{ "measurement": [ { "key": "model.layers.0.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.012533674947917461, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.012346061877906322, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.005992434918880463, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.005896490532904863, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.005778010934591293, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.002843402558937669, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.018051940947771072, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.012104665860533714, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.0057954383082687855, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.00571693480014801, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.006079370155930519, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.006583208218216896, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.0056831990368664265, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.003479550126940012, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.002743422519415617, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.0036582318134605885, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.00270731165073812, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.002413012320175767, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.0027017141692340374, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.0024073903914541006, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.002692453796043992, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.0026988168247044086, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.0023062746040523052, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.002401267411187291, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.012533674947917461, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.012533674947917461, "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": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.014539353549480438, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.01422488410025835, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.007164649665355682, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.006988872773945332, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.00678146630525589, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.0030296870972961187, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.01670859567821026, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.013855201192200184, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.006812590640038252, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.006670237518846989, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.007102499715983868, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.007426466327160597, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.0066098603419959545, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.0038290582597255707, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.0028213239274919033, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.003931448329240084, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.0027428509201854467, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.0023466097190976143, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.002729850821197033, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.0023335323203355074, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.002566391136497259, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.002724167425185442, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.001984803704544902, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.002319084480404854, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.014539353549480438, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.014539353549480438, "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": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.07006265968084335, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.0611538290977478, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.054120615124702454, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.04623882472515106, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.032406844198703766, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.026979593560099602, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.050468821078538895, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.040300410240888596, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.033254120498895645, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.02642974443733692, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.02612368017435074, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.026588745415210724, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.01939447596669197, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.015841884538531303, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.014888795092701912, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.013367951847612858, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.008589452132582664, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.00796777568757534, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.007473545614629984, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.006868835538625717, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.006903576198965311, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.006268922705203295, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.004735216964036226, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.004513110034167767, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.054120615124702454, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, "best_option": { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.054120615124702454, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.0.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.01943555288016796, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.012974992394447327, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.007315636612474918, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.007975679822266102, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.007659899070858955, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.0034277013037353754, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.01409465167671442, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.012390353716909885, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.009208342991769314, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.006235379725694656, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.006616911850869656, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.007165319751948118, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.006115911528468132, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.004105482250452042, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.0034854973200708628, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.0036959892604500055, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.002742876298725605, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.0024381510447710752, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.002555347979068756, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.002253161510452628, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.0022004954516887665, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.0025412351824343204, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.001628819270990789, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0021634630393236876, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.01943555288016796, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.01943555288016796, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.0.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.10988049954175949, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.10396689921617508, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.10196872800588608, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.09313695877790451, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.05016353353857994, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.04834284260869026, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.05552302300930023, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.051415372639894485, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.050547197461128235, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.04583032429218292, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.04413261637091637, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.02824174240231514, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.024730874225497246, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.024207955226302147, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.024068884551525116, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.014189786277711391, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.01306349691003561, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.012950925156474113, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.012281328439712524, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.01220549177378416, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.007885447703301907, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.008740054443478584, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.007722353562712669, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.00681580463424325, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.05552302300930023, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.05552302300930023, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.0.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.13230685889720917, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.1252252757549286, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.12286961078643799, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.11238593608140945, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.060141779482364655, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.05800469592213631, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.06637625396251678, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.06151352450251579, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.060592345893383026, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.05493772774934769, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.05280030146241188, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.03341109678149223, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.029119614511728287, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.02853522263467312, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.028394153341650963, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.016642054542899132, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.014543382450938225, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.014400722458958626, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.01351484190672636, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.01342998631298542, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.008627524599432945, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.008575920946896076, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.008429117500782013, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.005472726188600063, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.05493772774934769, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.05493772774934769, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.0.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.0800698921084404, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.06973877549171448, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.0653364434838295, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.05883060768246651, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.03502031788229942, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.031417809426784515, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.04279087856411934, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.03924262151122093, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.0366385392844677, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.030420592054724693, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.02897505834698677, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.02166684716939926, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.018834605813026428, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.01700284704566002, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.01654602214694023, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.01092282123863697, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.00931285135447979, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.009204053319990635, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.008480522781610489, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.00818348117172718, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.0060520782135427, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.006586800329387188, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.005481044761836529, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.0050888448022305965, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.04279087856411934, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.04279087856411934, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.1.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.024519024416804314, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.020326588302850723, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.015096781775355339, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.013438996858894825, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.010637822560966015, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.007343575358390808, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.019562164321541786, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.016181614249944687, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.011252823285758495, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.00875846017152071, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.00898953340947628, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.009249572642147541, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.007566903252154589, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.005347477737814188, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.0046944706700742245, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.004737146198749542, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.003157570958137512, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.002802447183057666, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.002852852689102292, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.0024951801169663668, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.0028084563091397285, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.002677216660231352, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.0021588150411844254, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0020996846724301577, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.024519024416804314, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.024519024416804314, "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": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.023534167557954788, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.01936352252960205, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.014226066879928112, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.012719433754682541, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.010175545699894428, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.006853640079498291, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.018252165988087654, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.015567125752568245, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.010794240981340408, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.00835895910859108, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.008586970157921314, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.00879611261188984, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.00728783942759037, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.00507791293784976, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.004408529959619045, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.0044382368214428425, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.002926300512626767, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.0025366719346493483, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.002623124746605754, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.0022225321736186743, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.0024840738624334335, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.0024595230352133512, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.001771475188434124, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.001830684021115303, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.023534167557954788, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.023534167557954788, "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": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11803343147039413, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.10292958468198776, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.09616676717996597, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.08369806408882141, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.054602671414613724, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.04807015880942345, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06620688736438751, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.060634925961494446, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.056141361594200134, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.0446806438267231, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.042027801275253296, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03393453732132912, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.029236245900392532, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.026450710371136665, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.025764847174286842, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.017029207199811935, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.014030290767550468, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.013475248590111732, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.012115059420466423, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.011659429408609867, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.009084070101380348, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.009391078725457191, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.008094697259366512, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.006703490391373634, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.054602671414613724, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.054602671414613724, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.1.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.13251593708992004, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.08599918335676193, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.062292538583278656, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.05986654385924339, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.059386275708675385, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.03689820319414139, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.08160928636789322, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.07217350602149963, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.06303615868091583, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.03672536090016365, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0391535721719265, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.041671350598335266, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.036076053977012634, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.030465418472886086, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.028972238302230835, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.02154509536921978, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.018527347594499588, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.017328590154647827, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.015177740715444088, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01427046861499548, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.012798071838915348, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.015098693780601025, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.011150890029966831, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.012924023903906345, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.03672536090016365, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.03672536090016365, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.1.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.12954510748386383, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.1229197233915329, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.12089580297470093, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.1103973537683487, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.05913330614566803, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.05720961466431618, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.06526234745979309, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.06020453944802284, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.059511151164770126, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.054155945777893066, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.05225178226828575, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.03326749429106712, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.029146535322070122, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.028690146282315254, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.028579751029610634, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.01676747389137745, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.015740638598799706, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.015636712312698364, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.014862991869449615, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.014807578176259995, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.0094954464584589, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.010813407599925995, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.009350830689072609, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.008706108666956425, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.054155945777893066, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.054155945777893066, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.1.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.17327646911144257, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.16458582878112793, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.16199371218681335, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.148367241024971, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.07943262159824371, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.07693756371736526, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.08728885650634766, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.08080534636974335, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.07995004206895828, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07287151366472244, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07013721019029617, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.044433727860450745, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.03886204585433006, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.03827083110809326, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.038130901753902435, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.02225903607904911, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02054392546415329, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02040642872452736, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.0193319171667099, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.019241610541939735, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.012237709946930408, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.013575009070336819, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.012054678983986378, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010461409576237202, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.044433727860450745, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.044433727860450745, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.1.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.04711565375328064, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.04686509817838669, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.009521067142486572, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.009322396479547024, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.008966642431914806, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.0035502579994499683, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.054198332130908966, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.04663226753473282, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.009022437036037445, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.00886114127933979, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.008826467208564281, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.008728297427296638, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.008751160465180874, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.002970051486045122, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.0031597029883414507, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.0028800403233617544, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.0030296649783849716, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.004632859956473112, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.003014975693076849, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.004625645000487566, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.003001069650053978, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.002994856331497431, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.004600713960826397, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.0046095396392047405, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.04711565375328064, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.04711565375328064, "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": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.04025686904788017, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.03766828402876854, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.036605868488550186, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.03304323926568031, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.0185259822756052, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.01759284920990467, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.021687516942620277, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.019391072914004326, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.018702100962400436, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.01659034937620163, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.01619199477136135, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.011055628769099712, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.009275869466364384, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.00890086404979229, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.008809263817965984, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.0055554513819515705, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.004610095638781786, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.004533872473984957, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.004230417776852846, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.004173010587692261, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.0029331219848245382, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.002846574643626809, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.0028129518032073975, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0019197161309421062, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.04025686904788017, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.04025686904788017, "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": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.035885460674762726, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.03350694477558136, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.03246503323316574, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.02927597053349018, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.01643582247197628, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.015558195300400257, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.01891106367111206, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.01732802763581276, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.016599833965301514, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.014682503417134285, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.014078456908464432, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.00956763606518507, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.008251267485320568, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.007865753956139088, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.007777090184390545, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.004798575770109892, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.004068384412676096, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.003994603641331196, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.003721510525792837, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.003661738010123372, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.0025182324461638927, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.0025192208122462034, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.002395290182903409, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0016924203373491764, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.035885460674762726, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.035885460674762726, "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": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.16004979610443115, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.15087606012821198, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.1478549689054489, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.13447609543800354, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.07435896247625351, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.0712767243385315, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.08269239217042923, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.07628437876701355, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.07496432960033417, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.06699524074792862, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.06413526087999344, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.041875164955854416, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.03629716858267784, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.0354132317006588, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.035210177302360535, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.020870955660939217, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.01798131875693798, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.01776406541466713, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.01646454446017742, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.016338078305125237, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.010717158205807209, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.010479345917701721, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.010376906022429466, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.00646671699360013, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.041875164955854416, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.041875164955854416, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.2.