769501, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.24129536747932434, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2362111657857895, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21381552517414093, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12196759134531021, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11574476957321167, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1353820264339447, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1252075433731079, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12359995394945145, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10884515941143036, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10357755422592163, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06917668879032135, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06011486053466797, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0586550235748291, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05830483138561249, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.034616775810718536, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.030513089150190353, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03041655942797661, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02789030596613884, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.027827970683574677, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.018434682860970497, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01890401542186737, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01796942576766014, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013157692737877369, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21381552517414093, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.43.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2777453362941742, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2600236237049103, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2545667290687561, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23049525916576385, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13120530545711517, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12450801581144333, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14548133313655853, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13467244803905487, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13297931849956512, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.1170622780919075, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11128906160593033, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07415127009153366, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06440093368291855, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06282950192689896, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06246120110154152, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03703755885362625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03211323171854019, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03200804442167282, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02920600026845932, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.029136618599295616, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.019322093576192856, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019038742408156395, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01880975440144539, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012188568711280823, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23049525916576385, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.43.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.22259965538978577, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.20271684229373932, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.19366908073425293, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.17385616898536682, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10121288895606995, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09364189952611923, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.11890622973442078, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.10937405377626419, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10361958295106888, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.08950264006853104, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08481597900390625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06003972887992859, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.052115298807621, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.04857955127954483, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04771014675498009, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.030162779614329338, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.025461522862315178, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.025124691426753998, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.023214077576994896, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.022726044058799744, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.016280945390462875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.016436316072940826, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.015174689702689648, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011531015858054161, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.22259965538978577, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.44.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1514178365468979, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14062097668647766, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13657481968402863, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.12277860939502716, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07054372876882553, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06620574742555618, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08011789619922638, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0738985538482666, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07160588353872299, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06242743879556656, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05931771546602249, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04070368781685829, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03541785478591919, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.0339069738984108, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.033553559333086014, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.020402558147907257, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.017656713724136353, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0175169724971056, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.016047483310103416, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015913471579551697, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.010808550752699375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.011073753237724304, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.010314292274415493, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.007662391755729914, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1514178365468979, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.44.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1220598816871643, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.11311434209346771, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.10914289951324463, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09781036525964737, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.056532297283411026, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.05256284773349762, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06550702452659607, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06058927997946739, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05741205811500549, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.049930639564991, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.047401312738657, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.033159881830215454, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.028954578563570976, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.027122503146529198, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.02667281962931156, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.016607077792286873, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.013939476571977139, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.013739983551204205, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.012607153505086899, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.012392898090183735, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008689936250448227, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008622126653790474, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00811703596264124, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00560436537489295, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1220598816871643, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.44.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.24744540452957153, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2306719273328781, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.22533723711967468, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20308150351047516, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11615359038114548, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10976562649011612, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13006962835788727, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11983472853899002, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11786037683486938, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10308758914470673, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09791883826255798, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06602086126804352, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05719782039523125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05549817904829979, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05511777848005295, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03293776139616966, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.028102703392505646, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02797982469201088, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.025372199714183807, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.025256413966417313, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01687467098236084, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.016325658187270164, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.016222745180130005, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.009826055727899075, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2306719273328781, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.44.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.28137555718421936, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.22051802277565002, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.19956813752651215, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1651187241077423, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.12710770964622498, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.10181543976068497, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.15533851087093353, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.13943181931972504, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.133610337972641, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.09088235348463058, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.08495137840509415, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07901560515165329, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06793547421693802, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.06242647394537926, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.061078887432813644, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.04101329669356346, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.034416716545820236, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.03405727073550224, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.027908815070986748, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.027198918163776398, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.024264594539999962, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0242998655885458, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.022770121693611145, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01902330294251442, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.22051802277565002, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.44.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.26362431049346924, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.24677512049674988, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.24160581827163696, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21867692470550537, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12458321452140808, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11824531108140945, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13799811899662018, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1278681457042694, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12627090513706207, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11116444319486618, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10560651123523712, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0703546330332756, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.061258070170879364, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05977373570203781, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05942142382264137, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03514706343412399, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.030864158645272255, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03076375275850296, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02814580872654915, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.028083158656954765, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.018477628007531166, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.018790874630212784, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01799730584025383, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012674011290073395, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21867692470550537, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.44.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.28562864661216736, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2673772871494293, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2618064284324646, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23686879873275757, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1348264217376709, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12795612215995789, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14928849041461945, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1383763551712036, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1366390734910965, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12026332318782806, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.1141839250922203, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07600494474172592, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06614499539136887, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06452895700931549, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06414639949798584, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03792427107691765, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.032888054847717285, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03277969732880592, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.029882241040468216, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.029811091721057892, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.019641771912574768, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019362425431609154, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01910788007080555, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012197834439575672, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14928849041461945, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.44.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.23353053629398346, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.21481695771217346, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.2055921107530594, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.18451669812202454, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10657508671283722, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.0991552546620369, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12480603158473969, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11518953740596771, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.1086188331246376, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.0948285311460495, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08984430134296417, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06294526904821396, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.054688822478055954, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.050959039479494095, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.05005447566509247, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03141069412231445, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02637534961104393, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.025985119864344597, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.024082573130726814, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.0235575120896101, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.016572924330830574, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.016594134271144867, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.015385124832391739, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011088138446211815, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.21481695771217346, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.45.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14932824671268463, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1386716067790985, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13433606922626495, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.12067066878080368, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06991344690322876, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0654100850224495, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07989861816167831, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07379528880119324, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07096181064844131, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.061821792274713516, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05874783918261528, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.040609873831272125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03538956120610237, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03364324942231178, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03321963548660278, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.020360413938760757, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.017541926354169846, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.017368506640195847, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015939105302095413, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015763480216264725, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.010796789079904556, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01108677126467228, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01023800764232874, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.007690511178225279, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14932824671268463, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.45.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.13335266709327698, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.12379960715770721, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.11913933604955673, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.10674311965703964, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.06233416125178337, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.057924702763557434, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.07312556356191635, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06711907684803009, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.06328965723514557, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.05512914061546326, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.052507318556308746, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.037117160856723785, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0321459025144577, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02998148649930954, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.029472259804606438, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.018626729026436806, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.015519937500357628, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.015278547070920467, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.014071271754801273, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.01381167396903038, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00988733395934105, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.009754491969943047, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00920943170785904, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.006482168100774288, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.13335266709327698, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.45.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.26142650842666626, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2435145229101181, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.23792794346809387, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.21415546536445618, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1228540912270546, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.11603304743766785, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13739055395126343, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.12661708891391754, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.12458613514900208, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10885128378868103, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10330260545015335, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06976262480020523, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.060443419963121414, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05870191007852554, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.0582856610417366, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03481264412403107, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.029709728434681892, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02958475798368454, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0268003698438406, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.026697201654314995, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01783604919910431, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.017233625054359436, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01717166043817997, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010355153121054173, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.21415546536445618, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.45.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.24784792959690094, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.17806923389434814, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.14647486805915833, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1229802668094635, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.10889311134815216, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07853057235479355, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.14579002559185028, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.131770059466362, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.11830753087997437, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0733586847782135, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.07321351766586304, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0747368186712265, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06408490240573883, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.05384122580289841, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.05114791542291641, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.038293398916721344, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.030199222266674042, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02934126928448677, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.024126147851347923, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.022483566775918007, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.022248340770602226, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0227340217679739, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.019327877089381218, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.017716629430651665, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.17806923389434814, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.45.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2662809491157532, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.24926869571208954, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2440713495016098, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2208186239004135, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12586084008216858, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11941920220851898, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1394892781972885, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12918104231357574, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12755438685417175, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11223965138196945, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10667455941438675, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07105910778045654, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06186172738671303, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.060354191809892654, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05999957025051117, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03550545871257782, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.031087327748537064, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03098515048623085, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02832338958978653, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.028257165104150772, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.018622122704982758, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.018806200474500656, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01813303492963314, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012532704509794712, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2208186239004135, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.45.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.287447988986969, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.26902076601982117, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2634044885635376, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23828957974910736, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1357220709323883, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12875765562057495, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1503080278635025, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13929669559001923, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13754191994667053, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12101710587739944, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11495453119277954, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07654966413974762, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06658557802438736, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0649513453245163, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0645652636885643, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03821873664855957, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.0331016406416893, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03299087658524513, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03007093444466591, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02999872714281082, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.019832290709018707, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019483165815472603, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01929379068315029, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012265953235328197, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1503080278635025, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.45.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2404063642024994, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.221479132771492, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.21210795640945435, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1903732568025589, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10980084538459778, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.10224667191505432, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.1284516155719757, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11848418414592743, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.11180736124515533, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09775560349225998, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.0926140695810318, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06473354995250702, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05624649301171303, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05248313397169113, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.051571451127529144, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03231231868267059, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.027150698006153107, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.026750832796096802, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.024795861914753914, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02426896058022976, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01702076755464077, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.017044672742486, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01581445522606373, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011364050209522247, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.221479132771492, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.46.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13415828347206116, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12423063814640045, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11986327171325684, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1074405312538147, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06256674230098724, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05820344761013985, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07205189764499664, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06674940884113312, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06358999013900757, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05516117811203003, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05236462503671646, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0365736149251461, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.031963493674993515, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.030076054856181145, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0296145211905241, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.018321853131055832, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.015630336478352547, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01543049979954958, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.014154672622680664, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013946943916380405, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009674337692558765, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00985962525010109, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009086045436561108, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006710696965456009, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13415828347206116, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.46.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11221825331449509, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.10400421917438507, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.09925387054681778, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.08902465552091599, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.052145056426525116, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.047982871532440186, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06233808770775795, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.057454004883766174, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05301310122013092, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04610694944858551, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.044031184166669846, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0315319262444973, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.027434492483735085, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.025079594925045967, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.024505142122507095, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01578695885837078, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.012979472056031227, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.012688500806689262, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.011750604026019573, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.011441660113632679, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00834500789642334, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008246642537415028, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007617654278874397, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0054077450186014175, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11221825331449509, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.46.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.22922256588935852, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2131042331457138, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.20800426602363586, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18685349822044373, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10707651078701019, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10100856423377991, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.12015780061483383, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11045800894498825, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10865186899900436, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09464100748300552, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08978044241666794, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.060887835919857025, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05265624448657036, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0511072613298893, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.050734732300043106, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.030370021238923073, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02580498531460762, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02569885365664959, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.023206550627946854, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.023126592859625816, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015506835654377937, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014884949661791325, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014873911626636982, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00882190652191639, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.22922256588935852, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.46.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2519764304161072, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.20540159940719604, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1858234852552414, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.13340474665164948, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.11429537087678909, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.09447423368692398, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.14448478817939758, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.1304846853017807, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.11927914619445801, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07966187596321106, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.07208281755447388, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07358488440513611, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06337723135948181, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.056151971220970154, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.05434483289718628, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03795190528035164, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.031063523143529892, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.030387992039322853, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.024955235421657562, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.02397446520626545, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.02235567569732666, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.022436678409576416, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.02034483291208744, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.017444156110286713, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.20540159940719604, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.46.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.270171195268631, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.25272005796432495, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.24734170734882355, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22360719740390778, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12785865366458893, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12121322751045227, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14173255860805511, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1312689483165741, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12960372865200043, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11388072371482849, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10812035948038101, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0722389668226242, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06287769973278046, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06133325397968292, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.060963086783885956, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0360911525785923, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.031592175364494324, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03148683160543442, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.028752418234944344, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02868608571588993, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.018936170265078545, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019114414229989052, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01843157783150673, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012741969898343086, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22360719740390778, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.46.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.29484042525291443, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2758237421512604, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.269985556602478, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.24408423900604248, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13936661183834076, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1321517527103424, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1543012261390686, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14309273660182953, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14126381278038025, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.1241513192653656, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11774025112390518, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07856278866529465, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06839260458946228, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06670650094747543, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06630218029022217, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03919796273112297, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.033982597291469574, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03386587277054787, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03083142451941967, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.030754348263144493, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.02028135396540165, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019976569339632988, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.0197250135242939, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012540997937321663, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1543012261390686, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.46.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2560400664806366, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2363966554403305, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.22772443294525146, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.20453830063343048, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.11688937991857529, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.10965100675821304, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.13452301919460297, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.12433849275112152, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.11889169365167618, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.10432949662208557, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.09869156777858734, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06768189370632172, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05889838561415672, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05571727454662323, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.05495164170861244, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03373925760388374, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.028544967994093895, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.028223494067788124, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.026060696691274643, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02561817318201065, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.017557010054588318, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01739545539021492, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01653778739273548, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011162035167217255, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.22772443294525146, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.47.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14729724824428558, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13711172342300415, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13235034048557281, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11915314942598343, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06921422481536865, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06460999697446823, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08005715161561966, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07382665574550629, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07025263458490372, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06137815862894058, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.058527667075395584, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04073648899793625, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03547987341880798, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03340265154838562, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03289691358804703, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.020555155351758003, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.017544960603117943, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.017322015017271042, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.016007153317332268, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015776636078953743, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.011121327057480812, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.011306955479085445, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.010484596714377403, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.007992523722350597, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14729724824428558, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.47.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.13245205581188202, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1232103481888771, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.11830969154834747, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.10614483058452606, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.062134016305208206, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.05758791044354439, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.0729907900094986, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06738053262233734, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.06306976079940796, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.05502668023109436, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.052396614104509354, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03696908429265022, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.032272860407829285, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.029882322996854782, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.02930636703968048, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.018543757498264313, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.015438311733305454, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.015164880082011223, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.013998246751725674, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.013702910393476486, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.009806622751057148, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.009693617932498455, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.009083343669772148, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.006362678483128548, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.13245205581188202, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.47.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.3043566644191742, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2841740846633911, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.27772292494773865, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.25088581442832947, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1440184861421585, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.1362358182668686, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.16131149232387543, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.14839676022529602, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.14608439803123474, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.1279481202363968, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.12158740311861038, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.08208538591861725, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.07093901187181473, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06893512606620789, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06845696270465851, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.040963828563690186, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.034898821264505386, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03476449474692345, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.03154423087835312, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.03141263499855995, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.02099250629544258, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.020264431834220886, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.020167915150523186, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.012223739176988602, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.16131149232387543, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.47.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.26394954323768616, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.22264595329761505, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.20948749780654907, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.17162953317165375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.1197071298956871, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.10383478552103043, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.14614811539649963, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.12909787893295288, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.1241668090224266, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.09284655749797821, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.08457939326763153, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07437030225992203, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06500475108623505, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.06096017733216286, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.05994231253862381, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.04015200585126877, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.03691391274333, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.036644842475652695, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.03271682932972908, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.032335978001356125, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.026181820780038834, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.02901693619787693, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.02516322024166584, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.02547033131122589, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.22264595329761505, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.47.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.27225786447525024, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2545514702796936, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2491416484117508, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22514839470386505, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12882356345653534, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12211628258228302, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14283227920532227, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13229018449783325, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1306033730506897, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11468929052352905, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10886996239423752, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07275283336639404, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06336324661970139, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.061793357133865356, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06141792610287666, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03635765612125397, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03180176019668579, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03169381991028786, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.028920549899339676, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.028851980343461037, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.019040530547499657, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019194507971405983, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.018534939736127853, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012731697410345078, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22514839470386505, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.47.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2960938811302185, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.27692168951034546, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.271009624004364, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.24500371515750885, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1400325745344162, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13275429606437683, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1551499366760254, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14377661049365997, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14195102453231812, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12465165555477142, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11829889565706253, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07901544868946075, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06876518577337265, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0670536607503891, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06664571911096573, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03944965824484825, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.034200798720121384, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03408225253224373, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03102501854300499, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.030950330197811127, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020458003506064415, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.02016608603298664, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019899381324648857, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012754849158227444, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1551499366760254, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.47.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.25655582547187805, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2368273138999939, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.22812238335609436, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.20467066764831543, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.11720956861972809, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.10989730060100555, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.13534584641456604, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.12466412037611008, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.11921511590480804, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.10450579226016998, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.09882944077253342, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06803721189498901, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05921342968940735, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.056006211787462234, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.055237218737602234, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.034088633954524994, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.028973110020160675, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02865503914654255, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.026497503742575645, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02606041356921196, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.018141789361834526, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.018047621473670006, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01713917776942253, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.01212510745972395, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.22812238335609436, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.48.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12710021436214447, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11779792606830597, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11312571913003922, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10140684992074966, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.059605568647384644, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.055160313844680786, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.0700179785490036, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06447570025920868, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.060564540326595306, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.052609264850616455, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05012794956564903, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03561926260590553, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03103141486644745, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.028796538710594177, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.028242824599146843, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.017972420901060104, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01514679566025734, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.014894360676407814, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.013771631754934788, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013507500290870667, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009724566712975502, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00986489187926054, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009053708054125309, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.00695668812841177, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12710021436214447, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.48.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10455942153930664, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.09686999768018723, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.09187197685241699, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.08219335973262787, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.04875713214278221, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04451772943139076, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.05919762700796127, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05459793284535408, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.04953916743397713, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04305019974708557, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.041115012019872665, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.029919631779193878, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.026169653981924057, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.023483717814087868, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.022822344675660133, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.015008995309472084, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.012183412909507751, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.011850109323859215, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.011037436313927174, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.010676688514649868, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.007965337485074997, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.007863068021833897, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007175047881901264, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005173753947019577, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10455942153930664, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.48.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2675948441028595, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.24833345413208008, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.24229179322719574, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.21743766963481903, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.12572894990444183, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.11836373805999756, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.141356959939003, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.1299188733100891, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.12770038843154907, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11094284057617188, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10505403578281403, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07185432314872742, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06205279752612114, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06014628708362579, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.059693727642297745, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03585212677717209, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.030447034165263176, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03030461259186268, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02735067531466484, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.027226535603404045, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.018357053399086, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01771553047001362, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.017557714134454727, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010661148466169834, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.21743766963481903, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.48.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2562255859375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.18637990951538086, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.15983334183692932, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.12283267825841904, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.11036620289087296, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0822041779756546, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1432340443134308, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.12927304208278656, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.12090398371219635, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07291580736637115, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0712064579129219, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07284089177846909, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06290696561336517, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.05449939891695976, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.05231533199548721, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03786592558026314, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.030566338449716568, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.03005046397447586, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.024162044748663902, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.022860145196318626, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.022476689890027046, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.022677656263113022, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.020177222788333893, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.017886465415358543, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.18637990951538086, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.48.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.27831029891967773, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2601284980773926, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2545560598373413, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23002949357032776, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13177691400051117, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12490971386432648, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14615967869758606, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13532686233520508, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13360735774040222, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11729513108730316, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11139556020498276, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07458791136741638, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0648505836725235, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0632396936416626, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06286101043224335, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.037282221019268036, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03260950744152069, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03250139579176903, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.029667425900697708, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02959533967077732, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.019638603553175926, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019773242995142937, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01911746710538864, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013231524266302586, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23002949357032776, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.48.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2989984154701233, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.27953040599823, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2736045718193054, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.24726897478103638, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1414283812046051, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1340685784816742, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1568061113357544, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14521875977516174, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14337792992591858, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12589935958385468, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11948603391647339, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07994692027568817, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06947724521160126, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06774449348449707, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06733822077512741, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03991993889212608, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03457462415099144, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03445587679743767, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.031367383897304535, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03129180520772934, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020780401304364204, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.02043081820011139, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.02021479234099388, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012961909174919128, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1568061113357544, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.48.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.26461824774742126, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2444555014371872, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.2362114042043686, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.2116944044828415, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12099268287420273, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.11380525678396225, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.1376802772283554, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.1274697184562683, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.1229577586054802, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.10783480107784271, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.10167061537504196, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06931398063898087, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.06052672863006592, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.057703886181116104, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.05703805759549141, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03461914509534836, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02967894822359085, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02940959483385086, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02709527127444744, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.026718085631728172, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01807914860546589, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01815485581755638, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01718852110207081, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011931032873690128, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.2116944044828415, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.49.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1152358129620552, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1068200170993805, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10207395255565643, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09152578562498093, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.05412393808364868, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04984317719936371, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06421314924955368, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.05938231945037842, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.054983630776405334, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04775932803750038, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.045539841055870056, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03258325159549713, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.028518840670585632, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.026153352111577988, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.025571132078766823, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.016406984999775887, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.013765441253781319, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01349044218659401, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.012514198198914528, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.012228456325829029, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.008849936537444592, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00902837049216032, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008167732506990433, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006351812276989222, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1152358129620552, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.49.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.09452883899211884, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.08761170506477356, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.08273018151521683, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.07408072054386139, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.04405096918344498, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.040014561265707016, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.05417155101895332, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04987743869423866, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.044736068695783615, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03890913352370262, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.037334930151700974, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.027413098141551018, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.023901840671896935, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02124679833650589, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.02057606168091297, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0137612484395504, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.011081011034548283, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.010740833356976509, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.010067003779113293, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.00970013253390789, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.007344646379351616, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.007277455180883408, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.006571565289050341, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004872691351920366, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.09452883899211884, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.49.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.26628392934799194, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.24753716588020325, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.241385817527771, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.21690733730793, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1255263239145279, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.11810846626758575, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.14121025800704956, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.12981738150119781, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.127370685338974, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.1107049360871315, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10496693104505539, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07166799157857895, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06200264021754265, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06001167371869087, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.059536222368478775, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03578820079565048, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.03034764900803566, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.030198141932487488, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02724658139050007, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.027115877717733383, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.018290065228939056, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.017589472234249115, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.017480097711086273, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010473288595676422, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.21690733730793, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.49.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.19936415553092957, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1392289102077484, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11126430332660675, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09583890438079834, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08529471606016159, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.058057837188243866, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.11697575449943542, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.10677812248468399, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0950300469994545, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05995972082018852, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.058856263756752014, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.06054847687482834, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.05257347598671913, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04311748594045639, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.040607087314128876, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03135376796126366, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.025587420910596848, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0248687993735075, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.021597959101200104, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.02018514648079872, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.018800456076860428, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.020589323714375496, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01620538718998432, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01707872748374939, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.19936415553092957, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.49.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2807512879371643, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2624519169330597, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.25682783126831055, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23218338191509247, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13296157121658325, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12603189051151276, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1475621610879898, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13653896749019623, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13479752838611603, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11834181845188141, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11241178214550018, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07529695332050323, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06544686108827591, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06381738930940628, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06343427300453186, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03763679787516594, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03292446210980415, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.032818134874105453, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02996089868247509, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02988707460463047, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01983623020350933, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019995570182800293, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01931210234761238, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013425111770629883, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1475621610879898, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.49.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2976706624031067, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2783215343952179, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.27238085865974426, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.24620306491851807, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1408366858959198, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13349418342113495, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1561051607131958, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14460314810276031, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14276091754436493, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12533798813819885, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11897460371255875, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07956773787736893, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06918632239103317, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0674610584974289, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06705085933208466, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.039731353521347046, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03443513065576553, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03431924059987068, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03123755194246769, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.031163884326815605, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020665999501943588, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.02035583183169365, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.02010161429643631, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012936081737279892, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1561051607131958, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.49.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.23094280064105988, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2131299078464508, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.2054394781589508, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1839967668056488, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.10586658120155334, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.09922736138105392, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.12186554819345474, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11246372014284134, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.10768254101276398, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09415868669748306, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.08882156759500504, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06144820153713226, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.053439911454916, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.050602059811353683, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.04991722106933594, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.030855825170874596, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.026120558381080627, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.025835061445832253, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.023817120119929314, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.023429812863469124, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.016521278768777847, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01617409661412239, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.015643084421753883, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010761738754808903, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2131299078464508, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.50.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12497734278440475, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1153586208820343, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11047638207674026, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09872698783874512, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.05842481553554535, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05385955050587654, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06882786750793457, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06356368213891983, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.05944474786520004, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05130894482135773, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04884796589612961, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03499200940132141, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.030511990189552307, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.028206106275320053, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.02764168754220009, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01759474352002144, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01480515580624342, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.014547780156135559, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.013406066223978996, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013131151907145977, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009413834661245346, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009616273455321789, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008720808662474155, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006723076570779085, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12497734278440475, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.50.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1038462221622467, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.09599420428276062, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.09038812667131424, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.08077183365821838, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.04831435903906822, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.043803002685308456, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.05988345295190811, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05495263636112213, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.04910852015018463, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.042571354657411575, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.040792316198349, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03025294840335846, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.026290645822882652, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.023314176127314568, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.022573523223400116, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01519495528191328, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.012166026048362255, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.011786366812884808, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.011026023887097836, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.010609681718051434, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008119110018014908, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008001493290066719, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007226990070194006, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005337182432413101, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1038462221622467, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.50.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2629230320453644, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.24336041510105133, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.23707804083824158, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2122534215450287, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.12339293956756592, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.11587979644536972, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1391102820634842, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.12775340676307678, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1254321187734604, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.108433797955513, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10266780108213425, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07062479108572006, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.061006225645542145, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05899280309677124, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05851852148771286, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.035251814872026443, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02984248287975788, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.029677897691726685, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02671889029443264, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.026571357622742653, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.018013108521699905, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.017306266352534294, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.017191991209983826, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010311387479305267, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2122534215450287, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.50.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.26358869671821594, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.17517024278640747, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1411621868610382, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.12846773862838745, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.11681026965379715, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07793320715427399, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.14344501495361328, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.13247190415859222, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.1251685470342636, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07174328714609146, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.07160373032093048, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07296110689640045, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06336396187543869, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.056265830993652344, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.054474618285894394, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03654560446739197, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.029274335131049156, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02878328040242195, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.020580248907208443, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.019214024767279625, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.019297383725643158, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01912584900856018, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01703360490500927, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.012625991366803646, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.17517024278640747, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.50.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2853913903236389, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.26681196689605713, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.261072039604187, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23590530455112457, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13527217507362366, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1281575858592987, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15022379159927368, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1389220952987671, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13713891804218292, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12034260481595993, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11435046046972275, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07671849429607391, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0666026771068573, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06493940949440002, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06454706937074661, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.038401421159505844, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03352362662553787, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03341391310095787, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030497469007968903, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.030422594398260117, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.02034623734652996, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.02038745768368244, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019819684326648712, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01372084952890873, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15022379159927368, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.50.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.300895631313324, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2812991142272949, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2752770781517029, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.24874503910541534, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.14244841039180756, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13496482372283936, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15790413320064545, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14631319046020508, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1444290727376938, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12674462795257568, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.12025295197963715, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08048585802316666, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.07000277191400528, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06824231147766113, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06782162189483643, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.04017786681652069, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03481949865818024, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03470069542527199, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03157060220837593, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.031491585075855255, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020870152860879898, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020557381212711334, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.020294997841119766, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013021514751017094, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15790413320064545, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.50.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2699972689151764, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.24925865232944489, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.24102653563022614, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.21570734679698944, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12335038185119629, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.1161232590675354, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.1404433697462082, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.1296815127134323, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.12537077069282532, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.10978728532791138, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.10342388600111008, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.0706457570195198, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.061431240290403366, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.058721691370010376, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.05807844176888466, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03523792698979378, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.029989738017320633, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.029739873483777046, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.027299001812934875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02693587727844715, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.018348785117268562, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01801990158855915, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01747281663119793, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011442349292337894, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.21570734679698944, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.51.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1312638372182846, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1214737594127655, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11664924770593643, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10449972003698349, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06158754229545593, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05699462071061134, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07195371389389038, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06651599705219269, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0625949278473854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05420830845832825, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.051580701023340225, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03657345473766327, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.031900133937597275, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.02969910204410553, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.029164167121052742, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.018364138901233673, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01553136482834816, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.015280967578291893, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.014068433083593845, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013807099312543869, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009805362671613693, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009980353526771069, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009134447202086449, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006902399472892284, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1312638372182846, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.51.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11489736288785934, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.10627715289592743, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.10096023231744766, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09032716602087021, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05362233519554138, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04908992722630501, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06472305953502655, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05965244770050049, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05451136827468872, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.047260183840990067, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04510748013854027, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.032715216279029846, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02852136269211769, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02588312141597271, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.025218214839696884, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.016440117731690407, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.013459541834890842, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.01312456838786602, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.012176113203167915, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.011824947781860828, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00875774584710598, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008697346784174442, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007950049825012684, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005817484110593796, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11489736288785934, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.51.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.24096354842185974, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2232910692691803, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.21768738329410553, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19520623981952667, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11328504979610443, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10648281872272491, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1273607760667801, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.11710084974765778, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.11506842821836472, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09969548881053925, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0944068431854248, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06475894153118134, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.055947914719581604, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.054185327142477036, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05376780778169632, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.032325513660907745, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.027423446998000145, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02729228138923645, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.024581044912338257, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.024469638243317604, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01655309833586216, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.015944242477416992, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.015830107033252716, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00956355407834053, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2232910692691803, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.51.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.21517135202884674, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16824497282505035, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1541779339313507, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10621004551649094, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.09841562062501907, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.08120888471603394, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.11342316120862961, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.10401196777820587, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.10204097628593445, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06157311797142029, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0558198019862175, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05785446986556053, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04991264268755913, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04739081487059593, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04677506536245346, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02910231053829193, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.024767756462097168, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.024647975340485573, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.017719801515340805, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.017552070319652557, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01583777368068695, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.015735605731606483, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.015064789913594723, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010981427505612373, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.21517135202884674, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.51.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.28587469458580017, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.267229288816452, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2614986300468445, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23628190159797668, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1355241984128952, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12840546667575836, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15033212304115295, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13919252157211304, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13740289211273193, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12051045894622803, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11439253389835358, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07665161788463593, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06668280810117722, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06501086056232452, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0646134540438652, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0383213609457016, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.033454351127147675, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03333890810608864, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030394172295928, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.030322100967168808, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.02013375610113144, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.02018963173031807, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019595513120293617, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013384975492954254, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15033212304115295, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.51.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.29972922801971436, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.28010547161102295, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.27409905195236206, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.24766477942466736, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.14191274344921112, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1344245970249176, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15733152627944946, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14578236639499664, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14387992024421692, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12621457874774933, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11972026526927948, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08020980656147003, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0697324350476265, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06797342002391815, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06755286455154419, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.04004549980163574, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03465530276298523, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03453754261136055, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03140665590763092, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03132643923163414, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020792411640286446, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020428184419870377, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.020217932760715485, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01288056094199419, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15733152627944946, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.51.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2629602551460266, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.24266879260540009, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.23469701409339905, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.21003688871860504, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12026926130056381, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.1131826788187027, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.13686303794384003, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.12639674544334412, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.12227866053581238, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.10698974132537842, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.100786492228508, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06892266869544983, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05994711071252823, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05729585886001587, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.05666082352399826, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03451498970389366, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02930743806064129, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.029057782143354416, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.026677528396248817, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.026320111006498337, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01820027269423008, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01767442561686039, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01736636459827423, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011310983449220657, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.21003688871860504, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.52.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1469070315361023, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13576164841651917, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13062448799610138, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11693285405635834, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06886964291334152, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06379402428865433, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08021533489227295, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07391863316297531, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06998699903488159, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06055206060409546, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05760413780808449, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.040834784507751465, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03560198098421097, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.033281221985816956, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.032706622034311295, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.020648015663027763, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.017543986439704895, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.017292369157075882, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01592271961271763, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015657391399145126, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.011208122596144676, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.011422701179981232, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.010503392666578293, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.008136147633194923, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1469070315361023, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.52.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12762823700904846, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.11794572323560715, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.11238650232553482, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.10063143819570541, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.059548404067754745, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.05462409928441048, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.07135840505361557, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06569375097751617, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.06053180247545242, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.05238353833556175, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.05001259222626686, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.036124784499406815, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.03141535818576813, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.028644435107707977, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.027950098738074303, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.018120886757969856, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.014801695942878723, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.014469470828771591, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.013351631350815296, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.012999395839869976, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.009546853601932526, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.009385108016431332, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.008687589317560196, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.006108371540904045, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12762823700904846, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.52.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2593996226787567, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.24052394926548004, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.23449482023715973, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.209980770945549, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.12177585065364838, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.1144501343369484, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13700050115585327, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.12590861320495605, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.12368262559175491, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10706333816051483, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10137299448251724, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0695686787366867, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.060068488121032715, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05819554254412651, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.057751379907131195, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03469588980078697, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02944253198802471, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02930796518921852, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.026390833780169487, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.026258382946252823, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.017744826152920723, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01708894968032837, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.016967151314020157, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010250222869217396, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.209980770945549, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.52.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.25543099641799927, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.2210647463798523, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.21068444848060608, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.16770879924297333, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.11691489815711975, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.10392734408378601, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1383596807718277, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.1233140230178833, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.12036144733428955, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.09059147536754608, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.07877220213413239, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.06974741816520691, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.060544826090335846, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.05771839991211891, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.057034026831388474, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03632226958870888, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.032405637204647064, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.032223135232925415, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.027747444808483124, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.027563173323869705, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.021691786125302315, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.023140063509345055, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0209011510014534, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.018902776762843132, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.2210647463798523, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.52.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.28305119276046753, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.26447737216949463, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.25874999165534973, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2334897518157959, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13418205082416534, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12709413468837738, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14894261956214905, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1378525346517563, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13606762886047363, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11922898888587952, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.1131332740187645, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07595799118280411, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06603126227855682, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06436334550380707, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06396491825580597, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03798331692814827, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.0330839641392231, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03297269344329834, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030031800270080566, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02995331585407257, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.019957218319177628, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.0199067872017622, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019418494775891304, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013114244677126408, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14894261956214905, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.52.