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.1810743659734726, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.1561564952135086, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.14680421352386475, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.12144589424133301, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.08447425067424774, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.07472558319568634, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.10003938525915146, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.09055664390325546, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.08618546277284622, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.06586272269487381, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.05846778675913811, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05156208202242851, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04436927288770676, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.04159112647175789, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.040921229869127274, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.02639877237379551, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.023219812661409378, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.022622982040047646, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.01993277482688427, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.019507907330989838, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.015334408730268478, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.016642054542899132, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.014526373706758022, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.013463239185512066, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05156208202242851, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05156208202242851, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.2.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.18287257850170135, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.17335787415504456, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.17038899660110474, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.155825674533844, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.08473430573940277, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.08178408443927765, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09337823837995529, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.08639161288738251, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.08534863591194153, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07714381814002991, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07415106892585754, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04741111397743225, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04119620844721794, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04046465829014778, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04029954597353935, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.023687344044446945, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.020918555557727814, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.020731735974550247, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.01942365989089012, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.019318416714668274, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.012561100535094738, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.012746758759021759, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.012319023720920086, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.00871619675308466, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04741111397743225, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04741111397743225, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.2.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.21390873193740845, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.20306913554668427, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.1996777504682541, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.1826879233121872, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.09924504905939102, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.09584125876426697, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10924623906612396, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.10114993900060654, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.09996603429317474, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.0904431939125061, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.08682123571634293, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05524888634681702, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04807177931070328, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.047240905463695526, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.047042280435562134, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.027541551738977432, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.023996742442250252, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02377784438431263, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.022196944802999496, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.022078538313508034, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.014233261346817017, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.013962404802441597, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.013954713940620422, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.008702096529304981, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05524888634681702, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05524888634681702, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.2.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.2030741423368454, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.18612615764141083, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.17978957295417786, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.16455137729644775, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.09176602214574814, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.08607544749975204, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.10687705129384995, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.09782016277313232, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.09425342828035355, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.08286319673061371, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.07933860272169113, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.05391616374254227, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.04669275879859924, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.04398078843951225, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.043327704071998596, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.02707565762102604, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.023028871044516563, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.022880440577864647, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.02130803093314171, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.020825738087296486, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.014599944464862347, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.01475688349455595, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.013746984302997589, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.010353147983551025, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.05391616374254227, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.05391616374254227, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.3.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.0758657306432724, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.07095514237880707, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.0690918117761612, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.062410157173871994, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.035323355346918106, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.0335603691637516, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.04114485904574394, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.036769431084394455, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.035675209015607834, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.0315256342291832, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.030754882842302322, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.021063901484012604, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.017631402239203453, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.016991816461086273, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.01684633642435074, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.01060356106609106, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.008776900358498096, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.008641255088150501, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.008026260882616043, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.00793051440268755, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.005532500799745321, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.0053267572075128555, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.005328514147549868, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0035072637256234884, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.04114485904574394, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.04114485904574394, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.3.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.065064437687397, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.060718368738889694, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.05903556942939758, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.0532984733581543, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.03012774884700775, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.028588270768523216, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.03423060104250908, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.03145677223801613, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.030443906784057617, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.026828018948435783, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.02567267417907715, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.017327995970845222, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.01499348133802414, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.014413298107683659, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.014274738729000092, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.008660931140184402, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.0074100615456700325, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.007291552145034075, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.0067466930486261845, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.00666304724290967, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.004519143607467413, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.004509781487286091, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.004324219189584255, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.00296837673522532, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.0532984733581543, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.0532984733581543, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.3.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.198655903339386, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.18682079017162323, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.18290190398693085, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.16598759591579437, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.09291940927505493, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.08892334997653961, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.10335855931043625, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.09543631970882416, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.09376943111419678, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.08331591635942459, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.07954390347003937, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.052329715341329575, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04546771198511124, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.04433491826057434, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.044067513197660446, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.02611576020717621, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.022519294172525406, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.022245824337005615, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.0205522533506155, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.020380457863211632, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.013431379571557045, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.013165238313376904, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.01301220990717411, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.00815560482442379, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.052329715341329575, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.052329715341329575, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.3.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.14963480830192566, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.13145555555820465, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.12191125005483627, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.10442030429840088, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.07042193412780762, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.061949558556079865, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.08748101443052292, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.07935179024934769, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.0720466747879982, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05724497511982918, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.05431065335869789, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.045389920473098755, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.03947852924466133, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.03514556586742401, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.034094810485839844, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.02312895655632019, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.020284907892346382, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.0195995531976223, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.018054412677884102, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.017426133155822754, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.013167866505682468, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.015427197329699993, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.011781311593949795, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.012789908796548843, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.05431065335869789, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, "best_option": { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.05431065335869789, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } } }, { "key": "model.layers.3.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.19613556563854218, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.18539582192897797, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.1819799393415451, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.16584327816963196, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.09205026924610138, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.0885138213634491, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10180389881134033, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09408468753099442, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.09283023327589035, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.08317454159259796, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07970661669969559, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.051858413964509964, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04501189664006233, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04412437602877617, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.043911486864089966, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.025927772745490074, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.022863591089844704, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.022634563967585564, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.021098563447594643, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02096402458846569, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.013781779445707798, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.014015082269906998, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.013489367440342903, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009653892368078232, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.051858413964509964, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.051858413964509964, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.3.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.23923811316490173, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.22633205354213715, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.22223293781280518, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.20268180966377258, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11213511973619461, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10782385617494583, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.1239449605345726, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11453372240066528, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11300958693027496, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10136960446834564, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09718077629804611, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06280703842639923, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0545913465321064, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05353008583188057, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05328807234764099, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.0313701331615448, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02724597230553627, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.026971979066729546, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.025035828351974487, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.024873778223991394, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.016306253150105476, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.015949642285704613, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01595398411154747, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010023899376392365, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0545913465321064, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0545913465321064, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.3.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.21234863996505737, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.19237516820430756, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.1848224699497223, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.1670869141817093, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.0967424139380455, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.08992607891559601, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.11247197538614273, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.1035236194729805, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.09961285442113876, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.08564728498458862, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.08102942258119583, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.05705646798014641, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.04959430918097496, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.04655041918158531, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.04582183435559273, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.028640829026699066, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.024525422602891922, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.024395307525992393, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.022387024015188217, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.021847836673259735, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.015479065477848053, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.015943415462970734, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.014521995559334755, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.011397339403629303, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.04959430918097496, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.04959430918097496, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.4.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.07774602621793747, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.07291102409362793, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.0709313303232193, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.06420067697763443, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.03635738417506218, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.03456169739365578, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.042379844933748245, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.037989553064107895, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.036695465445518494, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.03254716098308563, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.03175801783800125, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.021696727722883224, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.018218398094177246, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.017514487728476524, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.017346743494272232, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.010906383395195007, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.009009703062474728, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.008863011375069618, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.008242417126893997, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.008137524127960205, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.005676992703229189, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.005449641961604357, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.005446967203170061, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0035490733571350574, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.042379844933748245, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.042379844933748245, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.4.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.06295722723007202, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.05898814648389816, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.0572434738278389, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.051735226064920425, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.029312001541256905, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.027750112116336823, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.03358405455946922, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.030862707644701004, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.02958965301513672, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.026163581758737564, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.025085730478167534, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.01700642891228199, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.01472153328359127, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.014027342200279236, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.013863920234143734, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.008499354124069214, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.007214989513158798, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.0070769404992461205, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.006578410509973764, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.006472884211689234, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.004443936515599489, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.004411289934068918, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.004216024186462164, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0028838571161031723, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.051735226064920425, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.051735226064920425, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.4.