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.3043544292449951, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.28429555892944336, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.278163343667984, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2511471211910248, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1441534459590912, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13652049005031586, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1598757952451706, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14809349179267883, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14618298411369324, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12809249758720398, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.12139594554901123, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08142765611410141, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.07085780799388885, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06906521320343018, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06863868981599808, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.04072465002536774, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.035233065485954285, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03511318191885948, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.031911153346300125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03183059021830559, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.021247370168566704, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020797476172447205, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.020662054419517517, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013155302964150906, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1598757952451706, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.52.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.27553245425224304, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2551296651363373, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.24716021120548248, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.22192725539207458, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12599892914295197, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.11898934841156006, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.14288854598999023, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.13204605877399445, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.12794475257396698, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.11262916773557663, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.10631771385669708, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.07177966833114624, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.06261450052261353, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.0600409135222435, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.05941855162382126, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03580400347709656, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.030764512717723846, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.030523069202899933, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02813478745520115, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.027789944782853127, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.018592465668916702, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.018603764474391937, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.017764540389180183, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.012041399255394936, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.22192725539207458, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.53.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12326319515705109, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11391454190015793, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10887934267520905, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09735573083162308, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.057918377220630646, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0532861165702343, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06863022595643997, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06335842609405518, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.05887996032834053, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05088554322719574, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04844299703836441, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.034898217767477036, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03048066236078739, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.0280381441116333, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.027441449463367462, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01762421242892742, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.014841200783848763, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01456400379538536, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01347216498106718, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013177321292459965, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009605671279132366, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009825406596064568, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008875785395503044, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.007040182128548622, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12326319515705109, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.53.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.09968013316392899, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.0921807587146759, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.08656277507543564, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.07734712213277817, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.046641357243061066, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04211842641234398, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.058255620300769806, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.053428664803504944, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.04741203039884567, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04106142744421959, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03937968611717224, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02938716486096382, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02559278905391693, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.022541970014572144, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.021783286705613136, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01479496993124485, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.011822841130197048, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.011428575031459332, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.010718915611505508, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.010295934043824673, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.007966815494000912, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.007883823476731777, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007048636209219694, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00534208444878459, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.09968013316392899, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.53.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.28419938683509827, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.26332271099090576, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.25664669275283813, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2300623059272766, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.13408054411411285, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12599320709705353, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1507331132888794, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.1385699063539505, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.13620318472385406, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11775314807891846, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.11140947788953781, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0766991674900055, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0661962702870369, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06416813284158707, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06366930902004242, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03827909380197525, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.03249424695968628, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03234103322029114, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.029092419892549515, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02897208370268345, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01960837095975876, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01893717795610428, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.018759917467832565, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011434108950197697, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2300623059272766, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.53.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.23071032762527466, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1488458514213562, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1127668023109436, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10680707544088364, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.09742708504199982, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05692940205335617, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.13118264079093933, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.11994711309671402, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.10974059253931046, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06354913860559464, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06491202861070633, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.06706226617097855, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.057660285383462906, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.047567810863256454, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04489854723215103, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03384627029299736, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02571028657257557, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.025008918717503548, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01955791935324669, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.017650406807661057, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.018736321479082108, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.018603473901748657, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01571376621723175, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.013327018357813358, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.23071032762527466, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.53.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2851279079914093, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2663067877292633, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.26049625873565674, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23510277271270752, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13523942232131958, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12805470824241638, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1500372737646103, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13893844187259674, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13712197542190552, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12009598314762115, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11393283307552338, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07651069760322571, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06659387052059174, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.064908467233181, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06450872868299484, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03824438899755478, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.033433400094509125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03331952169537544, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030354101210832596, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.030277960002422333, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020078208297491074, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020223375409841537, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019535204395651817, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013462824746966362, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1500372737646103, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.53.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.30579155683517456, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.28562942147254944, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2794467806816101, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.252215176820755, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1448734849691391, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13717608153820038, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16054454445838928, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14885850250720978, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1469191163778305, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12867674231529236, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.12196718156337738, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08179877698421478, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.07120940089225769, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06939888000488281, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06896895170211792, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.04081827774643898, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03537687286734581, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03525207191705704, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.032019130885601044, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.031936950981616974, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.02112276293337345, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020833957940340042, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.0205246489495039, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013117576017975807, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16054454445838928, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.53.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2787564694881439, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.25817087292671204, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.2504834830760956, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.22483913600444794, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12749317288398743, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.12055115401744843, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.14402897655963898, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.13311634957790375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.12946341931819916, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.11396757513284683, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.10753996670246124, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.07243803143501282, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.0630570650100708, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.060651153326034546, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.060073550790548325, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.036101333796978, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.03089894913136959, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.03068483993411064, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02821694128215313, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02790268138051033, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.018695786595344543, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.018388228490948677, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01791737601161003, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011564728803932667, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.22483913600444794, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.54.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12553770840168, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11597363650798798, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11101625114679337, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09936508536338806, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.05894181877374649, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05432607978582382, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.0694187581539154, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06420476734638214, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.05994158983230591, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.051786333322525024, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04923471808433533, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03525508567690849, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.030834589153528214, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.028470776975154877, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.027897190302610397, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.017731083557009697, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.014961931854486465, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.014702093787491322, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.013555112294852734, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013276033103466034, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.00951451901346445, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009749113582074642, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008814298547804356, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0068458132445812225, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12553770840168, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.54.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.09880027920007706, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.09141421318054199, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.0858616977930069, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.07680051028728485, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.046210452914237976, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.041735149919986725, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.05788978189229965, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05292940139770508, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.04700165614485741, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.040740400552749634, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.039173901081085205, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02923755720257759, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0253790020942688, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02233748510479927, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.021576683968305588, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01468728482723236, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.01168221514672041, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.011294132098555565, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.01059669442474842, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.010171275585889816, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.007873240858316422, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.007752537727355957, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.006948849651962519, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005209760274738073, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.09880027920007706, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.54.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2828274071216583, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2619716227054596, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2552192211151123, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2287340760231018, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.13322973251342773, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.1251765340566635, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.14981776475906372, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.13792717456817627, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.13536590337753296, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.1169915646314621, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.11079224199056625, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07622087746858597, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06587150692939758, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06375555694103241, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06324436515569687, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03803025186061859, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.03229614719748497, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03212837874889374, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02890940196812153, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.028772519901394844, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01948157697916031, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.018814172595739365, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.018628234043717384, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011350488290190697, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2287340760231018, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.54.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.25492316484451294, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1576824188232422, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11789676547050476, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11960232257843018, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.11265512555837631, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07250843197107315, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.13949552178382874, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.1285249888896942, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.12110528349876404, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06696127355098724, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06833542883396149, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07097085565328598, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06155216693878174, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.05442097410559654, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.05261870101094246, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03563123196363449, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.028534794226288795, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.028032803907990456, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01986747235059738, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01845046505331993, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.018931768834590912, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01895133964717388, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01666194014251232, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.012854060158133507, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1576824188232422, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.54.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2886934280395508, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2695865333080292, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2637457251548767, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2378825545310974, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1369059979915619, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12961888313293457, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15180762112140656, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.140681654214859, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13881585001945496, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12154494225978851, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11518675088882446, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07734520733356476, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06736567616462708, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06564798951148987, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06523602455854416, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03862359747290611, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03368202969431877, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.033566705882549286, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030540071427822113, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03046158328652382, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.02012716978788376, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020183654502034187, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019565846771001816, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013182567432522774, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15180762112140656, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.54.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.30741795897483826, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.28709322214126587, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.28084588050842285, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.25340765714645386, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.14564277231693268, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13787472248077393, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16141743957996368, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14964509010314941, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14769582450389862, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.1293315291404724, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.12253503501415253, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0822448655962944, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.07159379124641418, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0697631686925888, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06932716816663742, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.041046351194381714, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03552977740764618, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03540361672639847, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03214693069458008, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03205973282456398, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.021235138177871704, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020884735509753227, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.02063702791929245, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013079453259706497, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16141743957996368, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.54.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2815282940864563, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2605219781398773, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.2527519762516022, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.22676752507686615, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12867195904254913, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.12166006118059158, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.14515265822410583, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.1342441886663437, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.13067303597927094, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.11491183936595917, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.1083618551492691, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.07302054017782211, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.06355063617229462, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.06117364764213562, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.060605816543102264, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.036355771124362946, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.031102396547794342, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.030891919508576393, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.028361069038510323, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02805395983159542, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.018777811899781227, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.018401524052023888, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01800677552819252, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011431852355599403, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.22676752507686615, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.55.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13978682458400726, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12955166399478912, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12461642175912857, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11172543466091156, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06584921479225159, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06104961410164833, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.0765160471200943, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07092365622520447, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06691637635231018, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05801703780889511, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.055086418986320496, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03888177126646042, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03405626863241196, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.031769104301929474, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.031211409717798233, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01954318769276142, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.016611158847808838, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.016356518492102623, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015051602385938168, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.014786841347813606, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0104450024664402, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.010654866695404053, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009753978811204433, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.007368455640971661, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13978682458400726, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.55.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1151808351278305, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.10677776485681534, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.10123060643672943, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09067843854427338, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05406879633665085, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.049367744475603104, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06574267148971558, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.060618914663791656, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05494307726621628, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.047677841037511826, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04561140760779381, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0332394503057003, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.028998607769608498, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.026089351624250412, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.025355570018291473, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.016678733751177788, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.013565878383815289, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.013204237446188927, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.012282604351639748, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.011890916153788567, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00886050146073103, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008774706162512302, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007978862151503563, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005809963680803776, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1151808351278305, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.55.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2919071316719055, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2712395191192627, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2647343873977661, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.23780879378318787, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.13790881633758545, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12988367676734924, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1548946499824524, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.14229269325733185, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.14002862572669983, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.12155933678150177, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.11522423475980759, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07879563421010971, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06799731403589249, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06596384197473526, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06549232453107834, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03931683301925659, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0333789698779583, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03323407098650932, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.029950492084026337, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.029839860275387764, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.020112095400691032, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.019317448139190674, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.019253404811024666, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011538655497133732, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1548946499824524, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.55.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.25845056772232056, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.17851552367210388, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.14276525378227234, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11430230736732483, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.10752971470355988, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07322350889444351, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1471337527036667, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.13578000664710999, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.1227642074227333, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07022549211978912, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.07197985798120499, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07515990734100342, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06510025262832642, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.05229145288467407, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.048803217709064484, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03762519732117653, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.027932634577155113, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.027003487572073936, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02104976773262024, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.018599027767777443, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.020073415711522102, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.02018936350941658, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01599237322807312, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.013841002248227596, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.17851552367210388, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.55.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2893090546131134, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2701909840106964, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.26428940892219543, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23844043910503387, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13728244602680206, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12994426488876343, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1523047387599945, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14109590649604797, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1392381340265274, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12189949303865433, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11562814563512802, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07768598198890686, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0676111951470375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06588239222764969, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06547493487596512, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03882719576358795, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03389010578393936, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03377407416701317, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030746957287192345, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.030665671452879906, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020338941365480423, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020432645455002785, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019776873290538788, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013506422750651836, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1523047387599945, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.55.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.3080991804599762, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.28769826889038086, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2814580798149109, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2539135217666626, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.14598345756530762, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1381983608007431, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16181446611881256, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.15001574158668518, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14806076884269714, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12962374091148376, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.12283355742692947, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08244866132736206, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0717754065990448, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06993430852890015, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06949903070926666, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.041155118495225906, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.035621657967567444, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.035498589277267456, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03222805634140968, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03214135393500328, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.021298713982105255, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020940689370036125, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.020694704726338387, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013114984147250652, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16181446611881256, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.55.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2777472734451294, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2572587728500366, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.2494494616985321, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.22382406890392303, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.1270790994167328, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.1201213076710701, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.1438886821269989, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.13283340632915497, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.12904022634029388, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.11357762664556503, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.10716594755649567, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.0724022313952446, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.06299328804016113, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.06051509082317352, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.05993080511689186, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03614092990756035, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.03094404749572277, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.030722184106707573, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02827589027583599, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.027955248951911926, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.018841814249753952, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01859321817755699, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.018028076738119125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011920956894755363, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.22382406890392303, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.56.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1245858371257782, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11540691554546356, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11027822643518448, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09882007539272308, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0588371604681015, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05412710830569267, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06974773854017258, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0644603818655014, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.059820424765348434, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05189160630106926, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04940589517354965, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03565600514411926, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03121628426015377, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.02862965129315853, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.028002021834254265, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.018164463341236115, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.015398247167468071, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.015098927542567253, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.014067973010241985, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01375471893697977, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.010190054774284363, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01048172265291214, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009448720142245293, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.007803869433701038, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1245858371257782, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.56.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10316844284534454, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.0955030769109726, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.09005361795425415, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.08055685460567474, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.04831141605973244, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04389344900846481, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.05955570191144943, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05489996820688248, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.049121830612421036, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04261269420385361, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04074354097247124, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.030088575556874275, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.026221303269267082, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.023338131606578827, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.02260519005358219, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.015136918053030968, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.012156936340034008, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.01178395003080368, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.011012089438736439, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.01060998160392046, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008059675805270672, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.007959507405757904, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007167429197579622, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005287130828946829, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10316844284534454, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.56.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2864077389240265, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2657105028629303, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.25919315218925476, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2324851155281067, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1351688802242279, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12719511985778809, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.15202410519123077, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.13977226614952087, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1372462958097458, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11904551833868027, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.11266740411520004, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07726192474365234, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06680149585008621, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06467607617378235, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06418272852897644, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03858504071831703, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.032731786370277405, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03256886824965477, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02934783510863781, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.029210269451141357, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01973232626914978, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01899590529501438, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.018865322694182396, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011358104646205902, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.15202410519123077, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.56.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2077006846666336, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13255369663238525, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10169234126806259, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09879206120967865, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08815331757068634, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05258411541581154, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.11870831996202469, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.10667454451322556, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.09898769855499268, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.060344815254211426, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06010912358760834, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.06113211810588837, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.053890325129032135, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04590381309390068, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.043803147971630096, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03275681659579277, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.028874604031443596, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.028462452813982964, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02492549456655979, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.023830272257328033, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.02091897837817669, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.024158991873264313, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.018797148019075394, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.021257735788822174, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2077006846666336, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.56.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.28927576541900635, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.27006760239601135, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2642141282558441, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23829452693462372, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13721075654029846, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1298869252204895, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15217730402946472, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14101827144622803, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13914288580417633, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12179706245660782, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11545702815055847, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07753979414701462, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0675397738814354, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06581136584281921, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06539709120988846, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03874441608786583, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03377938270568848, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03366350755095482, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03062068298459053, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.030539732426404953, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020213646814227104, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020262442529201508, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.0196501687169075, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013247598893940449, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15217730402946472, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.56.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.3087581396102905, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2882711589336395, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2819969058036804, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2543818950653076, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.14631979167461395, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13849703967571259, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1621231883764267, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.15037384629249573, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14837974309921265, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12986096739768982, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.1230192631483078, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08260155469179153, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0719306468963623, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.07008813321590424, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06965115666389465, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0412076897919178, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.035690486431121826, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.035559799522161484, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03226856142282486, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03218657150864601, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.021266629919409752, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020964354276657104, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.020656542852520943, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013111489824950695, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1621231883764267, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.56.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.280568927526474, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2598351538181305, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.2521013915538788, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.22635690867900848, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12827642261981964, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.12134598195552826, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.14503534138202667, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.13388285040855408, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.1302647739648819, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.11472617834806442, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.10823437571525574, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.07288888096809387, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.06343640387058258, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.06103726848959923, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.06046900153160095, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03639671951532364, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.031139828264713287, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.03092902898788452, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02845602110028267, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.028144605457782745, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.018984973430633545, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01858440600335598, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01821078546345234, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011776427738368511, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.22635690867900848, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.57.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1294657438993454, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11959417164325714, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11463796347379684, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1025913879275322, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06091558188199997, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05619160085916519, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07136480510234833, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06612792611122131, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06196232885122299, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05350999906659126, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05074404180049896, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.036238640546798706, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.031768228858709335, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.02941194176673889, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.02884068340063095, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01819382607936859, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01542538683861494, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.015164465643465519, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.013956338167190552, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013675874099135399, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009733559563755989, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009991749189794064, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009031856432557106, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006964751053601503, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1294657438993454, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.57.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.09922931343317032, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.09177446365356445, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.08614766597747803, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.07712850719690323, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.046519823372364044, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04195677861571312, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.058466412127017975, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05340421572327614, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.047320205718278885, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04104974493384361, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.039399247616529465, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.029567060992121696, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.025629956275224686, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.022491421550512314, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.021705426275730133, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.014822508208453655, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.011754411272704601, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.011355308815836906, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.010654673911631107, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.01022376213222742, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.007924663834273815, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0077949343249201775, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.006979992613196373, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005227343179285526, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.09922931343317032, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.57.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2924371659755707, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.27096354961395264, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2640496492385864, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.23657789826393127, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1380353718996048, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12967006862163544, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.15516677498817444, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.14263728260993958, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.14021006226539612, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.1212097629904747, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.11468709260225296, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07900106906890869, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06818366050720215, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06605210900306702, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06553097069263458, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.039408061653375626, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.033423684537410736, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03326624259352684, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02990342304110527, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02977007068693638, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.020151812583208084, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.019396118819713593, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.019277453422546387, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011608201079070568, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.15516677498817444, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.57.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.24822589755058289, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16085699200630188, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12248215079307556, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1171911358833313, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.10669796913862228, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06433311104774475, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.13925573229789734, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.12879955768585205, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.1178928092122078, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06859204173088074, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06931938230991364, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07081373035907745, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06151578947901726, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.0514909066259861, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04884728044271469, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.035398613661527634, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.026860792189836502, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.026096392422914505, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01952613703906536, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01750670187175274, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01849517971277237, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01827874779701233, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.015224010683596134, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011802817694842815, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16085699200630188, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.57.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2879258096218109, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2687739431858063, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.26287275552749634, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23709478974342346, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13666082918643951, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12934914231300354, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15162858366966248, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1404518485069275, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13860982656478882, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.1212870180606842, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11498161405324936, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0773400366306305, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06732282042503357, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06559853255748749, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06519310176372528, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0386584997177124, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.033778116106987, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.033662792295217514, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030640028417110443, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.030563538894057274, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020280834287405014, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020420486107468605, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019722266122698784, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013568629510700703, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15162858366966248, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.57.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.3095688223838806, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2890164852142334, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.28270554542541504, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2550109028816223, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.14674364030361176, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13889217376708984, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16273972392082214, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.15082786977291107, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14884482324123383, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.1302647739648819, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.12337782233953476, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08294613659381866, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.07217799127101898, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.07032519578933716, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06988613307476044, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.04141473397612572, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03584926947951317, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03572378680109978, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.032430678606033325, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03234463930130005, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.021489756181836128, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.021131081506609917, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.020885083824396133, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01330869272351265, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16273972392082214, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.57.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.27919715642929077, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.258494108915329, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.25078117847442627, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.2252236008644104, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12766261398792267, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.12070328742265701, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.14421381056308746, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.13330921530723572, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.12963703274726868, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.11417431384325027, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.10779912024736404, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.07264336198568344, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.06321225315332413, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.060764387249946594, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.06018001586198807, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03629206120967865, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.031025057658553123, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.030813638120889664, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02835877425968647, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02804422378540039, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.019014181569218636, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01857035979628563, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.018251119181513786, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.01183289010077715, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.2252236008644104, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.58.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1150948628783226, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.10630877315998077, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10130202025175095, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09059818089008331, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.05417976528406143, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04960981383919716, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06447523832321167, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.05977010726928711, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.05510874465107918, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04756699502468109, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.045216117054224014, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.032751139253377914, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02875429391860962, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.026216596364974976, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.025595681741833687, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.016485214233398438, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.013836551457643509, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.013545593246817589, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.012544671073555946, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01222930196672678, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.008899044245481491, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009152431972324848, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008159305900335312, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0064771343022584915, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1150948628783226, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.58.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08613157272338867, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07959123700857162, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07388315349817276, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06608542799949646, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.040323104709386826, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.035990048199892044, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.05203443020582199, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04752948880195618, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.04102770984172821, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03557409346103668, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.034298740327358246, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.026235593482851982, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.022788098081946373, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.019628241658210754, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01880098693072796, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.013266620226204395, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.010402869433164597, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.009970840066671371, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.00947384350001812, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.00900029856711626, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.007220299448817968, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.007159646600484848, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00626240624114871, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004955459386110306, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08613157272338867, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.58.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.28540685772895813, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.26450663805007935, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.25762447714805603, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2308371663093567, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1349107027053833, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.1266350895166397, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1519603580236435, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.1397707462310791, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.13702994585037231, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.1184573546051979, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.11211609095335007, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07729465514421463, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.066816046833992, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0645759105682373, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06403789669275284, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.038553208112716675, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.03272804617881775, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03253697603940964, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.029278842732310295, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.029127631336450577, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.019742220640182495, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01908148266375065, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.018857441842556, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011498775333166122, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2308371663093567, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.58.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2197347730398178, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12911784648895264, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.09023956209421158, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10147573798894882, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.09547233581542969, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.055878590792417526, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.12181707471609116, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.11107665300369263, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.10413116961717606, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05775703489780426, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.060108482837677, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.06240224838256836, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.05423401668667793, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04729573428630829, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04551875591278076, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03194329887628555, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02686305157840252, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02644979953765869, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02068405970931053, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.019490662962198257, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01863761804997921, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.020190833136439323, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.016666186973452568, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.016258658841252327, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2197347730398178, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.58.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.28768613934516907, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2686142027378082, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2627065181732178, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23698723316192627, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13657832145690918, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12926557660102844, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15161775052547455, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14035646617412567, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13852307200431824, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12119796127080917, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11493077129125595, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07733391970396042, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06728250533342361, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06555461883544922, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06514552235603333, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03867021203041077, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.033751264214515686, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03363781422376633, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03061934933066368, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.030537214130163193, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020297931507229805, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020396187901496887, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019747652113437653, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013541742227971554, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15161775052547455, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.58.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.309283584356308, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.28872138261795044, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.28239569067955017, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.25479426980018616, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.14663909375667572, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13878503441810608, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16264721751213074, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.15072236955165863, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14870695769786835, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.13012999296188354, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.12332583963871002, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08294545859098434, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0721210315823555, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.07026279717683792, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06982454657554626, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.04140440374612808, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03582827374339104, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03570205345749855, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.032402895390987396, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.032318562269210815, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.02149086631834507, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.021115221083164215, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.02088525891304016, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013305612839758396, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16264721751213074, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.58.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2767614722251892, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2562251687049866, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.24822597205638885, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.2229330837726593, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12649665772914886, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.11948146671056747, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.14341124892234802, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.13257944583892822, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.12849761545658112, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.11314459145069122, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.10681579262018204, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.07206335663795471, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.06284224987030029, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.06026970222592354, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.05964798480272293, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03591924160718918, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.030865628272294998, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.03062766045331955, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02823217771947384, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02789158746600151, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01863708533346653, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01863737776875496, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.017814038321375847, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.012035511434078217, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.2229330837726593, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.59.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12030623853206635, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11110960692167282, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10586576163768768, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09470243006944656, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.05674140900373459, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05198001489043236, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06784514337778091, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06252787262201309, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.05771419405937195, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0498230867087841, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.047512564808130264, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03464479744434357, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.030172115191817284, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.02754606492817402, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.02690986357629299, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01754879206418991, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.014657177031040192, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.014358367770910263, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.013315467163920403, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.012995134107768536, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009587118402123451, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009793956764042377, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00883211474865675, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.007061656564474106, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12030623853206635, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.59.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08346166461706161, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07706935703754425, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07142344117164612, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06388263404369354, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.0390470065176487, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03478896617889404, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.05039145052433014, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04612524434924126, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03974459320306778, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03442564234137535, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03319602832198143, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02544315531849861, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02209506928920746, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01894417218863964, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.018137319013476372, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01279541477560997, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.010001183487474918, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.009592934511601925, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.00909166969358921, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.008644228801131248, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.006912918295711279, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.006842080038040876, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005969653837382793, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004699801094830036, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08346166461706161, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.59.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.275991827249527, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2553529143333435, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.24871192872524261, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.22254879772663116, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.13014046847820282, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12207207828760147, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.14664340019226074, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.134872704744339, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.13228558003902435, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11409061402082443, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10782817006111145, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07448508590459824, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0644240751862526, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06223580613732338, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06171400099992752, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03715774416923523, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.031484149396419525, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03131301701068878, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02812102437019348, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02796645648777485, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.019004827365279198, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01828855648636818, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.018146032467484474, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0109094874933362, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.22254879772663116, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.59.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.21985235810279846, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1535825878381729, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12480711936950684, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11403492093086243, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0955110639333725, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06343815475702286, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.12551501393318176, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.11485624313354492, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.10361400246620178, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06578812003135681, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06515524536371231, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.06417714059352875, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.05525728315114975, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04665258899331093, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04442063346505165, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03243452310562134, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02542806789278984, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.024696242064237595, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.020100219175219536, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.018611954525113106, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.018158946186304092, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.018411308526992798, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01562376506626606, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01363285817205906, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.21985235810279846, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.59.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.28448012471199036, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2655210793018341, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.25968536734580994, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23427842557430267, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13510237634181976, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1278517246246338, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14993876218795776, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13885986804962158, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13701669871807098, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11988978087902069, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11362102627754211, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07647432386875153, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0665619969367981, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06485320627689362, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06445188820362091, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03822599723935127, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03338712453842163, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.033269234001636505, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030278783291578293, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03019915707409382, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.02003576047718525, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.0201672725379467, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01948630064725876, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013375870883464813, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14993876218795776, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.59.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.30941882729530334, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2888171672821045, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2824772596359253, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2548566162586212, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.14672617614269257, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13884267210960388, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16264794766902924, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.15079829096794128, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14882266521453857, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.13017074763774872, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.12331126630306244, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08288988471031189, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.07215303927659988, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0702960267663002, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06985113024711609, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.041366517543792725, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03582204505801201, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03569553419947624, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03238958865404129, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03230443596839905, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.02141641080379486, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.021088317036628723, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.02080695517361164, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013250917196273804, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16264794766902924, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.59.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2741503119468689, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.25382280349731445, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.24603497982025146, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.22113077342510223, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12535423040390015, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.11843372136354446, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.14178887009620667, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.13121411204338074, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.12732340395450592, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.11218581348657608, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.10593554377555847, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.07136908918619156, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.062273744493722916, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05970486253499985, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.05909692868590355, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03561495989561081, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.030559077858924866, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.030324457213282585, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.027965858578681946, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.027624255046248436, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.018526215106248856, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.018425041809678078, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.017718490213155746, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.01185454148799181, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.22113077342510223, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.60.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08572409301996231, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07876484841108322, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07349584996700287, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06562970578670502, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04029212146997452, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.036041002720594406, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.050476327538490295, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04668784141540527, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.04110146313905716, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03530523180961609, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.033707670867443085, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.025564854964613914, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02244376949965954, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.019575759768486023, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.018843311816453934, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.012923766858875751, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.010447906330227852, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.010094988159835339, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.009483917616307735, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.009091424755752087, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.007110069040209055, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.007205159869045019, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.006305279675871134, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.005149334669113159, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08572409301996231, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.60.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.04932940751314163, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.04543961212038994, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.03900879621505737, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.034925974905490875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.02294738031923771, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.018762752413749695, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03434322774410248, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.031191609799861908, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.02343510277569294, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02043977566063404, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02010861411690712, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.017211688682436943, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.014991038478910923, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.011297308839857578, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.010234340094029903, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.008776177652180195, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.00619848957285285, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.005655682645738125, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.005734565667808056, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.005135344807058573, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004963059443980455, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004793867468833923, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.003949368372559547, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003406223841011524, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.04932940751314163, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.60.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2981300354003906, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2754373252391815, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2678549587726593, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.23947583138942719, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.14084501564502716, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.13190920650959015, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.15916219353675842, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.1462278813123703, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.14333905279636383, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.1231115534901619, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.11629520356655121, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.08098790049552917, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.069942407310009, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06743824481964111, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06683681905269623, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.04040759801864624, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.034123580902814865, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03394170477986336, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.03042755275964737, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.030228836461901665, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.020664909854531288, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.019884679466485977, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.019659848883748055, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011867335066199303, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.15916219353675842, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.60.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2362430989742279, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1766476035118103, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.14311760663986206, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.12566711008548737, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.10533824563026428, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0748518779873848, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.14435157179832458, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.13258546590805054, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.11239438503980637, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07481511682271957, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.07417315989732742, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07378235459327698, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06362670660018921, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.05129594728350639, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04796614497900009, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03718535602092743, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.027509305626153946, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.026095997542142868, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.021818745881319046, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.019536644220352173, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.02031276933848858, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01992507465183735, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01652185246348381, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.013794670812785625, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1766476035118103, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.60.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2868174612522125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.267691433429718, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.261759489774704, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23614603281021118, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13622991740703583, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1288873255252838, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15121352672576904, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14002281427383423, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1381525844335556, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12084659188985825, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11456645280122757, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0770733430981636, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06712453812360764, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06539413332939148, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06498522311449051, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03856821730732918, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.033661358058452606, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.033543363213539124, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03052445501089096, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.030440455302596092, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020250948145985603, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020325232297182083, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019694628193974495, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013463491573929787, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15121352672576904, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.60.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.31053608655929565, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2898089587688446, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2834338843822479, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.25571209192276, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.14725537598133087, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13934151828289032, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16325198113918304, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1513613611459732, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14936810731887817, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.13065896928310394, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.12374933809041977, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08322390168905258, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0724358931183815, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.07056058198213577, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.07012099772691727, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.04154017195105553, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03598383441567421, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03585805743932724, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03254081308841705, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03245021402835846, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.021513646468520164, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.02122075855731964, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.020899532362818718, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013382771983742714, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16325198113918304, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.60.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2776035666465759, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2567650079727173, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.2487722486257553, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.2234296351671219, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12682847678661346, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.1197592094540596, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.14380958676338196, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.13286708295345306, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.12887383997440338, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.11336600035429001, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.10710690915584564, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.07241629809141159, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.06298427283763885, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.06041870266199112, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.059806566685438156, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03620491176843643, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.03093394637107849, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.030695658177137375, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.028277315199375153, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02793877199292183, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01900428719818592, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.018660619854927063, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.018173880875110626, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.012031873688101768, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.2234296351671219, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.61.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.11572402715682983, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.10675597935914993, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10181167721748352, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09106902778148651, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.054609011858701706, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05006168410181999, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06491227447986603, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06003029644489288, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.05559360608458519, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.047874998301267624, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0454721599817276, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03297773376107216, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02888680063188076, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.026448508724570274, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.025860615074634552, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01660821959376335, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.013997464440762997, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.013720950111746788, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.012683747336268425, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.012380718253552914, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009011304937303066, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009279166348278522, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008289264515042305, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.00662369467318058, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.11572402715682983, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.61.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07655107975006104, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07063761353492737, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06518372893333435, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05831008031964302, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03590182214975357, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.031784962862730026, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04713691398501396, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04286283999681473, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03655397146940231, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03170260041952133, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.030631041154265404, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.023812666535377502, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.020581217482686043, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.017450938001275063, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01663392409682274, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01201346330344677, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.009250552393496037, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00884038396179676, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.008428395725786686, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007975192740559578, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.006541253533214331, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.006407191511243582, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005606390070170164, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004421514924615622, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07655107975006104, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.61.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.29781800508499146, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2758943438529968, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2687808871269226, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.24059438705444336, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1407589167356491, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.13216935098171234, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.15827053785324097, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.1455891877412796, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.14303141832351685, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.1234440878033638, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.11669600009918213, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.08043276518583298, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06961306929588318, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06735372543334961, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06682570278644562, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.04013578221201897, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.034063920378685, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03390156850218773, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.030449653044342995, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.030285656452178955, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.02051899954676628, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.019773108884692192, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0196320042014122, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011789381504058838, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.15827053785324097, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.61.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2218725085258484, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.17148041725158691, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1447107493877411, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11610580235719681, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.09783822298049927, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07328624278306961, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.13387398421764374, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.12225557118654251, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.10601334273815155, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0701872780919075, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06812985986471176, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.06844618171453476, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.05879473313689232, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04783139377832413, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04490169510245323, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03463738039135933, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.025973672047257423, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.024881334975361824, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.020911039784550667, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.019085800275206566, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.019466649740934372, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.019025975838303566, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01624996028840542, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.013696364127099514, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2218725085258484, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.61.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2862136960029602, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.26706162095069885, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.26117444038391113, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23559610545635223, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13596157729625702, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1286182403564453, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15092085301876068, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13976873457431793, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13791708648204803, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12058880925178528, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11435391753911972, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07700280100107193, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06700710207223892, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06528031826019287, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06486751139163971, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03848807513713837, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.0336221344769001, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03350381925702095, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03048541583120823, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03040488436818123, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020186737179756165, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020334908738732338, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019628001376986504, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013518869876861572, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15092085301876068, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.61.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.3110913932323456, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2903127670288086, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2839556634426117, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.25609877705574036, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.14755553007125854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13960014283657074, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16360095143318176, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.15164121985435486, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1496528536081314, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.13087405264377594, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.12396825850009918, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08340585231781006, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.07256995141506195, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.07069901376962662, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0702497735619545, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0416298471391201, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03601182624697685, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.035886626690626144, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03254971653223038, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.0324619822204113, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.021552734076976776, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.021180856972932816, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.020938880741596222, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013278244063258171, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16360095143318176, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.61.