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.20701676607131958, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.19475987553596497, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.1906169354915619, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.17298978567123413, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.09712714701890945, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09283486753702164, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.10827794671058655, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.09981995075941086, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.09802142530679703, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.08694500476121902, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.08307089656591415, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05495520681142807, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04757118970155716, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.0463586188852787, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.04606754705309868, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.02741335518658161, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.023549599573016167, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.023252205923199654, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.021462883800268173, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.02128211408853531, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.014110750518739223, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.013779796659946442, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.01363262627273798, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.008538316003978252, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05495520681142807, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05495520681142807, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.4.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.17954444885253906, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.1582755148410797, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.15016795694828033, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.1267886608839035, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.08383560180664062, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.07563468813896179, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.09779345989227295, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.08975277841091156, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.08554564416408539, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.06842236965894699, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.061924081295728683, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.050530169159173965, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04410998895764351, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.04139597713947296, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.04075471684336662, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.02568252570927143, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02323932945728302, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.02264961414039135, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.020516134798526764, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.020115984603762627, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.014755280688405037, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.016762729734182358, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.013956968672573566, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.013677763752639294, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.050530169159173965, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.050530169159173965, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.4.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.18232585489749908, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.17205914855003357, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.1687707006931305, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.1536281853914261, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.08572132885456085, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.08222763240337372, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09479760378599167, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.08780191838741302, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.08642986416816711, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07717597484588623, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07378045469522476, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04820004478096962, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.041951991617679596, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04102693870663643, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.0408010333776474, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.024083659052848816, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.021151786670088768, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02092115394771099, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.019437406212091446, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.0192941315472126, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.012650426477193832, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.012820921838283539, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.012336067855358124, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.008613353595137596, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04820004478096962, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04820004478096962, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.4.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.23879048228263855, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.2255847007036209, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.22136178612709045, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.20159707963466644, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.1123020276427269, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10781683027744293, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.1240166574716568, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11483480781316757, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11316999047994614, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10122387856245041, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09674759209156036, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06296278536319733, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.054779596626758575, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05360719561576843, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05333650857210159, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.031406376510858536, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02725531905889511, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.026958152651786804, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.024970384314656258, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.024794017896056175, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.016226468607783318, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.015919845551252365, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01582713983952999, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009927479550242424, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.054779596626758575, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.054779596626758575, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.4.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.21836084127426147, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.19646842777729034, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.18857769668102264, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.16933228075504303, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.09997541457414627, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.09246878325939178, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.11645396053791046, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.1067005842924118, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.10302392393350601, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.08754357695579529, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.08240646123886108, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.059284575283527374, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.051273152232170105, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.048204001039266586, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.04744209349155426, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.029927609488368034, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.025471966713666916, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.02535821683704853, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.023083452135324478, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.02252560295164585, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.016441890969872475, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.016626162454485893, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.015473619103431702, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.011994299478828907, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.051273152232170105, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.051273152232170105, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.5.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.08512309938669205, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.07978091388940811, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.0776282325387001, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.07029180228710175, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.03986074775457382, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.037839125841856, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.04571587219834328, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.04158709943294525, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.040248945355415344, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.035581670701503754, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.034384943544864655, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.023279577493667603, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.019886882975697517, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.01912032999098301, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.018938379362225533, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.011657620780169964, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.009824794717133045, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.009667862206697464, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.008960035629570484, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.00884610041975975, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.0060856640338897705, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.005954098422080278, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.0058311051689088345, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.003891704138368368, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.04571587219834328, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.04571587219834328, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.5.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.06888818740844727, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.06450355798006058, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.06257648020982742, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.056511860340833664, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.03215451538562775, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.03043147549033165, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.036712296307086945, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.03385118395090103, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.0324682779610157, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.02862592600286007, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.02737642638385296, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.018584398552775383, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.016140326857566833, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.01538291946053505, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.015208696015179157, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.009280357509851456, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.00789988785982132, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.007748573087155819, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.007189564406871796, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.0070740170776844025, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.0048324232921004295, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.004805481992661953, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.004583387635648251, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0031177497003227472, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.036712296307086945, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.036712296307086945, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.5.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.22053022682666779, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.20758475363254547, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.20327459275722504, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.1841738075017929, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10378226637840271, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.0992419645190239, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11578356474637985, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.10661783814430237, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10472428798675537, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.09285872429609299, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.08867457509040833, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.058835092931985855, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05086831748485565, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.049594879150390625, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.049288101494312286, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.029362421482801437, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.025194155052304268, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.024876516312360764, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02293909154832363, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.022743510082364082, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.015170587226748466, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.014734246768057346, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.014667061157524586, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.009134004823863506, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05086831748485565, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05086831748485565, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.5.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.18268968164920807, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.16487522423267365, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.15179872512817383, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.1280980110168457, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.08679336309432983, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.07634290307760239, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.1138433963060379, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.10132052004337311, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.08793599903583527, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.07178954780101776, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.06727565824985504, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.059053633362054825, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.050280194729566574, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.04340524971485138, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.041688259690999985, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.029859978705644608, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.025124989449977875, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.02404707297682762, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.022591181099414825, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.021577784791588783, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.016524989157915115, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.019390907138586044, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.01397805567830801, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.01594647206366062, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.050280194729566574, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.050280194729566574, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.5.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.17497195303440094, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.1652022898197174, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.1619257777929306, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.14746591448783875, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.08240297436714172, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.07901757955551147, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09114759415388107, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.0843794122338295, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.08306261152029037, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07412108033895493, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07092240452766418, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.0463893860578537, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.040374018251895905, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.039483118802309036, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.039276812225580215, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.023186219856142998, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.020406795665621758, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02018214575946331, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.01874721050262451, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.01861891709268093, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.012223873287439346, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.012448226101696491, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.011928169056773186, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.008459329605102539, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.0463893860578537, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.0463893860578537, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.5.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.23616977035999298, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.22308605909347534, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.218900665640831, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.19940008223056793, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.1111411303281784, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.1066545695066452, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12298043072223663, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11366557329893112, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11205151677131653, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10009966045618057, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09581863135099411, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06238866597414017, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05425264686346054, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05310530960559845, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05282922834157944, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.031161881983280182, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02703160233795643, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.026729904115200043, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.024756833910942078, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02457904815673828, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.016184331849217415, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.015842068940401077, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.0157993845641613, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009944633580744267, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05425264686346054, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05425264686346054, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.5.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.23108132183551788, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.20884032547473907, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.20056281983852386, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.17970429360866547, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.10631470382213593, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.09857933223247528, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.1232687383890152, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.11356965452432632, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.10942032933235168, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09314463287591934, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.0875348448753357, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06296725571155548, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.054602157324552536, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05133393779397011, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05054383724927902, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.031694792211055756, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.02721390686929226, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.027059655636548996, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.02467193268239498, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.024066243320703506, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.017275461927056313, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.017855294048786163, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.016243312507867813, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.012963332235813141, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.054602157324552536, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.054602157324552536, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.6.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.09845563769340515, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.09198329597711563, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.08928504586219788, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.08057653903961182, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.045910678803920746, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.04343492537736893, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05287158861756325, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.04811450093984604, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.04636277258396149, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.0407986082136631, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.03931935504078865, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.026894554495811462, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.02304058149456978, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.022058293223381042, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.0218341127038002, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.013476994819939137, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.011448877863585949, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.011246697045862675, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.010431578382849693, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.010289463214576244, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007101547904312611, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007108449004590511, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.006780141964554787, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.00483478419482708, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05287158861756325, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05287158861756325, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.6.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.08443419635295868, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.07875501364469528, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.07633363455533981, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.06874935328960419, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.03928835690021515, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.03710686415433884, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.04477234557271004, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.041301239281892776, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.03970607370138168, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.03483833000063896, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.03319992870092392, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.022672047838568687, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.01972060091793537, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.01882188394665718, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.01861092448234558, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.011336587369441986, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.00971474964171648, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.00953003205358982, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.008819838985800743, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.008680004626512527, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.005927968770265579, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.005981328897178173, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.0056246062740683556, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.003981819376349449, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.04477234557271004, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.04477234557271004, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.6.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.21544241905212402, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.20263639092445374, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.19816412031650543, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.17946447432041168, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10136976838111877, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09678961336612701, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11359544843435287, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.10439668595790863, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10230608284473419, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.09057661145925522, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.08644663542509079, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05764446780085564, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04984940215945244, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.04847555235028267, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.04813610017299652, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.028791576623916626, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02468010224401951, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.0243527889251709, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.022463956847786903, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.022245723754167557, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.014880513772368431, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.014563610777258873, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.01432291604578495, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.009158825501799583, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04984940215945244, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04984940215945244, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.6.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.18955372273921967, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.1701090782880783, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.16223135590553284, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.1389719396829605, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.08981762826442719, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.08208782225847244, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.1055682897567749, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.09616002440452576, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.09150496870279312, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.07447188347578049, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.06866063922643661, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05473514273762703, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04753628745675087, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.044661231338977814, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.0439407080411911, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.028129566460847855, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.025124583393335342, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.02455884963274002, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02236907370388508, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.021945107728242874, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01640121266245842, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.018185781314969063, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.015590939670801163, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.01488382089883089, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05473514273762703, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05473514273762703, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.6.