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2807285487651825, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.259970098733902, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.25216493010520935, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.22660157084465027, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12829096615314484, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.12132659554481506, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.1448492705821991, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.13399051129817963, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.13028180599212646, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.11478982120752335, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.10838842391967773, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.07286687940359116, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.06342869251966476, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.06099319830536842, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.060415204614400864, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.036287881433963776, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.03101458214223385, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.03079691156744957, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02832484059035778, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.028006521984934807, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01874438486993313, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.018364788964390755, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.017954854294657707, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011401141993701458, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.22660157084465027, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.62.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12353222817182541, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11397966742515564, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10875651985406876, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09727513045072556, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.05826439708471298, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05346953496336937, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.0690612867474556, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06392186880111694, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.059311702847480774, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05108247324824333, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04848761484026909, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03508278355002403, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.030785400420427322, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.028259027749300003, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.027630729600787163, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.017688624560832977, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.014997053891420364, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.014707440510392189, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01359915267676115, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013285807333886623, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009601213969290257, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009987847879529, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00884917937219143, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0071962811052799225, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12353222817182541, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.62.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08322141319513321, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07676924765110016, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07119078189134598, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.0636049434542656, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.038969963788986206, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.034719668328762054, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.05035906657576561, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04592055454850197, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03971778228878975, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.034281112253665924, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03307799994945526, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02540997974574566, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.022018708288669586, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.018932079896330833, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.018151093274354935, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01278967596590519, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.010035376995801926, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00964404921978712, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.009122990071773529, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.008689816109836102, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.006940395571291447, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0069012828171253204, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.006018869578838348, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004805645905435085, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08322141319513321, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.62.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.29647189378738403, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.27457669377326965, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.26751846075057983, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.23939554393291473, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.14014023542404175, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.1315433531999588, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1577100157737732, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.14499853551387787, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.14240238070487976, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.12291845679283142, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.11609508097171783, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.08011654019355774, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06930402666330338, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06708571314811707, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06654152274131775, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03997819498181343, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.03390112891793251, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03373098745942116, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.03029475174844265, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.030154529958963394, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.020436767488718033, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.019647259265184402, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.019546043127775192, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011671779677271843, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1577100157737732, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.62.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.22566451132297516, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16197878122329712, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12958939373493195, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11584968119859695, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.09263493120670319, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06403519958257675, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1348852515220642, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.1233435720205307, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.10730422288179398, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07113762944936752, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.07068182528018951, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0691937804222107, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.059229351580142975, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.045414362102746964, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04155683144927025, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03480639308691025, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.024843936786055565, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02368275821208954, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.020937586203217506, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01849379763007164, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01913757063448429, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.018946316093206406, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.014915538020431995, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.013450554572045803, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.22566451132297516, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.62.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2859615683555603, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2668083608150482, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.26091083884239197, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2352733612060547, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1358737200498581, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1285453736782074, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15087300539016724, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13969294726848602, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13783897459506989, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12050656229257584, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11427264660596848, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0770009383559227, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06699371337890625, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06526515632867813, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0648578330874443, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.038513366132974625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03367117419838905, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03355273976922035, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030545013025403023, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.030463065952062607, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020272573456168175, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020455116406083107, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019717657938599586, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013708575628697872, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15087300539016724, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.62.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.31144165992736816, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2906072437763214, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2842356860637665, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.25632449984550476, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.14772573113441467, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13975024223327637, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16374099254608154, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.15184080600738525, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14983566105365753, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.13100093603134155, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.12408334016799927, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08345726132392883, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.07264707237482071, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.07077000290155411, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.07032323628664017, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.04165315255522728, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03606681153178215, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03593924641609192, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.032599132508039474, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03251132369041443, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.02154557593166828, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.021231425926089287, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.020928112789988518, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013333865441381931, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16374099254608154, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.62.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.27868878841400146, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2578515410423279, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.24978217482566833, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.22440072894096375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12738212943077087, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.1202949583530426, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.14461497962474823, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.13344348967075348, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.12941007316112518, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.11392594873905182, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.10757116228342056, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.07258367538452148, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.06328804790973663, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.060708675533533096, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.06008777767419815, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03616247698664665, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.03113584592938423, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.03089822828769684, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02848704904317856, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02814664877951145, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.018717141821980476, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01885954663157463, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.017866333946585655, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.012267074547708035, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.22440072894096375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.63.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.11269523203372955, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1041511595249176, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.09879195690155029, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0884903147816658, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.05322486162185669, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04853072389960289, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06480662524700165, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.05941447988152504, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.05415932461619377, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04682213068008423, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04475928097963333, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03305388242006302, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02872762829065323, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.02590716816484928, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.02520902454853058, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.016804201528429985, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.013859020546078682, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.013522359542548656, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.012622321024537086, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.012259691953659058, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009281773120164871, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009395573288202286, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008470174856483936, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006811784114688635, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.11269523203372955, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.63.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07762976735830307, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07164102047681808, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06612175703048706, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05920257791876793, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.0364844873547554, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.032281357795000076, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04776252061128616, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04340464621782303, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.0371161587536335, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.032157544046640396, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03115718439221382, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.024198288097977638, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.020902026444673538, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01774117350578308, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.016914477571845055, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.012235056608915329, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.009408538229763508, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0089810686185956, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.008570271544158459, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.008099275641143322, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0066790771670639515, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.006526164710521698, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0057409354485571384, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004506292287260294, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07762976735830307, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.63.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2854895293712616, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2642478942871094, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2573316693305969, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.23024819791316986, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.13503822684288025, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12661825120449066, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.15198858082294464, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.14000609517097473, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.13720501959323883, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11836548894643784, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.11179541796445847, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07733840495347977, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06699591130018234, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06469018012285233, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06414884328842163, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03860665112733841, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.03282161429524422, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03264089673757553, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.029338661581277847, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.029171999543905258, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.019808363169431686, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.019250430166721344, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01893155463039875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011726455762982368, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.23024819791316986, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.63.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.21633067727088928, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14795726537704468, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11226003617048264, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10011059045791626, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08885987102985382, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05690888687968254, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.12991976737976074, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.11920134723186493, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.10323207080364227, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06476445496082306, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06537502259016037, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.06688498705625534, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.05822567269206047, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04488902539014816, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04114852100610733, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03433970361948013, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.026519110426306725, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.025479953736066818, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.022662313655018806, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.02053435891866684, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.019754312932491302, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.021735290065407753, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01586718112230301, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.017532851547002792, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.21633067727088928, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.63.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.28767186403274536, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2684115469455719, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.26250648498535156, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2367849051952362, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13670234382152557, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12932142615318298, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15196821093559265, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1405351758003235, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13867713510990143, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12122602015733719, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.1149398684501648, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07758677750825882, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06741920858621597, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06568334251642227, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0652727410197258, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03883989900350571, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.033921051770448685, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03380551561713219, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030782025307416916, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03069976717233658, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020563652738928795, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020655682310461998, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.020001886412501335, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01391175203025341, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15196821093559265, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.63.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.311556339263916, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.290743887424469, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.28437888622283936, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2565128207206726, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.14777930080890656, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13981561362743378, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16398751735687256, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.15189363062381744, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14990025758743286, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.13108408451080322, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.12416701763868332, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08364587277173996, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.07270918786525726, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0708293542265892, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.07038136571645737, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.041754577308893204, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03611911088228226, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03599217161536217, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03266023099422455, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03257150202989578, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.021682851016521454, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.021299991756677628, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.021068045869469643, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013431797735393047, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16398751735687256, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.63.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.27874955534935, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.25792431831359863, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.24997346103191376, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.22444981336593628, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12743273377418518, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.12035352736711502, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.14423811435699463, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.13341784477233887, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.1294512152671814, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.11393170058727264, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.10753986984491348, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.0726052075624466, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.0633407011628151, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.06075560674071312, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.060136064887046814, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03629447892308235, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.0312015600502491, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.030966252088546753, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.028554227203130722, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.028212303295731544, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.018986299633979797, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.018963346257805824, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01815377175807953, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.012422334402799606, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.22444981336593628, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.64.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12220434099435806, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11288151144981384, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10780563950538635, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09661616384983063, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.05763024836778641, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05297897383570671, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06819562613964081, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06320039182901382, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.058634888380765915, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.050672516226768494, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04813145846128464, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03465288504958153, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.030380917713046074, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.027897793799638748, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.027287017554044724, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.017449375241994858, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01473549660295248, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.014453931711614132, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01337253674864769, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013071347028017044, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.00943868886679411, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00972731038928032, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008705580607056618, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0069215549156069756, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12220434099435806, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.64.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08736513555049896, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.080778107047081, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.07535260915756226, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.0674099326133728, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.04086781665682793, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03668033704161644, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.05194113031029701, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04760527238249779, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.04159091040492058, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03604567050933838, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03464018553495407, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.026201045140624046, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.022743988782167435, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.019758455455303192, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.018998561426997185, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.013152172788977623, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.01034805178642273, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.009948859922587872, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.00939082633703947, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.008956884033977985, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.007052745670080185, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0069345757365226746, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.006119328085333109, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004635717254132032, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08736513555049896, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.64.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2885989546775818, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.26757195591926575, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.26085853576660156, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2338254600763321, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.13634777069091797, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12815625965595245, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.15320032835006714, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.14105622470378876, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.13850723206996918, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11986687034368515, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.11331365257501602, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07783807814121246, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06738188862800598, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06520488858222961, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.0646897479891777, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03882797062397003, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.032958827912807465, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.032795943319797516, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.029494263231754303, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02935113199055195, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01984400674700737, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.019064879044890404, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.018983542919158936, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011285392567515373, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.15320032835006714, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.64.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.24156953394412994, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16102637350559235, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12077782303094864, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11260829865932465, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.098879873752594, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.058802250772714615, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1427653282880783, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.13059699535369873, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.1145368292927742, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0704304650425911, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.07132396847009659, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0730694830417633, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06270898878574371, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04852869361639023, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04457659274339676, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.036673881113529205, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02662568911910057, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02548968233168125, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.021548734977841377, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.018920961767435074, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01986709050834179, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.020289761945605278, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01532200537621975, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.014590874314308167, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16102637350559235, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.64.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.28432580828666687, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2652533948421478, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2593832314014435, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23392310738563538, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13514377176761627, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12779594957828522, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1500435769557953, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1389234960079193, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13706208765506744, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.1197950690984726, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11355191469192505, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07658899575471878, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06663601100444794, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06491174548864365, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06450840830802917, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03829726576805115, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.033481549471616745, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03336995840072632, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030368097126483917, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.030289355665445328, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.02015678584575653, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020336046814918518, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01959848962724209, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013629606924951077, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1500435769557953, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.64.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.3098642826080322, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2890971899032593, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2827838063240051, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.25498560070991516, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.14702872931957245, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1390688419342041, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16311466693878174, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.15114106237888336, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14914768934249878, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.1303972452878952, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.12351972609758377, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08316104859113693, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.07235734164714813, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0704803466796875, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.07003691792488098, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0415487214922905, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03597301244735718, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03584614396095276, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.032526880502700806, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.032440733164548874, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.021612146869301796, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.02126234769821167, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.021004868671298027, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013471897691488266, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16311466693878174, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.64.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2749439477920532, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2540750801563263, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.24589316546916962, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.22078728675842285, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.1255464404821396, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.11837509274482727, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.14276337623596191, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.1318930685520172, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.12760919332504272, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.11214552074670792, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.10598602890968323, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.07190457731485367, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.06249689310789108, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05977185443043709, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.059119079262018204, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03592044115066528, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.030539102852344513, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.030284002423286438, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02789117768406868, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02752433903515339, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.018792925402522087, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01836780458688736, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.017904410138726234, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011692628264427185, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.22078728675842285, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.65.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.10401985794305801, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.09577327221632004, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.09067945927381516, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08099115639925003, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04904898628592491, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04456740990281105, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.059306107461452484, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0548701249063015, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.049968499690294266, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.042932312935590744, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.040804795920848846, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.030113602057099342, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02638029307126999, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.023748785257339478, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.02309497259557247, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.015174401924014091, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.012547716498374939, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01223763171583414, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.011352520436048508, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01100876647979021, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.008219128474593163, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.008364902809262276, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.007441922090947628, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.005891431123018265, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.10401985794305801, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.65.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06474306434392929, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.05947691202163696, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05376068502664566, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.048003315925598145, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.030184349045157433, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02615339681506157, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04127034917473793, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.037417780607938766, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03081454336643219, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.026562152430415154, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.025787455961108208, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02077111043035984, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.017922230064868927, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014743964187800884, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.013883532956242561, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.010521678254008293, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007906528189778328, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007456100080162287, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007214622106403112, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006714399438351393, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005812785588204861, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0056733316741883755, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004866527859121561, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0039581675082445145, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06474306434392929, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.65.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.29186469316482544, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2698037922382355, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2627289593219757, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.23475566506385803, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.13784581422805786, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12926946580410004, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.15502296388149261, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.14275068044662476, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.14016763865947723, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.12060703337192535, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.11377167701721191, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07881328463554382, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06826742738485336, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06599460542201996, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06546242535114288, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0393381230533123, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.03343822807073593, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03326992318034172, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02982516400516033, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02967977523803711, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.02015260048210621, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01949234865605831, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01926538534462452, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011763926595449448, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.15502296388149261, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.65.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2568970024585724, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.18233898282051086, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.15061187744140625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.12623251974582672, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.11308486014604568, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.08042117953300476, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.14579547941684723, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.13372588157653809, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.12284445762634277, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07381200045347214, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0728331208229065, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07448969781398773, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0640864372253418, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.05473915487527847, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.05231948569417, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03735225275158882, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02880295179784298, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.028065793216228485, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0213087797164917, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01958901435136795, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.020000509917736053, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.019633110612630844, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01705346442759037, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.013256574980914593, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.18233898282051086, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.65.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.28680554032325745, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.26756519079208374, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2616947591304779, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23600800335407257, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13631179928779602, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12894462049007416, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1512511670589447, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1401292383670807, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1382693499326706, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.1208522766828537, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11450938135385513, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07716164737939835, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0672033429145813, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06547168642282486, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06506209820508957, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03857248276472092, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03374071791768074, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03362567350268364, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030594680458307266, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.030515171587467194, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.02022673934698105, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.02044750563800335, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019663359969854355, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013645080849528313, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1512511670589447, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.65.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.31013253331184387, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2894205152988434, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.28303930163383484, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.255231648683548, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.14719854295253754, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13925163447856903, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1633714735507965, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1513116955757141, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14929905533790588, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.13054132461547852, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.12367302179336548, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08336414396762848, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.07244914025068283, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.07058286666870117, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.07013864815235138, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.04165034741163254, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03607960417866707, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03595355525612831, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03263477236032486, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.032550010830163956, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.02172793447971344, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.02140044793486595, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.021118154749274254, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.0136808380484581, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1633714735507965, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.65.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.275871217250824, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.25484374165534973, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.24671953916549683, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.2212778925895691, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12594999372959137, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.118798166513443, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.14288583397865295, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.132071852684021, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.1280323714017868, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.11237543821334839, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.10604558140039444, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.0718943178653717, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.06251978874206543, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05990797281265259, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.05928264558315277, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03580121695995331, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.030499812215566635, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.030261583626270294, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02779986523091793, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02745066024363041, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01852821558713913, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.018164480105042458, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.017679816111922264, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011352147907018661, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.2212778925895691, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.66.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.11262901872396469, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1041533574461937, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0987737625837326, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08839038014411926, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.05314573273062706, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04851432889699936, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06415338814258575, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.05936254933476448, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0540907084941864, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04672976955771446, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.044474683701992035, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03255469352006912, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02854089066386223, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.025768505409359932, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.025079596787691116, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.016396017745137215, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.013631947338581085, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.013302967883646488, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.012375977821648121, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.012017779052257538, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.008864074014127254, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009103335440158844, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008039381355047226, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006446653977036476, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.11262901872396469, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.66.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07249224931001663, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06722061336040497, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.061396725475788116, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.054925862699747086, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.034064508974552155, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02990555763244629, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04549401253461838, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04160391166806221, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.034674908965826035, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03020431287586689, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.029279207810759544, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.023017071187496185, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.019936230033636093, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01657734252512455, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.015683932229876518, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01158845704048872, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.008776784874498844, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008302495814859867, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0080171599984169, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007501394022256136, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.006239882204681635, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.006143428385257721, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005213452968746424, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00416194461286068, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07249224931001663, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.66.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2968151867389679, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.27502816915512085, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.26791173219680786, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.23989087343215942, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1403748095035553, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.131829172372818, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.15795879065990448, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.14536425471305847, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.14257460832595825, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.12313959747552872, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.1162928119301796, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.08033297210931778, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06945089250802994, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06717769801616669, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06661617755889893, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.04008909687399864, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.033979497849941254, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0338030643761158, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.030364342033863068, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.030212419107556343, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.020484255626797676, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.019703516736626625, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.019579440355300903, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011714685708284378, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.15795879065990448, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.66.