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.16716116666793823, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.15775065124034882, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.15457403659820557, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.14059722423553467, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.07886151969432831, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.0755530521273613, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.08733660727739334, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.08086954802274704, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.07954201102256775, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07085856050252914, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.06771925836801529, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.044481098651885986, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.03870061784982681, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.03779464215040207, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.037582673132419586, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.022228501737117767, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.019501926377415657, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.01927804946899414, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.017887165769934654, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.01775527186691761, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.011689381673932076, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01184877473860979, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.011387013830244541, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.007980391383171082, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.044481098651885986, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.044481098651885986, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.6.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.23768509924411774, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.2244160920381546, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.22010710835456848, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.20042294263839722, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11202716082334518, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10741394758224487, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.1242920309305191, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11466553807258606, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11293349415063858, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.1007310226559639, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09645884484052658, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06301391124725342, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05476238578557968, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.053547099232673645, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05326472967863083, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03154074773192406, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.0272939745336771, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02698660083115101, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02497749961912632, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.024789411574602127, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01645590178668499, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.016050072386860847, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.016051465645432472, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010130466893315315, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05476238578557968, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05476238578557968, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.6.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.22949574887752533, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.2056838572025299, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.19669301807880402, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.17584092915058136, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.10529697686433792, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.09681645780801773, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.12346263229846954, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.11345440149307251, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.10873305797576904, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09167252480983734, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.08611638098955154, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06322788447141647, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.05468655005097389, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.0508979931473732, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.04997522383928299, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03202928602695465, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.027079744264483452, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.026899287477135658, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.024468399584293365, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.023785537108778954, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.017732784152030945, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.017970286309719086, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.016541490331292152, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.01313059963285923, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.05468655005097389, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.05468655005097389, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.7.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10184335708618164, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.09504364430904388, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.09225382655858994, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.08318037539720535, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.047364991158246994, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.044784046709537506, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.054359663277864456, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.04967182129621506, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.04787962883710861, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.04201287403702736, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04036158323287964, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.027617568150162697, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.023767590522766113, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.022743944078683853, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.022499801591038704, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.013830862939357758, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.011791280470788479, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.011578885838389397, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.010728172957897186, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01057133823633194, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007293413858860731, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007328340318053961, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.006953178904950619, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.004980811383575201, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.054359663277864456, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.054359663277864456, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.7.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.08729824423789978, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.08142032474279404, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.07892189919948578, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.0710521787405014, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.0406150184571743, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.038315340876579285, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.04632006958127022, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.04277624934911728, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.04104562848806381, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.035976286977529526, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.03430945798754692, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.023462891578674316, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.020403511822223663, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.01943891867995262, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.019202644005417824, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.011722957715392113, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.010005464777350426, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.009809175506234169, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.00907273218035698, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.008925527334213257, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.00611146679148078, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.006133812479674816, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.005789071787148714, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.004038625862449408, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.04632006958127022, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.04632006958127022, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.7.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.21866445243358612, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.20552292466163635, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.20096096396446228, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.18207472562789917, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10288004577159882, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09819462895393372, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11563747376203537, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.10609856247901917, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10382872819900513, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.0918964073061943, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.08781995624303818, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.058675240725278854, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05064156651496887, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.049172889441251755, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.048830822110176086, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.029301606118679047, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.025046277791261673, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.02470063604414463, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.0227820947766304, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.022558089345693588, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.015103607438504696, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.014761107973754406, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.014489569701254368, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.009244582615792751, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05064156651496887, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05064156651496887, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.7.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.20621490478515625, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.18393081426620483, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.17516346275806427, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.14810088276863098, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.09736552834510803, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.08891380578279495, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.1143469512462616, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.10439682006835938, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.09934660792350769, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.0798226147890091, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.07366646826267242, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05894462764263153, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.051085103303194046, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.047989025712013245, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.047257088124752045, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.02986821159720421, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02657957561314106, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.025938894599676132, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02339094504714012, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.022938333451747894, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.016869839280843735, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.01880691759288311, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.015933755785226822, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.015059903264045715, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.051085103303194046, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.051085103303194046, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.7.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.16529476642608643, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.1560569554567337, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.15290027856826782, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.13907451927661896, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.0780198872089386, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.0747465118765831, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.0864333063364029, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.08005785942077637, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.07869257032871246, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07008159905672073, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.06704109907150269, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04400331899523735, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.03833302855491638, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.0374111570417881, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.037193335592746735, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.021985333412885666, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.019364040344953537, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.01914563961327076, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.017776578664779663, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.017642883583903313, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.011575866490602493, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.011863457970321178, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.011269367299973965, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.008091982454061508, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04400331899523735, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04400331899523735, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.7.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.23421674966812134, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.22119775414466858, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.21694090962409973, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.19742250442504883, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.1105773001909256, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10600894689559937, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12230250984430313, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11324401199817657, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11147980391979218, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09943161904811859, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09495532512664795, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.062041424214839935, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05404484272003174, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05283349007368088, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05254771560430527, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.030957741662859917, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.026835257187485695, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.026527592912316322, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.024526335299015045, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.024340452626347542, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.015925297513604164, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.015642235055565834, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.015499722212553024, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009673284366726875, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05404484272003174, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05404484272003174, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.7.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.23695607483386993, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.2120014727115631, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.20255771279335022, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.18090598285198212, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.10878326743841171, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.09985578060150146, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.12757454812526703, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.1172880008816719, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11233802884817123, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09435313194990158, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.08848435431718826, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06517326086759567, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.05645231157541275, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.052599284797906876, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.051666490733623505, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.032885488122701645, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.02797127142548561, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.02779972366988659, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.02519911900162697, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.024509016424417496, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.01810542680323124, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.018549710512161255, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.016901737079024315, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.013566879555583, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.052599284797906876, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.052599284797906876, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.8.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10660949349403381, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.09986569732427597, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.09698982536792755, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.0875658318400383, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.049837928265333176, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.04720121622085571, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.0568976104259491, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05234062671661377, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05033755674958229, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.044394996017217636, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04222710430622101, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.028900280594825745, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.025089388713240623, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.023992540314793587, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.023718303069472313, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.014499671757221222, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.012562558986246586, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.012340222485363483, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.011495338752865791, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.011332252062857151, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007697694934904575, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007977765053510666, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007324327249079943, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005636970512568951, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05234062671661377, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05234062671661377, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.8.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.08779329061508179, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.08221397548913956, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.07964207977056503, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.07192295044660568, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.040986139327287674, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.038691386580467224, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.046819981187582016, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.043310340493917465, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.041366275399923325, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.036497414112091064, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.03477073460817337, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.023715049028396606, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.0206780806183815, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.019631467759609222, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.01937868632376194, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.011854634620249271, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.010126459412276745, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.009917527437210083, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.00922316499054432, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.009062853641808033, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.0061953687109053135, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.006244628690183163, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.00585577730089426, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.004135984927415848, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.046819981187582016, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.046819981187582016, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.8.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.2144792526960373, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.2016303539276123, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.19695669412612915, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.1783994883298874, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10074613243341446, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09606488794088364, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11334003508090973, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.10429476201534271, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10170286893844604, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.09009747952222824, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.08594026416540146, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05751880258321762, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04977520555257797, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.048163942992687225, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.04779091849923134, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.028673620894551277, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02453814446926117, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.024173785001039505, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.022340690717101097, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.022084930911660194, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.014808718115091324, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.014515748247504234, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.014188284054398537, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.009104017168283463, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04977520555257797, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04977520555257797, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.8.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.21312062442302704, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.1889798641204834, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.17889127135276794, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.15196631848812103, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.09976480901241302, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09041709452867508, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11945781111717224, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.10815493762493134, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10198765248060226, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.08154378831386566, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.07598429173231125, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.061225276440382004, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05314325913786888, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.04944229871034622, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.0485311895608902, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03128433600068092, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.027681533247232437, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.026975322514772415, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.024364953860640526, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.023805206641554832, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.018116939812898636, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.019950343295931816, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.01701359637081623, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0161424707621336, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05314325913786888, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05314325913786888, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.8.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.1675155609846115, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.15802247822284698, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.1548094004392624, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.14077234268188477, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.07908222824335098, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.07572297751903534, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.08770499378442764, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.0812222883105278, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.07976614683866501, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.0710403174161911, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.06789162009954453, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.0446629524230957, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.03890034928917885, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.03793127089738846, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.03770630434155464, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.02232459932565689, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.019645126536488533, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.019414113834500313, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.01802688278257847, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.017888037487864494, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.011770842596888542, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.012054555118083954, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01145151723176241, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.008242088370025158, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.0446629524230957, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.0446629524230957, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.8.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.22900786995887756, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.2161126583814621, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.21178913116455078, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.19271841645240784, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.10800553858280182, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10349906980991364, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11963319033384323, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11071155220270157, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.10893340408802032, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09704786539077759, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09270332008600235, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.060720816254615784, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05283007398247719, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.0516192726790905, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.051316797733306885, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.030313869938254356, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.026252448558807373, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02594086341559887, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02398933842778206, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02380099892616272, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.015677111223340034, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.015368604101240635, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.015249955467879772, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009586483240127563, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05283007398247719, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05283007398247719, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.8.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.24474163353443146, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.2192118763923645, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.20965036749839783, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.18653713166713715, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11268772929906845, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10357073694467545, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.1315699815750122, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12115095555782318, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.1163102462887764, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09747559577226639, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09109711647033691, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06715734302997589, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.05840413272380829, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05449794977903366, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.053564343601465225, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03371765464544296, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.0290260910987854, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.028866823762655258, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.02611115574836731, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.025421250611543655, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.018248973414301872, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.019274931401014328, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.01700744964182377, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.014171706512570381, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05449794977903366, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05449794977903366, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.9.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11328333616256714, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.10602722316980362, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.10309509187936783, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.09325914084911346, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.053033698350191116, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.050334133207798004, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06009851396083832, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05534280091524124, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05353463813662529, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.047157250344753265, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04498961195349693, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.030518870800733566, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.026434630155563354, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.025403421372175217, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.025153756141662598, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.015263470821082592, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.013087745755910873, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.01287071406841278, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.011899321340024471, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01174547802656889, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007989969104528427, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007991945371031761, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007633639965206385, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005298607982695103, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05534280091524124, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05534280091524124, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.9.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.0921054482460022, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.08630695194005966, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.08371967077255249, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.07567576318979263, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.04304802045226097, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.040721382945775986, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.04904581233859062, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.0453953817486763, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.043466296046972275, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.03836606442928314, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.03657406195998192, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.024840734899044037, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.021661261096596718, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.02061557210981846, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.020368149504065514, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.012413399294018745, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.010592764243483543, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.010388263501226902, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.009645748883485794, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.009487379342317581, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.006472536362707615, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.006478391587734222, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.006130056455731392, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0042205918580293655, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.04904581233859062, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.04904581233859062, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.9.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.22251014411449432, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.20927345752716064, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.20420604944229126, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.18492133915424347, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10471786558628082, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09978479146957397, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11784599721431732, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.10850488394498825, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10569902509450912, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.09359019249677658, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.08923453837633133, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05976952984929085, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.051841266453266144, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.050114668905735016, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.0496906153857708, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.0298763494938612, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02557878941297531, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.02519550360739231, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.023283226415514946, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.023019544780254364, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.015474229119718075, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.015206633135676384, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.014839858748018742, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.009619629010558128, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.051841266453266144, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.051841266453266144, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.9.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.2296706885099411, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.20317816734313965, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.19318929314613342, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.1666373610496521, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10825809836387634, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09809865057468414, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.1254790723323822, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11522924154996872, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.11023116856813431, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.08852215856313705, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.08203747868537903, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06504452973604202, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05655466392636299, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.0534023642539978, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05262627825140953, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03313198685646057, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.029552865773439407, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.028890814632177353, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02593965083360672, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.025490133091807365, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01896643079817295, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.020836174488067627, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.01805119216442108, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.016712069511413574, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.0534023642539978, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.0534023642539978, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.9.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.16790157556533813, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.15829038619995117, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.15492241084575653, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.14074699580669403, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.07937368750572205, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.07593651115894318, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.08804106712341309, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.08159441500902176, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.08003455400466919, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07119500637054443, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.06799714267253876, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04491860419511795, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.039156436920166016, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.03816234692931175, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.03792373090982437, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.02245444618165493, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.01990632712841034, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.01967144012451172, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.018292715772986412, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.018143342807888985, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.011928102001547813, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.012432040646672249, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.011602964252233505, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.008747303858399391, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04491860419511795, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04491860419511795, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.9.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2286377102136612, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.21557381749153137, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.21131226420402527, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.19200672209262848, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.10792212188243866, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10334702581167221, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11963847279548645, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11065705865621567, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.10883525758981705, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09682455658912659, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09239726513624191, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.060742661356925964, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05282925069332123, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05157856270670891, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05127891153097153, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.030338790267705917, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02624843828380108, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.025937732309103012, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.023968467488884926, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02377542108297348, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.015706244856119156, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01539503876119852, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.015268903225660324, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009634745307266712, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05282925069332123, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05282925069332123, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.9.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.25115105509757996, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.22499220073223114, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.2154412716627121, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.1915069967508316, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11575251072645187, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10658886283636093, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.1345599740743637, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12382151186466217, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11934740841388702, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09998884052038193, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.0932273119688034, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06868241727352142, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.05965328961610794, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05599680915474892, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.055116407573223114, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.034594710916280746, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.029834875836968422, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.029680274426937103, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.026809528470039368, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.026132458820939064, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.018902651965618134, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.01976672001183033, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.01776697300374508, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.014551535248756409, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.055116407573223114, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.055116407573223114, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.10.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11575520038604736, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.10830280929803848, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.10529862344264984, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.09494943916797638, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05408012494444847, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.051277052611112595, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.061571553349494934, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.056573957204818726, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05460824444890022, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.04811084270477295, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04581746459007263, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03127510845661163, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.027010323479771614, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.02589535340666771, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.02562478370964527, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.015623875893652439, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.013311820104718208, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.013072656467556953, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.012097697705030441, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.011921028606593609, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.008139267563819885, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.00809741672128439, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007706001866608858, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005306062288582325, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05460824444890022, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05408012494444847, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.10.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.09460285305976868, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.08858682960271835, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.08576082438230515, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.07742301374673843, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.04415910691022873, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.04169267416000366, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05042199790477753, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.046748578548431396, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.04459258168935776, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.03931369632482529, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.03744832053780556, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.025515025481581688, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.022300170734524727, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.02113872393965721, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.020860273391008377, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.012755300849676132, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.010861330665647984, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.010635227896273136, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.009880303405225277, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.009700285270810127, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.006642933003604412, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.006647471338510513, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.006269354373216629, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.004320094361901283, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05042199790477753, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05042199790477753, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.10.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.223289355635643, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.20976650714874268, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.20428943634033203, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.18480798602104187, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10501758009195328, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09980623424053192, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11917788535356522, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.10949775576591492, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.1060168668627739, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.09376703202724457, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0894089788198471, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.0603877492249012, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.052446700632572174, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.050339967012405396, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.04983196035027504, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03028920292854309, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02585027553141117, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.025409942492842674, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.023558665066957474, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.023234182968735695, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.015871357172727585, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.01568124257028103, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.015126675367355347, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.010214882902801037, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.052446700632572174, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.052446700632572174, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.10.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.21797791123390198, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.19632403552532196, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.1860702931880951, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.16467061638832092, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.1027672290802002, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09259948134422302, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.12281669676303864, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11237675696611404, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10449301451444626, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.08647936582565308, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0827653557062149, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.0632287785410881, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05435016751289368, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.04996966943144798, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.04887823387980461, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03170134127140045, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.026791896671056747, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.025908032432198524, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.023682063445448875, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.02299375645816326, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01714516058564186, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.018156763166189194, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.015656284987926483, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.013457292690873146, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05435016751289368, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05435016751289368, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.10.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.1667235642671585, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.1571052372455597, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.1537599265575409, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.13970673084259033, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.07890576869249344, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.07548347115516663, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.0876510888338089, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.08111801743507385, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.0795988216996193, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07079646736383438, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.0676412433385849, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.044788066297769547, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.03902141749858856, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.03802255913615227, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.0377795435488224, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.02239993028342724, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.019962675869464874, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.019730178639292717, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.01837006025016308, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.01822344958782196, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.011984460055828094, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.012648467905819416, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.011653481982648373, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009112266823649406, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.044788066297769547, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.044788066297769547, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.10.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2247200757265091, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.21189796924591064, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.20768170058727264, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.18865767121315002, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.10607732832431793, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10155405104160309, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11762430518865585, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.10877615213394165, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.10698239505290985, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09518075734376907, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09083583205938339, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.059728384017944336, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05193166062235832, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05069771781563759, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05040421709418297, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.02980130724608898, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.025780489668250084, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.025472749024629593, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02352249249815941, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02333802729845047, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01536722481250763, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01509019173681736, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.014930975623428822, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009387494064867496, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05193166062235832, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05193166062235832, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.10.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.24643345177173615, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.22087419033050537, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.21195237338542938, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.1881127655506134, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11357598006725311, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10482512414455414, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.1318797618150711, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12093394994735718, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11712303012609482, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09816911071538925, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09150116890668869, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06742849200963974, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.058232277631759644, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05489034578204155, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05407306179404259, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03397141396999359, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.029079744592308998, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.02896697446703911, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.026098191738128662, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.025462590157985687, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.018711237236857414, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.019051633775234222, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.01766197569668293, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.013859409838914871, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05489034578204155, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05407306179404259, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.11.