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.24803949892520905, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.17538368701934814, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1455221176147461, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1262890100479126, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.10498959571123123, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07275702804327011, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.139091357588768, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.12793776392936707, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.11775527149438858, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07402969896793365, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.07244297116994858, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07115225493907928, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06133463233709335, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.05100155249238014, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04827537387609482, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03569970652461052, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.027186056599020958, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.026490893214941025, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02137027680873871, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01953502930700779, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.019136572256684303, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.019203094765543938, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.015817968174815178, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0133051797747612, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.17538368701934814, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.66.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.28776371479034424, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.268478661775589, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2625061869621277, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23671895265579224, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13681218028068542, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12939542531967163, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1518903225660324, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1406436711549759, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13879193365573883, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12127307057380676, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11500930786132812, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0775289386510849, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06749911606311798, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06575789302587509, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06534641981124878, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03877630829811096, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.033978234976530075, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03385932371020317, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030826546251773834, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.030746297910809517, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020415738224983215, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020723039284348488, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019857926294207573, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013993509113788605, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1518903225660324, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.66.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.3092322051525116, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.28847458958625793, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2821086049079895, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.25431957840919495, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.14673873782157898, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13881826400756836, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16282027959823608, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.15082868933677673, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1488489806652069, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.13008178770542145, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.12322442978620529, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08304033428430557, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.07221417129039764, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.07035655528306961, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06991136074066162, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.04145466908812523, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.035913415253162384, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.035786811262369156, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.032465845346450806, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.032377585768699646, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.021527182310819626, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.02123652771115303, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.020917214453220367, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013473615981638432, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16282027959823608, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.66.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.26752588152885437, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2470085471868515, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.23901152610778809, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.2144407480955124, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12216732650995255, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.11512839794158936, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.13865722715854645, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.12824590504169464, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.12419254332780838, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.10898604989051819, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.10284365713596344, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06977824121713638, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.060815852135419846, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.058204278349876404, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.05758211389183998, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.034863270819187164, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.029827000573277473, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02958332560956478, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.027233432978391647, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02688363939523697, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.018230054527521133, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.018055956810712814, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01739432103931904, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011677604168653488, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.2144407480955124, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.67.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0891561508178711, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.08276621997356415, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07628058642148972, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0684899240732193, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04252439737319946, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.037665508687496185, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.05550737679004669, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.05079019442200661, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.04325881972908974, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03756999969482422, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.036478716880083084, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.028391186147928238, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.024649541825056076, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.020973574370145798, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.019998053088784218, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.014583933167159557, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.011595805175602436, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.011116167530417442, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.010678107850253582, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.010157402604818344, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.008354765363037586, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.008528105914592743, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.007326540071517229, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006509040016680956, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0891561508178711, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.67.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06124389171600342, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.05676637962460518, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05044756084680557, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.04519791901111603, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.028939969837665558, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.024598268792033195, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.040292080491781235, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03716576471924782, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.029459567740559578, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02563818171620369, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.025034040212631226, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.020411383360624313, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.017881784588098526, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01415948010981083, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.013129888102412224, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.010396378114819527, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007591302506625652, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007048578467220068, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006953907199203968, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006363231223076582, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005797164514660835, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005562438629567623, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0047827837988734245, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003812272334471345, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06124389171600342, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.67.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2713235914707184, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.24982033669948578, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.24270537495613098, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2160821557044983, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1278417706489563, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.11932112276554108, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.14476686716079712, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.13288819789886475, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.13010668754577637, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11121472716331482, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10495848953723907, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07374883443117142, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06360246241092682, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06126910075545311, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06070113182067871, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03687819465994835, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.031123390421271324, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.030932754278182983, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.027653483673930168, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02747403457760811, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.018898501992225647, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.018306512385606766, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.017936717718839645, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011193117126822472, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2160821557044983, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.67.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1370086371898651, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1042231023311615, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0788583904504776, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0660802498459816, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.05672396719455719, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.039024919271469116, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09321063756942749, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08473828434944153, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06607697159051895, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.047483865171670914, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04776899144053459, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04886161908507347, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04259302467107773, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03043997287750244, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.02689172700047493, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.026441901922225952, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02031087875366211, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.019195176661014557, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.019052650779485703, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.017401939257979393, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01721678115427494, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.018454337492585182, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01437597256153822, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.016030211001634598, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1370086371898651, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.67.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2827201187610626, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.26374688744544983, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2578694224357605, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23255102336406708, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13440310955047607, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12712633609771729, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14935413002967834, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13815546035766602, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1363406926393509, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11914496123790741, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11296749114990234, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07634316384792328, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06632545590400696, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0646151453256607, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06421278417110443, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03821491822600365, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03343156725168228, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03331923112273216, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030345359817147255, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03026903234422207, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020275436341762543, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.0204453207552433, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01973246969282627, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013883174397051334, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14935413002967834, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.67.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.3087863326072693, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2881038784980774, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2817194163799286, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2540605962276459, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.14657241106033325, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13864582777023315, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16263459622859955, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.15063004195690155, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14866137504577637, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12996776401996613, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.12312977015972137, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08301616460084915, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.07217536866664886, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.07030410319566727, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06986687332391739, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.041460998356342316, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03596358001232147, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.035840027034282684, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.032540179789066315, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03245581313967705, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.021631265059113503, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.021382933482527733, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.021024459972977638, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013738620094954967, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16263459622859955, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.67.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2736017107963562, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.25257521867752075, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.2447250485420227, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.2194739282131195, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12491324543952942, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.11787547916173935, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.14167529344558716, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.13074897229671478, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.126998171210289, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.11140594631433487, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.10504916310310364, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.07125013321638107, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.061947889626026154, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05946188047528267, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.058864519000053406, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.035589106380939484, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.030364152044057846, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.030148450285196304, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.027690649032592773, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.027361182495951653, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.018646951764822006, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01819978840649128, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.017851442098617554, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011580404825508595, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.2194739282131195, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.68.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1312934309244156, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12147044390439987, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11686758697032928, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10464057326316833, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.062049202620983124, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.057509638369083405, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07208657264709473, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06666556745767593, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.063067726790905, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.054489172995090485, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.051637664437294006, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.036646630614995956, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03200614079833031, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.02994311787188053, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.02944008633494377, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.018388960510492325, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.015647901222109795, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.015425145626068115, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01414050068706274, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013907653279602528, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009795370511710644, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.010012137703597546, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009155124425888062, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.00692744692787528, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1312934309244156, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.68.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10684342682361603, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.09873140603303909, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.09389762580394745, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.08394943922758102, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.050210561603307724, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04594986513257027, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.060874853283166885, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05571190267801285, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05102771148085594, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.044144243001937866, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.042114514857530594, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.030791200697422028, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0266704298555851, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02422531507909298, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.023598751053214073, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.015454297885298729, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.012582231312990189, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.012280315160751343, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.011355113238096237, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.011034082621335983, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008188005536794662, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008080578409135342, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007404771167784929, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0053665838204324245, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10684342682361603, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.68.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2915981709957123, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2702966630458832, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.26345789432525635, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.23573212325572968, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.13773731887340546, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.1294582039117813, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.15421779453754425, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.14227768778800964, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.13993680477142334, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.12084536254405975, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.1139754205942154, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07839461416006088, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0680418387055397, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06593392789363861, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.0654485821723938, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03913979232311249, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.033507123589515686, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03334708511829376, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02997608296573162, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.029850177466869354, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.020107146352529526, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.019642192870378494, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.019321193918585777, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.012061158195137978, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.15421779453754425, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.68.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2562641203403473, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.18053032457828522, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.14322996139526367, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.12830738723278046, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.11175843328237534, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07447902113199234, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.15222370624542236, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.13877958059310913, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.12226078659296036, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07739228755235672, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.07684282213449478, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07819525897502899, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06698457151651382, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.054887011647224426, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.05163703113794327, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03945942968130112, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.030230794101953506, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.029131505638360977, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02415013685822487, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.022035248577594757, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.022005677223205566, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.022506102919578552, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01832050271332264, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01678069867193699, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.18053032457828522, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.68.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2811136543750763, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.262208491563797, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2563595771789551, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23121793568134308, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13366056978702545, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1264282464981079, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14857251942157745, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13741019368171692, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1355915516614914, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.1185191199183464, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11234615743160248, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07592593133449554, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06595419347286224, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06425587832927704, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0638522207736969, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.038008153438568115, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03324763476848602, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03313414379954338, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030177682638168335, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03010113351047039, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020173324272036552, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020332371816039085, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01963270828127861, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013806028291583061, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14857251942157745, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.68.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.30763906240463257, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.28699007630348206, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.28064921498298645, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.25313061475753784, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.14610013365745544, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13818331062793732, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16238759458065033, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.15016433596611023, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14816801249980927, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12951341271400452, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.12272756546735764, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08297815173864365, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.07196703553199768, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.07011585682630539, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06967531144618988, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.04154385253787041, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.035950545221567154, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03582567721605301, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03254307806491852, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03245756775140762, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.021914871409535408, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.02148834802210331, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.02132505737245083, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013965414837002754, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16238759458065033, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.68.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.27592942118644714, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2544609606266022, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.24626277387142181, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.2207624912261963, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12589102983474731, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.11861728876829147, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.1430438905954361, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.13206052780151367, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.12806423008441925, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.11214901506900787, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.105827197432518, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.07194193452596664, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.06258232891559601, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.059963077306747437, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.0593336820602417, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03586497902870178, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.030667850747704506, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.03043515980243683, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.027948861941695213, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.027598973363637924, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01867271214723587, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01846933364868164, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.017823874950408936, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011843171902000904, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.2207624912261963, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.69.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14726808667182922, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13653583824634552, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13170717656612396, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11804518103599548, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06968154013156891, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06481127440929413, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08132879436016083, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07433824241161346, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07080716639757156, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06136605516076088, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.058569855988025665, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04170139506459236, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0358440987765789, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.0337345227599144, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03322655335068703, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02113160490989685, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01767219789326191, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.017445476725697517, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01602291688323021, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01579098030924797, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.011297957971692085, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.011254262179136276, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.010610707104206085, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.00777813233435154, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14726808667182922, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.69.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1143532246351242, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.10590261220932007, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.10134077072143555, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09072668105363846, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.053846973925828934, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04965013265609741, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06415749341249466, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05878201127052307, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05471134930849075, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04739004373550415, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.045117657631635666, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.032433174550533295, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.028140122070908546, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02592364139854908, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.025376569479703903, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.016259247437119484, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.01338895782828331, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.01313028484582901, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.012069935910403728, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.011788654141128063, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008556991815567017, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008431758731603622, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007832846604287624, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005509101785719395, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1143532246351242, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.69.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2769654393196106, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.25693950057029724, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2505273222923279, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2242622971534729, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.13068456947803497, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12289288640022278, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.14628922939300537, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.13501335680484772, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.13273084163665771, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.1148226410150528, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10846050083637238, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07429296523332596, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0645267441868782, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06252501159906387, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06205487996339798, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03706125542521477, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.031628698110580444, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03147929906845093, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.028320081532001495, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.028199633583426476, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01897960901260376, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01836112141609192, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.018242312595248222, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010980487801134586, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2242622971534729, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.69.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.26295584440231323, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.18316157162189484, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.15260957181453705, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1352255493402481, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.11698338389396667, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.08126645535230637, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1446862518787384, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.13286355137825012, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.1251170039176941, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07505885511636734, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.07435362786054611, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07388801872730255, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06359458714723587, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.05638435482978821, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.05455850437283516, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.036948125809431076, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.029440417885780334, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.028926558792591095, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02137352153658867, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.020049620419740677, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.019469816237688065, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.019382787868380547, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01711881533265114, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.012968122027814388, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.18316157162189484, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.69.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.28187981247901917, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.26293012499809265, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.25715845823287964, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23196342587471008, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13404269516468048, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1268130987882614, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14892086386680603, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13779067993164062, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13596323132514954, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11888258159160614, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11269968003034592, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07611390948295593, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06616832315921783, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0644707903265953, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0640665739774704, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03812463581562042, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.0333876758813858, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03327437490224838, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030325565487146378, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.0302448533475399, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020244574174284935, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020469272509217262, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019704002887010574, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013958398252725601, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14892086386680603, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.69.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.3062233626842499, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.28568392992019653, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.27940794825553894, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2519659698009491, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.14540083706378937, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1375458985567093, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16139905154705048, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14946670830249786, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14749346673488617, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12892180681228638, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.12212871015071869, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08243700116872787, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.07160395383834839, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06976481527090073, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06933017075061798, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.04117608070373535, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03574410453438759, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03562121465802193, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03236255794763565, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.032275132834911346, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.021545715630054474, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.021338650956749916, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.02094685472548008, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013825057074427605, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.16139905154705048, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.69.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2704172432422638, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.24919980764389038, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.24089883267879486, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.2158665806055069, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12350236624479294, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.11611446738243103, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.14048317074775696, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.12995645403862, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.12565630674362183, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.10993362963199615, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.10362017154693604, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.07076580077409744, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.061728086322546005, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05890044942498207, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.05823080241680145, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.035296544432640076, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.030299164354801178, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.030039945617318153, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02763756364583969, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.027261126786470413, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01840047352015972, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.01853226311504841, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.017487358301877975, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.012189752422273159, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.2158665806055069, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.70.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12452439963817596, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11501089483499527, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11025090515613556, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09871017187833786, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.05865022912621498, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05416114628314972, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06931544095277786, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06363361328840256, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.059787169098854065, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05154429003596306, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04898705333471298, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.035251714289188385, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03055168129503727, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.028300439938902855, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.027752144262194633, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01769852824509144, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.014786820858716965, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.014546837657690048, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01337443944066763, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013105196878314018, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009427236393094063, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009508159011602402, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008679674938321114, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006540708709508181, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12452439963817596, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.70.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10469913482666016, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.09676994383335114, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.09217025339603424, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.08241045475006104, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.049183592200279236, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04505382850766182, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.05903268977999687, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05435160920023918, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05007851868867874, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.043259285390377045, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.041145700961351395, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.029931847006082535, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02605592831969261, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.023712532594799995, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.02313142642378807, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.015010782517492771, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.012293435633182526, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.01201525516808033, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.01109737902879715, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.010793136432766914, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.007945310324430466, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.007860582321882248, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00720361340790987, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0051896595396101475, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10469913482666016, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.70.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.27446791529655457, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.25449228286743164, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2479725182056427, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2221347987651825, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1294778436422348, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12173638492822647, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.14518380165100098, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.13408254086971283, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.13147957623004913, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11379318684339523, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.1074945405125618, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07378348708152771, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06407129764556885, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06195458397269249, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06145148724317551, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03682301938533783, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.03136499598622322, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.031201167032122612, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.028098968788981438, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.027956509962677956, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.018858913332223892, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.018261801451444626, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01808778941631317, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010985623113811016, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.2221347987651825, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.70.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.21633204817771912, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16739316284656525, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1465446799993515, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1122334748506546, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.09591206908226013, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07459425181150436, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1227741464972496, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.11250366270542145, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.10185688734054565, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06609916687011719, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.062300894409418106, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0628066286444664, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0539863146841526, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.046557843685150146, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04464216157793999, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03159615769982338, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.024979112669825554, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02428426221013069, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.019508659839630127, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01837235875427723, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01727813482284546, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.017505625262856483, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.014981647953391075, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.012624770402908325, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.21633204817771912, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.70.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2819560766220093, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.26303669810295105, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.25718382000923157, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2319934070110321, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13405932486057281, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1268133670091629, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1488851010799408, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13779279589653015, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13597312569618225, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11886310577392578, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11272601783275604, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07608490437269211, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0661371573805809, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06444233655929565, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06404278427362442, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03806164115667343, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.033313967287540436, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03320034220814705, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030236534774303436, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.030156491324305534, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.02009112387895584, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020339883863925934, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019542543217539787, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013771969825029373, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1488851010799408, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.70.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.3062562346458435, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.28574827313423157, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2794382870197296, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.25202009081840515, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.14545989036560059, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13760745525360107, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1614660918712616, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14948974549770355, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14754565060138702, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12894748151302338, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.12226730585098267, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08244980871677399, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.07163027673959732, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06979101896286011, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06935865432024002, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0412299819290638, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03577009588479996, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.0356491357088089, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03237885236740112, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03229362145066261, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.021621838212013245, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.021364586427807808, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.021027272567152977, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013862987980246544, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1614660918712616, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.70.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2690739035606384, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.24713093042373657, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.23839624226093292, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.21341313421726227, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12260384112596512, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.1149757131934166, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.14042344689369202, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.12946262955665588, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.1248965635895729, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.10880181938409805, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.10262826085090637, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.07071816921234131, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.06137041747570038, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.058438610285520554, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.057729657739400864, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.035290803760290146, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.029984191060066223, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.029717830941081047, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.027267970144748688, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.026872579008340836, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.018527960404753685, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.018267398700118065, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.017580876126885414, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011867765337228775, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.21341313421726227, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.71.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13367263972759247, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12370910495519638, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11879311501979828, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10627345740795135, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0630951076745987, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.058360129594802856, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07360024005174637, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06821179389953613, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06416187435388565, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05545833706855774, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05256354808807373, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03739885240793228, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.032722536474466324, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.030430637300014496, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.029874086380004883, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01875629462301731, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.015886733308434486, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.015631619840860367, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01436559297144413, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.014093623496592045, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009976527653634548, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.010174881666898727, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009281406179070473, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006987602915614843, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13367263972759247, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.71.