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.12013468146324158, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.11223112046718597, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.10898546874523163, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.09833802282810211, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05612140893936157, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.053170882165431976, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.063884437084198, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.058845311403274536, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05677877739071846, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.04992857947945595, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04759090021252632, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03247286006808281, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.028234105557203293, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.027028240263462067, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.026741620153188705, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.016288001090288162, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.014168472960591316, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.013943113386631012, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.012957679107785225, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.012781649827957153, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.008696329779922962, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.009030334651470184, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.008274218998849392, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.006425471045076847, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.04992857947945595, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.04992857947945595, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.11.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.09857289493083954, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.0922558456659317, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.08933562785387039, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.08062046766281128, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.04600818082690239, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.04343699291348457, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05253827944397926, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.04862171411514282, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.04646604508161545, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.04094399884343147, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0390017069876194, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.026591399684548378, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.023206891492009163, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.022052988409996033, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.021774668246507645, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.013292500749230385, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.01138047594577074, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.011152585968375206, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.010365087538957596, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.010186469182372093, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.006957405712455511, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007034855429083109, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.006578077096492052, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.004679253790527582, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05253827944397926, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05253827944397926, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.11.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.23099572956562042, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.21697869896888733, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.21169403195381165, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.1914314329624176, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10869614034891129, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.10341185331344604, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.12289614975452423, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.1127980649471283, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10968709737062454, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.0970531776547432, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.09254349768161774, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06241994723677635, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05402009189128876, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.051991645246744156, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05150841921567917, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03119262121617794, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.026576917618513107, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.026154421269893646, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.024208571761846542, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.023890340700745583, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.016218725591897964, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.015920676290988922, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.015472875908017159, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.010140799917280674, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05402009189128876, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05402009189128876, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.11.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.22974367439746857, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.21051307022571564, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.20284363627433777, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.17809493839740753, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10792499780654907, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.10028769075870514, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.12597304582595825, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11429228633642197, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10957165062427521, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.09261023253202438, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0867241844534874, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06484522670507431, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05584903806447983, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.053008463233709335, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05229606851935387, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.033103544265031815, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02920886129140854, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.028653979301452637, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.026383385062217712, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.025951886549592018, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.019088787958025932, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.020454131066799164, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.018273109570145607, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0162691380828619, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.053008463233709335, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.053008463233709335, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.11.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.16964486241340637, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.15983115136623383, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.15653876960277557, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.14217126369476318, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.08040313422679901, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.07689076662063599, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.08930283039808273, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.08271743357181549, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.08113682270050049, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07210689038038254, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.06882214546203613, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04560073837637901, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.03978803753852844, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.03875354677438736, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.03850709646940231, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.022819282487034798, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.020335625857114792, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.020096395164728165, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.018694467842578888, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.018544277176260948, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.012175414711236954, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.012864950112998486, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.011833637952804565, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009250564500689507, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04560073837637901, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04560073837637901, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.11.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.23181182146072388, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.21845240890979767, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.2140103131532669, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.19446209073066711, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.1095413789153099, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10485126823186874, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12139355391263962, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11238618195056915, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11049038916826248, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09822939336299896, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.0936540812253952, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06170522794127464, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.053685300052165985, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.052388109266757965, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.052075378596782684, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.030788231641054153, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.026658879593014717, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.0263306125998497, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02430974319577217, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02411551959812641, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.015877675265073776, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.015628907829523087, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.015428347513079643, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.00976134929805994, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.053685300052165985, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.053685300052165985, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.11.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.2506487965583801, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.2242073118686676, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.21476896107196808, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.19050998985767365, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11555956304073334, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10631059110164642, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.13420893251895905, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12349814176559448, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.1192471981048584, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09962307661771774, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09284418076276779, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06854186952114105, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.05950034782290459, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.055860381573438644, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05498473718762398, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.0344901904463768, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.02971925027668476, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.029584672302007675, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.026666052639484406, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.025989070534706116, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.018811900168657303, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.019652780145406723, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.017674937844276428, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.014421431347727776, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05498473718762398, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05498473718762398, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.12.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.12416593730449677, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.11621439456939697, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11306313425302505, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.10210097581148148, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05821267515420914, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.05515816807746887, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06593962758779526, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.060788899660110474, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05878668278455734, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05179722234606743, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.049348849803209305, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03347526863217354, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.02906990237534046, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.027924692258238792, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.02764442004263401, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.01674547605216503, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.014425316825509071, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.014189456589519978, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.013128497637808323, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.012958195060491562, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.008772403933107853, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.008877690881490707, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.008382236585021019, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0059613934718072414, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05179722234606743, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05179722234606743, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.12.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10337495803833008, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.0968264788389206, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.09394378960132599, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.08472255617380142, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.0483437180519104, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.045713379979133606, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05502081662416458, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05095774680376053, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.04880066215991974, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.0430448092520237, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0410052165389061, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.027879180386662483, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.024337971583008766, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.023148853331804276, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.022859858348965645, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.013928670436143875, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.01189111452549696, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.011663790792226791, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.010819475166499615, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.010639344342052937, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007253736257553101, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007267061620950699, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.006864569149911404, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.004734116140753031, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05502081662416458, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05502081662416458, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.12.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.233722984790802, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.2197664976119995, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.21437935531139374, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.19396743178367615, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.1101539134979248, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.10477549582719803, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.12471885234117508, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.1145096868276596, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.1111585721373558, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.09851690381765366, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.09394589066505432, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06337975710630417, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05489477887749672, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05277217552065849, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.0522448755800724, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03172106295824051, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.027044476941227913, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.026586251333355904, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02465032786130905, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.024305345490574837, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.016560565680265427, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.016306709498167038, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.015773629769682884, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.010483073070645332, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05489477887749672, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05489477887749672, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.12.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.23274850845336914, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.21189408004283905, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.20472612977027893, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.17786821722984314, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.11021587252616882, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.10262955725193024, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.12586864829063416, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11527027189731598, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.11163853853940964, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.09355314075946808, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.08641223609447479, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.0649508535861969, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05617186054587364, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.053870782256126404, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05332336947321892, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.032847460359334946, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.029233738780021667, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.02870405837893486, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.026084892451763153, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.025740472599864006, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.018477454781532288, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.019851859658956528, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.017795505002141, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.01542653888463974, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.053870782256126404, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.053870782256126404, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.12.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.1761312633752823, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.16600100696086884, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.16254912316799164, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.14761778712272644, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.08349128812551498, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.0798494815826416, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09266044944524765, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.08583036810159683, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.08421371877193451, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07486599683761597, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07145937532186508, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04728928580880165, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04128485918045044, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.0402311310172081, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.03997332602739334, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.023666363209486008, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.021092643961310387, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02084241434931755, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.019398504868149757, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.01924082264304161, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.012613714672625065, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01332453265786171, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.012265476398169994, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009561100043356419, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04728928580880165, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04728928580880165, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.12.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.23398733139038086, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.22068427503108978, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.2162487804889679, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.1964343637228012, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11060921102762222, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10593333840370178, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12255144864320755, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11349467933177948, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11159119009971619, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09923016279935837, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09466513246297836, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06227894872426987, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05416695401072502, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05287613347172737, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05256786197423935, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.031064452603459358, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.026882119476795197, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.026555484160780907, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.024518121033906937, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.024317480623722076, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01601451076567173, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.015719298273324966, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01556489709764719, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009768064133822918, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05416695401072502, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05416695401072502, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.12.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.25690487027168274, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.22924146056175232, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.21956290304660797, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.1945238709449768, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11829499900341034, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.1087358295917511, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.13731832802295685, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12627524137496948, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12212639302015305, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10165292769670486, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09449983388185501, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.0700778141617775, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.06064693257212639, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05705789849162102, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05619371682405472, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03516390174627304, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.03002944588661194, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.029905973002314568, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.026783427223563194, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.026095449924468994, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.018997035920619965, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.01944311521947384, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.01786058023571968, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.013842159882187843, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03516390174627304, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03516390174627304, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.13.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.12737974524497986, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.11952639371156693, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11637944728136063, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.10530716180801392, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05979878082871437, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.05684050917625427, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06778743863105774, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.06236205995082855, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.06036613509058952, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05337495729327202, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.050721850246191025, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03444444760680199, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.029823381453752518, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.028675569221377373, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.028401175513863564, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.01722598262131214, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.01480015553534031, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.01456363219767809, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.013505863957107067, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01333580631762743, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.009043676778674126, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.00906770396977663, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.00866397749632597, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0060590230859816074, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05337495729327202, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05337495729327202, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.13.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.1059303879737854, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.09938158839941025, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.09644481539726257, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.08720505237579346, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.04968523979187012, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.04706841707229614, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05640461668372154, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05227169021964073, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.050147201865911484, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.044301778078079224, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04222197085618973, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.02858336828649044, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.02495630271732807, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.023800130933523178, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.023538250476121902, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.01427834015339613, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.012246443890035152, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.01201498880982399, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.01115673128515482, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.010980430990457535, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007449113763868809, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007485765963792801, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.00707654282450676, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.004904462955892086, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05227169021964073, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05227169021964073, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.13.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.24742157757282257, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.23289933800697327, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.22746673226356506, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.2060794234275818, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.11686781793832779, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.11140629649162292, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.13118253648281097, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.12082646787166595, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.11787980794906616, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.10448215156793594, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.09965366870164871, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06664159148931503, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05775140970945358, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05591568350791931, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05547553673386574, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03329803794622421, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.0285025667399168, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.02808557078242302, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.025951478630304337, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.02566509135067463, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.017202366143465042, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.016864966601133347, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.016509732231497765, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.010589123703539371, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05547553673386574, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05547553673386574, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.13.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.22659045457839966, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.20729190111160278, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.20062460005283356, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.17192061245441437, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10669413954019547, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09994251281023026, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.12214876711368561, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11183875799179077, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10823892056941986, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.09065482020378113, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0831962451338768, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.063250333070755, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05466105043888092, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05240471661090851, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.051840778440237045, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03205862641334534, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.028808174654841423, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.028338907286524773, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02585459128022194, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.02553727477788925, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.018271394073963165, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.0200175940990448, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.017599785700440407, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.015948588028550148, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05466105043888092, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05466105043888092, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.13.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.17392826080322266, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.163940891623497, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.16044890880584717, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.14574159681797028, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.08236288279294968, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.07876268774271011, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09150349348783493, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.08475159108638763, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.08309942483901978, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07388166338205338, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07057437300682068, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04670930281281471, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.040704283863306046, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.03963096812367439, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.039385341107845306, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.023350823670625687, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.020684830844402313, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02043813280761242, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.018997814506292343, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.018837599083781242, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.012381031177937984, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01293931994587183, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.012031744234263897, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009120429866015911, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04670930281281471, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04670930281281471, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.13.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.23159655928611755, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.21838650107383728, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.21393774449825287, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.19436757266521454, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.10943730920553207, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.1048007383942604, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12136166542768478, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11228766292333603, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11037635803222656, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09815843403339386, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.0936410203576088, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06165686249732971, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0536210834980011, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05232381448149681, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05201483517885208, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03076857514679432, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02662092261016369, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02629205770790577, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.0242854505777359, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.024092117324471474, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01588122919201851, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.015594377182424068, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.015425285324454308, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009719962254166603, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0536210834980011, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0536210834980011, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.13.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.25390625, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.22631841897964478, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.21597445011138916, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.19145940244197845, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11678752303123474, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10689250379800797, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.1367070972919464, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.1256645917892456, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12069740146398544, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10030035674571991, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.093394435942173, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06983862072229385, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.060411158949136734, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05636109784245491, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05536354333162308, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.0351734459400177, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.029732823371887207, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.029552830383181572, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.026533590629696846, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.025776199996471405, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.019175436347723007, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.019420314580202103, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.017900966107845306, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.013869773596525192, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05536354333162308, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05536354333162308, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.14.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.12544545531272888, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.11755776405334473, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11452525854110718, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.10343476384878159, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05869847908616066, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.055760595947504044, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06654619425535202, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.06119132414460182, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.059245266020298004, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.052331842482089996, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04990033432841301, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.033843908458948135, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.029212551191449165, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.02810794487595558, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.027844706550240517, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.016924748197197914, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.014456058852374554, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.01422027312219143, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.013171939179301262, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.013000795617699623, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.008836332708597183, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.008801177144050598, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.00839376449584961, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005796974524855614, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.052331842482089996, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.052331842482089996, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.14.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10232672095298767, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.09582652896642685, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.09279412776231766, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.08380669355392456, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.047764282673597336, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.04512177035212517, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05449947714805603, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05051400512456894, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.04822247847914696, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.042567066848278046, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04058414697647095, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.027569063007831573, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.024108625948429108, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.022864416241645813, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.02256491221487522, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.013778818771243095, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.011755087412893772, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.011515045538544655, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.010702794417738914, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.010513617657124996, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007181297522038221, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007203647401183844, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.006787263322621584, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.004692533053457737, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05449947714805603, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05449947714805603, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.14.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.24727283418178558, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.23270830512046814, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.2273685485124588, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.20585182309150696, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.11664842069149017, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.1112423688173294, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.13123446702957153, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.12079180777072906, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.11764666438102722, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.10435649007558823, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0996059849858284, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06672246009111404, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05778435245156288, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.055822696536779404, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.055360134690999985, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.033348243683576584, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02848481573164463, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.028050195425748825, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.025962799787521362, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.02565223164856434, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.0172954723238945, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.016948342323303223, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.016558725386857986, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.010711267590522766, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.055360134690999985, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.055360134690999985, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.14.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.23562301695346832, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.21548303961753845, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.2077796310186386, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.18297800421714783, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.11136942356824875, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.10353206843137741, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.1275085210800171, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.1172449067234993, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.11324009299278259, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.09576085954904556, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0884755477309227, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06575103104114532, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05686778202652931, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.054183926433324814, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05351896956562996, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.033156294375658035, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.028987308964133263, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.028400737792253494, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.025975368916988373, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.02557283639907837, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.018334897235035896, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.019277963787317276, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.017535215243697166, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.014466906897723675, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.054183926433324814, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.054183926433324814, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.14.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.17784477770328522, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.16754589974880219, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.16397330164909363, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.14902576804161072, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.08430559188127518, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.0806126669049263, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09356632828712463, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.08671831339597702, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.08503930270671844, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07556167244911194, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.0721340999007225, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04775281995534897, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04167211800813675, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04059070348739624, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04033271595835686, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.023875975981354713, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.021234046667814255, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02098165825009346, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.019507041200995445, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.01934870146214962, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.012686440721154213, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.013356138952076435, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.012328021228313446, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009503725916147232, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04775281995534897, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04775281995534897, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.14.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2345641702413559, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.22117598354816437, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.21666403114795685, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.1968408226966858, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11086520552635193, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10615295171737671, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12290678918361664, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11377697438001633, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11182735860347748, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09946697950363159, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.0948508232831955, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06246713921427727, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.054331373423337936, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.053024616092443466, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05270949751138687, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.031151745468378067, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02698369510471821, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.026656191796064377, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02461693435907364, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02441822737455368, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01607854664325714, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.015826139599084854, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.015624849125742912, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009891537949442863, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.054331373423337936, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.054331373423337936, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.14.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.2580743134021759, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.2297893613576889, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.21919845044612885, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.19421564042568207, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.1186269223690033, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10848231613636017, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.13913092017173767, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12787014245986938, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12269410490989685, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10184568166732788, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09499222785234451, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.07089388370513916, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.061339858919382095, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05714346468448639, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05612809583544731, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.035535778850317, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.029974980279803276, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.0298004113137722, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.026696424931287766, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.025908229872584343, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.019097093492746353, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.0193780530244112, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.017760369926691055, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.013572805561125278, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.035535778850317, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.035535778850317, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.15.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.12184853106737137, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.11426477879285812, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11115772277116776, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.10062526911497116, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05718734860420227, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.05428460240364075, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06536620110273361, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.059785276651382446, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05780843645334244, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.051069509238004684, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0490424782037735, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.033255551010370255, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.02858944982290268, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.027422217652201653, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.027144744992256165, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.01662612520158291, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.014122091233730316, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.013885816559195518, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.012893421575427055, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.012717541307210922, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.008717727847397327, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.00862840935587883, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.00830375961959362, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005700133740901947, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.051069509238004684, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.051069509238004684, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.15.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10217344760894775, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.0958079844713211, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.0928269773721695, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.08382687717676163, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.04780348390340805, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.04519207775592804, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.0545656755566597, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.0506138801574707, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.04827575758099556, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.042646605521440506, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04067663475871086, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.02762668766081333, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.02414286881685257, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.022893080487847328, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.022588057443499565, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.0138116255402565, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.011759566143155098, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.01152023859322071, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.010711329989135265, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.010522154159843922, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007200552150607109, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007190415635704994, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.006800884380936623, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.004667708184570074, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.0545656755566597, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.0545656755566597, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.15.