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11594388633966446, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.10739153623580933, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.10244021564722061, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09160847216844559, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05461420118808746, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.05017947778105736, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.0653071328997612, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06013014540076256, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.055530160665512085, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04802664369344711, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04577217251062393, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.033087264746427536, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02883502095937729, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.026308653876185417, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.025689540430903435, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.016595397144556046, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.013627678155899048, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.013327471911907196, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.012300108559429646, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.011972488835453987, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008798130787909031, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008671559393405914, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.008020042441785336, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005695880390703678, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11594388633966446, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.71.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2840076982975006, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.26342737674713135, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.25691020488739014, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.23015907406806946, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.13426604866981506, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12632517516613007, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.15052174031734467, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.1387948840856552, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.13635507225990295, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11810898780822754, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.1115870401263237, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07653042674064636, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06639265269041061, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06427644938230515, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.063778355717659, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.038216304033994675, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.032562628388404846, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.032398659735918045, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.029189493507146835, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.029054541140794754, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01957099325954914, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0189571101218462, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.018782302737236023, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011432812549173832, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.23015907406806946, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.71.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2235497087240219, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16716815531253815, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.14577129483222961, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.12105284631252289, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.09922589361667633, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0751013457775116, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.12410762906074524, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.1133156567811966, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.10619688779115677, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06951368600130081, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06425270438194275, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.06368085741996765, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.054791487753391266, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04855472594499588, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04698668047785759, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.032195188105106354, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.026534181088209152, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.026095280423760414, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02119559608399868, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.020270617678761482, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.018048571422696114, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01885879784822464, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.016177034005522728, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.014361029490828514, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2235497087240219, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.71.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.27863675355911255, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.25998058915138245, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2542167007923126, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22925615310668945, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13247300684452057, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1253264844417572, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14714133739471436, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13617442548274994, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13439109921455383, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11745687574148178, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11133784055709839, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07512906193733215, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06533768028020859, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06366344541311264, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06326809525489807, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.037571653723716736, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.032855622470378876, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03274109587073326, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02979646623134613, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.029719658195972443, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01979650929570198, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019970158115029335, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019249893724918365, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013405010104179382, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22925615310668945, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.71.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.3015996813774109, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.28143277764320374, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.27524229884147644, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.24815590679645538, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1432347297668457, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13548675179481506, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15890462696552277, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14720416069030762, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.1452813744544983, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12698517739772797, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.12037903815507889, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08115445822477341, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.07050451636314392, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06870628148317337, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0682796984910965, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.04053613170981407, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.035154785960912704, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03503142297267914, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.031804539263248444, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03172391653060913, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.021168189123272896, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.02091081254184246, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.020577745512127876, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013443815521895885, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15890462696552277, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.71.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.26555052399635315, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.2440180629491806, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.23505116999149323, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.21054285764694214, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12099259346723557, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.11331279575824738, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.13904759287834167, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.1283671110868454, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.1232958436012268, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.10749203711748123, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.10150310397148132, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.07005930691957474, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.060834042727947235, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05767149105668068, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.056907933205366135, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.0349159762263298, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.029567880555987358, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.029267048463225365, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.026901407167315483, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.026471145451068878, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.0182280782610178, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.018024463206529617, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.017201988026499748, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011635735630989075, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.21054285764694214, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.72.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14396469295024872, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1334523856639862, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12859579920768738, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11522909253835678, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0680498406291008, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0631980150938034, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07860656082630157, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07286456972360611, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06914963573217392, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05989113077521324, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0567622035741806, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03995669260621071, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.034939832985401154, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03277799114584923, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03225734457373619, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.020027929916977882, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.017030000686645508, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01680062524974346, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015391145832836628, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0151509465649724, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.010597380809485912, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01074051670730114, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00994456373155117, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.007282067555934191, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14396469295024872, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.72.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12198624759912491, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.11307823657989502, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.108186274766922, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09685970842838287, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.057489991188049316, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.052976466715335846, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06806159019470215, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06283411383628845, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05842086672782898, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.050645604729652405, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04809479042887688, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03449292853474617, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.030122904106974602, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02765912003815174, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.027066756039857864, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.017280753701925278, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.014271187596023083, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.013975128531455994, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.012867293320596218, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.012561156414449215, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.009100187569856644, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008969388902187347, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.008344233967363834, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0058019994758069515, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12198624759912491, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.72.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.28887975215911865, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2686292827129364, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2620158791542053, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.23504677414894104, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.13661852478981018, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12863394618034363, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.153285950422287, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.14110702276229858, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1387905329465866, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.12036609649658203, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.11382845044136047, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07792415469884872, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06747333705425262, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0653948187828064, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06489074975252151, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.038877394050359726, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.03309761732816696, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03293425589799881, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02967752330005169, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02954411320388317, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.019892269745469093, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.019192779436707497, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.019097115844488144, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011497041210532188, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.153285950422287, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.72.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.23975394666194916, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.180545836687088, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1593715250492096, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.12618595361709595, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.10600270330905914, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.08336976915597916, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.13057473301887512, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.12008176743984222, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.11399281024932861, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0725899264216423, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06742628663778305, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.06704523414373398, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.05786227434873581, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.05154137685894966, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04995351657271385, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.033708930015563965, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.027717778459191322, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.027327777817845345, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.021572016179561615, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.02062315121293068, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01852772757411003, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.019162967801094055, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.016556207090616226, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.014097297564148903, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.180545836687088, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.72.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.27703869342803955, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.25848111510276794, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2527642846107483, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22785718739032745, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13178591430187225, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12465435266494751, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14633572101593018, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13544559478759766, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13368216156959534, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11685194075107574, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11077025532722473, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07479681074619293, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06503695994615555, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06336534023284912, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06297309696674347, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03741452842950821, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03279748558998108, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03269174322485924, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02977265976369381, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02969771809875965, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.019789498299360275, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.02007284201681614, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01924828253686428, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013660309836268425, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22785718739032745, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.72.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.30092477798461914, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.28074735403060913, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2745743691921234, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.24746714532375336, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1428951919078827, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13520847260951996, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15846878290176392, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.14685814082622528, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14496083557605743, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.1266910582780838, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11997389048337936, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.08088242262601852, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.07034505158662796, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06854885816574097, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06812039017677307, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0403972826898098, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.035096313804388046, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03497534617781639, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.031752247363328934, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03166967257857323, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.02106468752026558, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020906340330839157, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.0204777829349041, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01348668709397316, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15846878290176392, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.72.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.2647360563278198, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.24269530177116394, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.2335871309041977, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.20909424126148224, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.12044458091259003, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.11262993514537811, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.1386595368385315, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.12788374722003937, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.12285522371530533, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.10679652541875839, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.10079722106456757, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.069847971200943, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.06059828773140907, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.057408079504966736, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.056643225252628326, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.034835346043109894, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.02944425866007805, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.029147827997803688, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.0267548318952322, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02631346881389618, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01822105422616005, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.017978720366954803, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.017177006229758263, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011620169505476952, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.20909424126148224, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.73.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1419893503189087, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13163402676582336, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12715288996696472, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11395428329706192, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06709806621074677, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.062464673072099686, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07726897299289703, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07137051969766617, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06817876547574997, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05904539301991463, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.056038983166217804, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03929236903786659, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03422471508383751, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.032304778695106506, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03183990344405174, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.019716784358024597, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.016734469681978226, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.016534004360437393, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015111790969967842, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01490690652281046, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.010428089648485184, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.010447006672620773, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009833845309913158, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006996565498411655, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1419893503189087, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.73.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11775165051221848, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.10912832617759705, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.10470528900623322, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09369882941246033, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.055464040488004684, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.051292236894369125, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.0653543695807457, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06011981889605522, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.056386981159448624, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.048836011439561844, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.046436335891485214, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.033102381974458694, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.028795603662729263, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.026696590706706047, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.026180433109402657, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.016572143882513046, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.013739566318690777, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.013506599701941013, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.012387272901833057, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.012137091718614101, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008692332543432713, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008564682677388191, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.008037684485316277, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005545791704207659, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11775165051221848, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.73.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.28217950463294983, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.262004554271698, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2556016147136688, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.22918805480003357, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1332523226737976, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12550003826618195, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1491662561893463, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.13764922320842743, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.13529366254806519, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11732590943574905, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.11092199385166168, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07580152153968811, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06577788293361664, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0637882649898529, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.063308946788311, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.037840183824300766, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.032276008278131485, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03212231025099754, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.028923287987709045, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.028803599998354912, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.019352154806256294, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.018696915358304977, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01860934868454933, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011173835024237633, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.22918805480003357, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.73.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2451051026582718, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.18868359923362732, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.16486258804798126, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.12661291658878326, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.10896071791648865, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.08418170362710953, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.13837364315986633, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.1274164915084839, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.1164579689502716, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07498655468225479, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06992670148611069, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07073020190000534, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06121153011918068, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.05278380587697029, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.050637710839509964, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03548603132367134, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.027995361015200615, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.027259904891252518, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.021586259827017784, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.020213929936289787, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01895512267947197, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.019230114296078682, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.016276011243462563, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.013386680744588375, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.18868359923362732, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.73.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.274600625038147, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2561965882778168, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2505073845386505, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22584961354732513, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13055630028247833, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12349708378314972, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14522722363471985, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13418933749198914, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13243456184864044, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11573586612939835, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10974743217229843, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07415805757045746, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06439121067523956, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0627347007393837, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0623445063829422, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03715576231479645, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03239540383219719, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03228315711021423, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.029385633766651154, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.029309626668691635, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01969263143837452, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019716776907444, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01914752833545208, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013271686621010303, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22584961354732513, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.73.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.29710152745246887, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.27716588973999023, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2710457742214203, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2443268746137619, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1411089152097702, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1334589570760727, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15651127696037292, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1450134813785553, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14313052594661713, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.12507636845111847, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11849862337112427, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07991001754999161, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06943902373313904, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06766095012426376, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06724074482917786, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.039910681545734406, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03461701050400734, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.034500353038311005, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.0313129760324955, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.03123406693339348, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.02081957831978798, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020586568862199783, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.020234275609254837, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013235896825790405, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15651127696037292, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.73.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.24889105558395386, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.22776375710964203, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.21870355308055878, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.19567513465881348, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.11325617879629135, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.10555683076381683, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.13144737482070923, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.12100634723901749, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.11561912298202515, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.10026226937770844, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.09464488923549652, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06607186794281006, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.05741891264915466, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05409299209713936, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.05329069867730141, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03304731845855713, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.027919529005885124, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.027600500732660294, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.025377066805958748, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.024911466985940933, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.017484765499830246, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.017346499487757683, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01643051952123642, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.011510027572512627, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.22776375710964203, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.74.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12932483851909637, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11974930018186569, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1151660606265068, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10318226367235184, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0610801987349987, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05660758540034294, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07162743806838989, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06567320227622986, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06208759546279907, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.053755152970552444, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05111386254429817, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03652630001306534, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.031592272222042084, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.02949056215584278, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.028975632041692734, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.018358716741204262, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.015429852530360222, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.015197725035250187, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.013971466571092606, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013729607686400414, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009815193712711334, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009902814403176308, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009116856381297112, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006864763330668211, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12932483851909637, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.74.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1134762167930603, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.10509993135929108, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.10061592608690262, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09001704305410385, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05347519367933273, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04931528866291046, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06334131956100464, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.058280304074287415, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05434450879693031, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04706821218132973, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04477902874350548, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0321381539106369, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.027927620336413383, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02573019079864025, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.025202468037605286, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01612972281873226, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.013302958570420742, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.01305432990193367, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.012009983882308006, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.011742944829165936, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0085385050624609, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008386073634028435, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007859545759856701, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005504874046891928, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1134762167930603, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.74.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2747653126716614, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2551455795764923, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2486000955104828, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.22299471497535706, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.13000114262104034, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12214849144220352, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.14583206176757812, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.13454025983810425, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.13200460374355316, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11436162889003754, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10816583037376404, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07417915016412735, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06436436623334885, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06219620630145073, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.061700694262981415, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03701982647180557, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0315203033387661, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03134589269757271, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0282489862293005, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02809952199459076, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.018955973908305168, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.018364114686846733, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01816159300506115, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011044114828109741, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.22299471497535706, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.74.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2284470945596695, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.18213964998722076, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.16688281297683716, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.12866225838661194, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.10323365777730942, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.08575709909200668, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.12293031066656113, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.1123567596077919, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.10822119563817978, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07176020741462708, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0660589262843132, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.06315844506025314, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.05433519557118416, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.05027448758482933, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04928193241357803, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03181641176342964, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.027215898036956787, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.026969704777002335, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.021484188735485077, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.02103036642074585, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.017663685604929924, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01870504766702652, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.016420528292655945, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.014249504543840885, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2284470945596695, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.74.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.267956018447876, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.24997609853744507, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.24447011947631836, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22042380273342133, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12746015191078186, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.1205688938498497, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14176242053508759, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13100320100784302, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.129292294383049, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11300116777420044, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10721620172262192, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07258140295743942, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06295135617256165, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06134379282593727, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06096267327666283, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03631960228085518, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.031848032027482986, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03174159675836563, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.028935829177498817, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.028864756226539612, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01936570182442665, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019635779783129692, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.018842963501811028, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01354212500154972, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.22042380273342133, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.74.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.29307395219802856, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.27347221970558167, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2674272060394287, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.24111178517341614, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.13933026790618896, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.13180695474147797, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15464912354946136, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1431581974029541, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.14132122695446014, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.1235046312212944, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11708904802799225, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07911238819360733, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0686606913805008, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06690903007984161, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06649045646190643, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03955572098493576, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03447540104389191, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03435622528195381, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.03125131502747536, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.031170709058642387, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.020879466086626053, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.020870262756943703, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.02030789852142334, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013911336660385132, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.15464912354946136, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.74.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.24501898884773254, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.22350461781024933, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.21385300159454346, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.1915394514799118, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.11131082475185394, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.10325580090284348, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.1302202194929123, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.11989430338144302, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.11385134607553482, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.09842222183942795, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.09311611205339432, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.06571204215288162, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.056920185685157776, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.05321435630321503, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.052313584834337234, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.03281876817345619, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.027541037648916245, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.02717617340385914, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.02502964250743389, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.02450089529156685, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.01735498383641243, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.017295610159635544, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01615162007510662, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.01156799029558897, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.22350461781024933, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.75.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12746486067771912, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11792982369661331, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11358164995908737, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10179857164621353, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.060159023851156235, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05575511232018471, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06978694349527359, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06437686830759048, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06113370507955551, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05284189432859421, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.050181809812784195, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.035481229424476624, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.030900467187166214, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.02899816818535328, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.028535787016153336, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.017815914005041122, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.015106281265616417, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.014904121868312359, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.013640591874718666, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013430007733404636, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009488647803664207, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009581109508872032, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008918470703065395, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006562560796737671, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12746486067771912, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.75.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1122841164469719, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.10403435677289963, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.09962280094623566, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.08917572349309921, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05284582078456879, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04872583970427513, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06241617724299431, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05757102370262146, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05369904264807701, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.046487994492053986, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.044190775603055954, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03164968267083168, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02757207676768303, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02546062134206295, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.024932458996772766, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.015871770679950714, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.013164238072931767, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.012913111597299576, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.011879367753863335, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.011612031608819962, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008399919606745243, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00830626580864191, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007739478722214699, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005457377061247826, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1122841164469719, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.75.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.25330764055252075, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.23512502014636993, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2292860746383667, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20541155338287354, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.11947537213563919, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.11248276382684708, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1339627355337143, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.12359658628702164, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1213691309094429, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10511115938425064, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09941384941339493, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0680951252579689, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05907759070396423, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.057211607694625854, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05675649270415306, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03399668633937836, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.029072459787130356, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.028930630534887314, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.026088835671544075, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.025973668321967125, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.017457954585552216, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.017045896500349045, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.016772903501987457, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010469069704413414, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2292860746383667, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.75.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.22363731265068054, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1681390255689621, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.14525306224822998, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11618667095899582, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.09703168272972107, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07493627071380615, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.12660425901412964, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.11536762118339539, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.10618969053030014, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06927280128002167, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0639268010854721, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.06501181423664093, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.05570540949702263, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.047483254224061966, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04534430429339409, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.03284411132335663, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.025999953970313072, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.025405660271644592, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.021032530814409256, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.019772715866565704, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01829078420996666, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01886194385588169, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01578519493341446, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0141786253079772, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.22363731265068054, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.75.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.25728800892829895, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.24010254442691803, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.23480254411697388, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21175529062747955, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12229076027870178, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11569199711084366, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13585177063941956, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.125737726688385, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12402938306331635, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10845822095870972, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.1028200164437294, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06937949359416962, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06033465638756752, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05877835676074028, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.058406371623277664, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03470467031002045, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.030371258035302162, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03026692196726799, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.027568604797124863, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.027494831010699272, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01828072965145111, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01852318085730076, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01776343584060669, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012518569827079773, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.21175529062747955, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.75.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.28039637207984924, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2616342604160309, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.25584426522254944, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23072978854179382, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.1332155019044876, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12602221965789795, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14802025258541107, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1369110494852066, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13511040806770325, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11810809373855591, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.11192376911640167, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07561898976564407, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06564250588417053, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.06395159661769867, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0635540634393692, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03783153370022774, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.03292851522564888, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03281287103891373, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.029851889237761497, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.029773371294140816, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.019937563687562943, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01989288628101349, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.019373690709471703, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013209301047027111, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.23072978854179382, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.75.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.17978180944919586, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.16291269659996033, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.15457464754581451, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.13820655643939972, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.0822543203830719, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.07520383596420288, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.09833548218011856, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.09033665060997009, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.08438006043434143, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.07226989418268204, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.06837821751832962, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.04991111531853676, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04365759715437889, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.039943717420101166, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.03904984891414642, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.025295797735452652, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.021712616086006165, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.021338922902941704, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.019881075248122215, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.01936866156756878, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.013919985853135586, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.015053410083055496, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01270123478025198, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.01149788312613964, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.17978180944919586, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.76.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13466835021972656, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12468986213207245, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12043137848377228, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10790809243917465, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06331516057252884, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05898332968354225, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07325165718793869, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06738182157278061, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06445353478193283, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05580083280801773, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05295703932642937, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03723296895623207, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03230004757642746, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.030445829033851624, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.029999850317835808, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01865413784980774, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.015767043456435204, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.015585453249514103, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.014249163679778576, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.014054411090910435, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009835375472903252, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009843632578849792, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009227391332387924, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.006595886778086424, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13466835021972656, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.76.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11844545602798462, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.10998128354549408, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.10549402236938477, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.09446665644645691, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05566589534282684, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.05146179348230362, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.06542134284973145, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.0605052225291729, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05652923136949539, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.049110934138298035, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.046623796224594116, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.033110786229372025, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.028943007811903954, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.026740314438939095, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.026187989860773087, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01656944677233696, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0137073565274477, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.013441142626106739, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.012368336319923401, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.012087762355804443, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00865270011126995, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008488611318171024, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007969806902110577, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0053641051054000854, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11844545602798462, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.76.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2696874439716339, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.25062230229377747, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.24478760361671448, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.21968042850494385, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1271132230758667, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.1198204830288887, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.14229872822761536, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.13108964264392853, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.12898102402687073, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11219869554042816, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10632834583520889, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07241392135620117, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0626605823636055, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06083790585398674, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06037977337837219, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03616693243384361, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.030897239223122597, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.03076203353703022, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.027783824130892754, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02767227217555046, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.018564049154520035, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0180585328489542, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.017841976135969162, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011086207814514637, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.21968042850494385, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.76.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.27274319529533386, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.21753178536891937, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.19746311008930206, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1622014194726944, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.1213071346282959, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.09948624670505524, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1514672487974167, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.13717736303806305, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.13047519326210022, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.09002631157636642, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.08369042724370956, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.07785697281360626, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.06616352498531342, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.058977898210287094, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.057179778814315796, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.039124421775341034, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.031723178923130035, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.03128901496529579, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.025901105254888535, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0248944703489542, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.02131642773747444, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.02193397283554077, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.018989674746990204, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.016164664179086685, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.21753178536891937, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.76.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2483474612236023, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.23184426128864288, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22663791477680206, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2045717090368271, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11824818700551987, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11192736774682999, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1313725858926773, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12157861888408661, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11992952972650528, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.10497269034385681, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09965305030345917, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06737896800041199, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.058654434978961945, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05716121196746826, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05681167170405388, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.033723391592502594, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.030148137360811234, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.030051304027438164, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.02754460833966732, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02747710794210434, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.018141446635127068, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.01927662268280983, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017666393890976906, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.014076231978833675, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.22663791477680206, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.76.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.270382821559906, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2523466944694519, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.24673447012901306, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2226150631904602, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.12891869246959686, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.12203893810510635, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.14319771528244019, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.13249701261520386, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.13076597452163696, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.11447790265083313, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10870572924613953, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.07362347096204758, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.06413157284259796, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.0625128298997879, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.06213492900133133, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.0368974395096302, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.033315371721982956, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.03320518881082535, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.030517537146806717, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.030444631353020668, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.02009795419871807, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.021742651239037514, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01959121599793434, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01634400710463524, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.2226150631904602, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.76.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.16593217849731445, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.1494312733411789, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.13919462263584137, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.12410079687833786, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.07566101849079132, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.06791787594556808, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.09534946829080582, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.08664564788341522, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.07778605073690414, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.066020168364048, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.06273200362920761, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.04760395735502243, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04131603240966797, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.0366390123963356, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.035489920526742935, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.024046946316957474, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.019677666947245598, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.019207442179322243, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.01786901243031025, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.017239956185221672, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.013322178274393082, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.013584858737885952, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.01200268603861332, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.010043543763458729, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.16593217849731445, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.77.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12400460988283157, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11505033820867538, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11100860685110092, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09941510111093521, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.058430835604667664, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05440489947795868, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06784792989492416, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06237061694264412, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.05933268740773201, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05149633064866066, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04905273765325546, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.034510452300310135, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.029925579205155373, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.028162730857729912, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.027734320610761642, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.017327791079878807, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.014675959944725037, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.014484153129160404, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01329550426453352, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013102492317557335, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009237446822226048, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009304012171924114, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008705860003829002, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0063893902115523815, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12400460988283157, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.77.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10621999949216843, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.09853976219892502, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.09439680725336075, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.0844394788146019, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.049999311566352844, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.046143561601638794, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.05888872593641281, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.0543852299451828, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05075319856405258, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04402031749486923, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04180429130792618, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02984236367046833, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.026063106954097748, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02405085600912571, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.023566439747810364, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.014966046437621117, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.012460232712328434, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.01222543977200985, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.01125643402338028, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.011009400710463524, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.007932324893772602, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00787665881216526, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00733262300491333, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0052102976478636265, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10621999949216843, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.77.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.26244449615478516, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.24383413791656494, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.23787625133991241, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.21332299709320068, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.12400031089782715, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.11673588305711746, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1388774812221527, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.12803010642528534, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1258331835269928, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10911872237920761, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10330309718847275, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.07059147953987122, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06118111312389374, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.059307172894477844, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05885017663240433, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03524601086974144, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.030054740607738495, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02991507388651371, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02696632221341133, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.026865489780902863, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01807532086968422, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.017520589753985405, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.017371540889143944, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.010638768784701824, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.21332299709320068, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.77.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.21797828376293182, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1548042893409729, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.133097305893898, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11302489787340164, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.09915295243263245, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07151917368173599, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.11885145306587219, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.10756815969944, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.10424504429101944, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06299765408039093, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.060138601809740067, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.06128836050629616, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.051906272768974304, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.048120737075805664, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04720359668135643, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.030975835397839546, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.025752529501914978, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02555149234831333, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01908527873456478, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.018567034974694252, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01717466488480568, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01730785146355629, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01596635766327381, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.012803921476006508, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.21797828376293182, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.77.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.23391370475292206, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2182614803314209, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.21339504420757294, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.19243597984313965, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11104439198970795, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10505764931440353, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.12365913391113281, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.1141960546374321, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.11261707544326782, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09853627532720566, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.09349630773067474, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06321704387664795, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.05488002300262451, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.05344517529010773, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.053107425570487976, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.031639207154512405, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.0277999434620142, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.027700720354914665, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.025291211903095245, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.025222837924957275, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01681571453809738, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017227867618203163, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.016339078545570374, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.011971283704042435, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2182614803314209, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.77.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.25055140256881714, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.2338244616985321, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2286243587732315, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.20625551044940948, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.11916659772396088, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.11276830732822418, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.13267949223518372, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.12250141054391861, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.12085863947868347, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.1057848259806633, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.10053741931915283, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0681050643324852, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.0590660385787487, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.057563818991184235, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.057206686586141586, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.034112125635147095, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.030299322679638863, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.030200669541954994, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.027666369453072548, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.027600232511758804, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.018451163545250893, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.019277753308415413, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.017969224601984024, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.01398398820310831, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.2286243587732315, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.77.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.19271273910999298, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.1720791459083557, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.16200777888298035, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.14438924193382263, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.08743413537740707, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.07911129295825958, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.10818586498498917, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.0971844270825386, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.09004779160022736, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.07616811245679855, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.07227348536252975, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.054648809134960175, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.04717259854078293, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.0428834930062294, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.041821375489234924, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02816939540207386, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.024022242054343224, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.023589370772242546, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.022048529237508774, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.021447233855724335, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.016600973904132843, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.017476191744208336, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.015321833081543446, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.014016568660736084, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.19271273910999298, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.78.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12047673016786575, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11171885579824448, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10788977891206741, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09644776582717896, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.056678079068660736, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.052795447409152985, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06529118865728378, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.060357026755809784, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.057573091238737106, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04987389221787453, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04720389097929001, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0331575870513916, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.028922345489263535, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.027267534285783768, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.026868807151913643, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.0166314784437418, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.014145899564027786, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.013970838859677315, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01277713943272829, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.012599406763911247, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.008818392641842365, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.008867544122040272, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00831684097647667, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.005980315618216991, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12047673016786575, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.78.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10408199578523636, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.09664198756217957, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.09282004833221436, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.08297348767518997, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.04894557222723961, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04534178972244263, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.05737312138080597, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05294773355126381, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.04970456659793854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04309692606329918, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04090442508459091, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.02911192737519741, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02534755878150463, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.023538639768958092, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.023090118542313576, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.014586191624403, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.012121091596782207, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.011915379203855991, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.010928583331406116, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.010714665986597538, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.007678636349737644, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.007546965964138508, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007126193027943373, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004897558130323887, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.10408199578523636, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.78.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.277534157037735, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2575535774230957, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2513088583946228, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.22479155659675598, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.1309710592031479, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.12331729382276535, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.14677172899246216, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.13520506024360657, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1329594850540161, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.11520794779062271, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10879992693662643, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0746411457657814, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06469911336898804, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.06270439922809601, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.06222066655755043, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03728172928094864, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.031869854778051376, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.031720276921987534, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.028595147654414177, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02847839519381523, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.019177427515387535, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.018712788820266724, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01844046264886856, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0115584097802639, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.22479155659675598, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.78.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.22478540241718292, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.19347164034843445, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.18330101668834686, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1462886929512024, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.10042687505483627, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.08939047157764435, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.12730476260185242, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.10886527597904205, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.10449855029582977, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07945942878723145, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0735517218708992, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.06341741234064102, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.05415543168783188, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.05052992329001427, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.049618594348430634, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.034420356154441833, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.030021335929632187, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.029781047254800797, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.026662234216928482, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.026344135403633118, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.022744527086615562, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.02318553812801838, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.02180139347910881, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.019988102838397026, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.22478540241718292, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.78.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2085414081811905, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.19458162784576416, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.19021476805210114, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.1715673804283142, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.09930963069200516, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.09398514777421951, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.1106167808175087, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10211510211229324, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10072017461061478, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.08819229900836945, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08380011469125748, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.05700400844216347, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04962560907006264, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.04835762083530426, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.0480622760951519, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.02865118905901909, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.026111844927072525, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.026027144864201546, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.024010421708226204, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02395363338291645, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.015880528837442398, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017469827085733414, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.015485418029129505, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.013553942553699017, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2085414081811905, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.78.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2226492166519165, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.20764797925949097, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.20296645164489746, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.18311162292957306, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.10571272671222687, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.10001856833696365, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.11838063597679138, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.10877805948257446, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.10722773522138596, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.09383781999349594, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.08927524089813232, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.06080581247806549, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.052473437041044235, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.051091209053993225, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.05076346546411514, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.03064633160829544, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.02697681076824665, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.026879940181970596, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.024650806561112404, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.02458156831562519, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.01694219931960106, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.017272524535655975, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01650087535381317, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.012629123404622078, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.2226492166519165, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.78.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.15451893210411072, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.13693906366825104, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.12728208303451538, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.11347693204879761, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.07025444507598877, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.06246095895767212, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.08943019807338715, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.08027191460132599, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.07269593328237534, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.06086166948080063, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.05830835551023483, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.04571777209639549, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.039140354841947556, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.03474931791424751, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.03364855796098709, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.02343609370291233, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.019832199439406395, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.019387852400541306, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.018225859850645065, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.017619045451283455, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.013640866614878178, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.014905768446624279, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.012261870317161083, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.012180594727396965, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.15451893210411072, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.79.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.11049583554267883, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.10197807848453522, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.09809538722038269, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08751266449689865, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.05183076113462448, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04801740124821663, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.060355886816978455, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.05566856265068054, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.052733298391103745, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.045412998646497726, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04295765608549118, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.030675865709781647, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02672259695827961, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.024995209649205208, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.02457672357559204, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.015416039153933525, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01309415977448225, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01291095931082964, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.011827831156551838, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.011636506766080856, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.008281514048576355, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00842370931059122, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.007752530742436647, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.005886981729418039, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.11049583554267883, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.79.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.09290586411952972, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.08574298769235611, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.08224596828222275, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.073328897356987, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.04348917677998543, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04017339646816254, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.05084908753633499, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04698113352060318, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.04418490454554558, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03804560750722885, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03597826138138771, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.025761160999536514, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02249726466834545, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.020934194326400757, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.020555445924401283, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.012939115054905415, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.010857262648642063, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.010683642700314522, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.009765415452420712, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.009584572166204453, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.006889273878186941, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.006864152383059263, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00641682930290699, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004606175236403942, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.09290586411952972, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.79.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.258738249540329, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.23968563973903656, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.23313286900520325, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20803433656692505, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.12196579575538635, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.1143883764743805, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.13728998601436615, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.1266387403011322, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.12390514463186264, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.10686974227428436, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.10078546404838562, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.06978010386228561, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.06061090901494026, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05846478044986725, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05797003582119942, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.03490506857633591, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.029848629608750343, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.029669035226106644, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.026724958792328835, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.026568781584501266, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.018021028488874435, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01777993142604828, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.017270416021347046, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.011220128275454044, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.20803433656692505, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.79.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18930931389331818, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14193077385425568, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12192320823669434, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09535778313875198, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08051984012126923, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.060143884271383286, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10874420404434204, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09831223636865616, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08952610194683075, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05826124548912048, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0559905506670475, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.056256942451000214, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.048227351158857346, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04034813866019249, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.038303859531879425, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.028729500249028206, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.023578409105539322, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.023030700162053108, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.019989674910902977, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.018872832879424095, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01667035184800625, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01858609914779663, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.014368658885359764, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.015291282907128334, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18930931389331818, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.79.mlp.gate_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.16560108959674835, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.15380989015102386, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.15003371238708496, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.13463439047336578, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.0782589241862297, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.073744997382164, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.08766112476587296, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.08080941438674927, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.0794435515999794, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.06899980455636978, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.06526754796504974, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.0448606051504612, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.03886052593588829, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.037705253809690475, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.037430163472890854, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.022456925362348557, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.019719764590263367, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.019631410017609596, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.01788550615310669, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.01781512424349785, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.012019248679280281, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.012402205727994442, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.01163962110877037, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.00878218375146389, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.16560108959674835, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.79.mlp.up_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.16947759687900543, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.3750523158482144, + "total_bits": 557854720.0, + "err": 0.15728743374347687, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.6250523158482144, + "total_bits": 616574976.0, + "err": 0.15334567427635193, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7188023158482144, + "total_bits": 638595072.0, + "err": 0.13758158683776855, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2188023158482144, + "total_bits": 756035584.0, + "err": 0.08023729175329208, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7344273158482144, + "total_bits": 877146112.0, + "err": 0.0755416601896286, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312630789620534, + "total_bits": 711986176.0, + "err": 0.09025322645902634, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.1250523158482144, + "total_bits": 734015488.0, + "err": 0.082933709025383, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.1719273158482144, + "total_bits": 745025536.0, + "err": 0.08147387951612473, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5313023158482144, + "total_bits": 829435904.0, + "err": 0.07071564346551895, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656276157924107, + "total_bits": 858789888.0, + "err": 0.06703951209783554, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.031263078962054, + "total_bits": 946867200.0, + "err": 0.046414025127887726, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125052315848214, + "total_bits": 968896512.0, + "err": 0.04010428488254547, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218802315848214, + "total_bits": 990916608.0, + "err": 0.03889217972755432, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312552315848214, + "total_bits": 1012936704.0, + "err": 0.038599371910095215, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.031263078962054, + "total_bits": 1181748224.0, + "err": 0.023359855636954308, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218802315848214, + "total_bits": 1225797632.0, + "err": 0.020749235525727272, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312552315848214, + "total_bits": 1247817728.0, + "err": 0.020665457472205162, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531302315848214, + "total_bits": 1299197952.0, + "err": 0.018920868635177612, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703177315848214, + "total_bits": 1339568128.0, + "err": 0.0188515093177557, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.031263078962054, + "total_bits": 1416629248.0, + "err": 0.012901939451694489, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125052315848214, + "total_bits": 1438658560.0, + "err": 0.013604938983917236, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.218763078962054, + "total_bits": 1460669440.0, + "err": 0.012512187473475933, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125052315848214, + "total_bits": 1908420607.9999998, + "err": 0.010239314287900925, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.1719273158482144, + "total_bits": 510144512.0, + "err": 0.16947759687900543, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.79.mlp.down_proj", + "numel": 234881024, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.06399326026439667, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 557885440.0, + "err": 0.05537997931241989, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 616605696.0, + "err": 0.049420423805713654, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7233973911830356, + "total_bits": 639674368.0, + "err": 0.043506547808647156, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2233973911830356, + "total_bits": 757114880.0, + "err": 0.029209868982434273, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7278616768973216, + "total_bits": 875603968.0, + "err": 0.02467299811542034, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 711993856.0, + "err": 0.041056059300899506, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 734046208.0, + "err": 0.036153219640254974, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.174290248325893, + "total_bits": 745580544.0, + "err": 0.030538776889443398, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5269688197544644, + "total_bits": 828418048.0, + "err": 0.025280628353357315, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6608058384486606, + "total_bits": 859853824.0, + "err": 0.024576101452112198, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 946874880.0, + "err": 0.02150842733681202, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 968927232.0, + "err": 0.018533281981945038, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.223397391183036, + "total_bits": 991995904.0, + "err": 0.015499984845519066, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.321611676897321, + "total_bits": 1015064576.0, + "err": 0.01470866147428751, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 1181755904.0, + "err": 0.011751788668334484, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.223397391183036, + "total_bits": 1226876928.0, + "err": 0.010181655175983906, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.321611676897321, + "total_bits": 1249945600.0, + "err": 0.009892857633531094, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526968819754464, + "total_bits": 1298180096.0, + "err": 0.009625906124711037, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.71893310546875, + "total_bits": 1343268864.0, + "err": 0.009258586913347244, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1416636928.0, + "err": 0.007768807001411915, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1438689280.0, + "err": 0.00880548357963562, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.227724347795759, + "total_bits": 1462774272.0, + "err": 0.006988705601543188, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1908451328.0, + "err": 0.007911194115877151, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.174290248325893, + "total_bits": 510699520.0, + "err": 0.06399326026439667, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + } + ], + "last_module_idx": 162, + "base_perplexity": 1051.0014796236303, + "reuse_measurement": "migtissera_Synthia-70B-v1.2-2.3bpw\\measurement.json", + "cal_filename": "C:\\Users\\user\\Downloads\\oobabooga_windows\\oobabooga_windows\\migtissera_Synthia-70B-v1.2-2.55bpw\\cal_data.safetensors", + "q_last_module_idx": 162, + "cal_perplexity": 1782.6083328489458 +} \ No newline at end of file