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.24366679787635803, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.22935734689235687, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.2239731103181839, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.2029893547296524, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.11499956250190735, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.1096528097987175, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.129389226436615, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.1190263032913208, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.11600479483604431, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.10293638706207275, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.09820596128702164, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06581667065620422, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.056907329708337784, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05498838424682617, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.054514940828084946, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03282862529158592, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.027972310781478882, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.027543917298316956, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.025478465482592583, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.02518751658499241, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.016946783289313316, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.01651027426123619, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.016207538545131683, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.010244163684546947, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05498838424682617, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.054514940828084946, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.15.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.2276056557893753, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.2066173404455185, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.19769421219825745, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.17434866726398468, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10777036100625992, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09900964051485062, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.12559452652931213, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11486023664474487, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10912070423364639, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.0913262739777565, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0856088325381279, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06485233455896378, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05576727166771889, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05246071517467499, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.051649559289216995, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03268028795719147, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.028228821232914925, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.027545690536499023, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.025093451142311096, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.02459992840886116, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01810375414788723, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.01904609613120556, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.01702992618083954, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.01442215871065855, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05576727166771889, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05576727166771889, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.15.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.17779076099395752, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.16753120720386505, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.1639314442873001, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.14898189902305603, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.08418850600719452, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.0805230513215065, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09351586550474167, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.08663728088140488, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.08494924753904343, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07547587901353836, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07208690792322159, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.047678813338279724, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0415974035859108, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04050225764513016, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.040238186717033386, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.023846406489610672, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.021126501262187958, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02086876891553402, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.019392674788832664, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.01923428289592266, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.0126313716173172, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01320100948214531, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.012268545106053352, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.00928643811494112, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.047678813338279724, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.047678813338279724, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.15.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.23192739486694336, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.21860171854496002, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.21417371928691864, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.1945660412311554, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.10963404923677444, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10490051656961441, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12159926444292068, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11250996589660645, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11056723445653915, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09829708933830261, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09377047419548035, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06178216636180878, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05372793599963188, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.052410148084163666, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05210075527429581, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.030818432569503784, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.026708291843533516, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.026381265372037888, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.024369388818740845, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02416980266571045, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.015936624258756638, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01570681296288967, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.015484807081520557, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009876268915832043, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05372793599963188, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05372793599963188, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.15.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.2564137876033783, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.22750484943389893, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.2167920470237732, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.19214056432247162, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11775375157594681, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10751695930957794, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.1382044404745102, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.1269780993461609, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.1219959706068039, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10093853622674942, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09419186413288116, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.07063528150320053, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.06109276041388512, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.056838322430849075, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05581580474972725, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.0356125645339489, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.03001324273645878, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.02986009418964386, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.026746133342385292, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.02597801573574543, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.019471263512969017, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.019649308174848557, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.01816253922879696, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.014076781459152699, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05581580474972725, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05581580474972725, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.16.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.12456850707530975, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.11708056181669235, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11401226371526718, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.10294981300830841, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05858221650123596, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.055731773376464844, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06708700954914093, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.06107744947075844, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05912451818585396, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.052283938974142075, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.05033528059720993, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.034142035990953445, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.02928842604160309, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.028178280219435692, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.02791585586965084, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.01716155745089054, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.014587676152586937, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.014358000829815865, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.013316038995981216, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.013150149956345558, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.009106720797717571, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.009004245512187481, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.008746032603085041, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.006092672236263752, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.052283938974142075, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.052283938974142075, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.16.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10053832828998566, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.09427779167890549, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.09130527824163437, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.08251979202032089, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.04705847054719925, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.044441837817430496, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.0538698248565197, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.04989807307720184, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.04749618098139763, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.04200732335448265, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04007437080144882, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.027233464643359184, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.02380598522722721, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.022531181573867798, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.022228218615055084, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.013613772578537464, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.011574880219995975, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.011326915584504604, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.010547368787229061, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.010346529074013233, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007085199002176523, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007080895826220512, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.006684840191155672, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.004580037202686071, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.0538698248565197, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.0538698248565197, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.16.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.2598143517971039, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.24481624364852905, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.2392958253622055, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.21725815534591675, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.12297270447015762, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.11733942478895187, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.13783937692642212, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.12698742747306824, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.12403133511543274, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.1101292073726654, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.1051798090338707, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.07014167308807373, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.060720864683389664, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.058827899396419525, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05836230888962746, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.034991584718227386, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.029925476759672165, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.029496099799871445, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.027271326631307602, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.026984086260199547, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01801556535065174, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.0176200233399868, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.01728886552155018, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.01094543281942606, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.034991584718227386, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.034991584718227386, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.16.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.2214815318584442, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.1989356130361557, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.1914217472076416, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.1644086241722107, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10433138161897659, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09607933461666107, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11857375502586365, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.10921580344438553, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10621238499879837, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.0867103859782219, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0799918994307518, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.061213843524456024, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.053236398845911026, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.051058992743492126, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.050537340342998505, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03091214969754219, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.027838721871376038, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.027318745851516724, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.024560488760471344, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.024233371019363403, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01733780838549137, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.019065197557210922, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.01667795516550541, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.01497691310942173, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.053236398845911026, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.053236398845911026, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.16.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.17441916465759277, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.16444909572601318, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.16104142367839813, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.14637330174446106, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.0825732871890068, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.07900498062372208, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.0916096568107605, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.08483733236789703, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.08327886462211609, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07405845820903778, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07078280299901962, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.046718429774045944, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04072459042072296, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.039694301784038544, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.03945635259151459, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.023352133110165596, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02069253847002983, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.020447446033358574, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.019003864377737045, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.018854830414056778, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.012391185387969017, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.012891451828181744, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.012053207494318485, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009049600921571255, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.046718429774045944, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.046718429774045944, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.16.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2268962264060974, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.21402230858802795, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.20974808931350708, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.1906869262456894, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.10719393938779831, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10267983376979828, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11870633810758591, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.10992981493473053, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.10810956358909607, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09621351212263107, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.0918264165520668, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.060325104743242264, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0524941049516201, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05124523490667343, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05094806104898453, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.030094088986516, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.026077402755618095, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.025765802711248398, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.023809684440493584, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02361944317817688, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.015534733422100544, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.015296039171516895, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.015104849822819233, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009563595056533813, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0524941049516201, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0524941049516201, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.16.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.24556539952754974, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.217830628156662, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.2069813460111618, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.1838536262512207, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11267393082380295, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10250446200370789, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.13286708295345306, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.1223670095205307, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11671176552772522, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09655257314443588, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09039022773504257, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06787773221731186, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.05876461789011955, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.054384078830480576, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05331544578075409, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03409901261329651, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.028702011331915855, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.028483891859650612, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.02556738257408142, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.024765189737081528, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.018406644463539124, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.018823375925421715, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.0169904213398695, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.013393591158092022, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.054384078830480576, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.054384078830480576, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.17.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.12492015212774277, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.11769643425941467, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11476157605648041, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.10415618121623993, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.058717940002679825, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.05591683089733124, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06680257618427277, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.06112510338425636, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.0592399463057518, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05264671891927719, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.05067705363035202, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.033947452902793884, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.029253128916025162, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.028175443410873413, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.027924664318561554, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.017048683017492294, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.014501756988465786, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.014275239780545235, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.013277937658131123, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01311474945396185, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.008965063840150833, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.008813287131488323, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.008621817454695702, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005799137521535158, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05264671891927719, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05264671891927719, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.17.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10446635633707047, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.09813770651817322, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.09527302533388138, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.08629649132490158, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.04894807189702988, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.04639124497771263, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.055677030235528946, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05158892273902893, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.049397535622119904, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.043790895491838455, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0418223962187767, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.028198732063174248, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.024613704532384872, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.02344845049083233, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.02317136526107788, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.014091952703893185, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.012059474363923073, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.011830800212919712, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.011018330231308937, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.010836876928806305, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.00735792052000761, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.0073785376735031605, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.006983059924095869, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0048273298889398575, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.055677030235528946, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.055677030235528946, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.17.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.24835455417633057, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.2342352271080017, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.22906416654586792, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.20810402929782867, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.11734279245138168, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.1120728999376297, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.1313760131597519, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.12114274501800537, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.11834364384412766, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.10529898852109909, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.10066960006952286, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.0667576938867569, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05787770077586174, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05607801303267479, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05565608665347099, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03332441300153732, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.028492745012044907, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.028086015954613686, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.025990791618824005, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.0257167536765337, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.017132068052887917, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.016703015193343163, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.01645379140973091, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.010275859385728836, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05565608665347099, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05565608665347099, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.17.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.22057615220546722, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.20078155398368835, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.19374071061611176, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.16555438935756683, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10406278818845749, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09694267809391022, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11843913048505783, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.10840486735105515, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10549495369195938, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.08661575615406036, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0801931694149971, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06115410104393959, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.053012192249298096, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05107659474015236, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.050629131495952606, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.030918342992663383, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.027992375195026398, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.027567358687520027, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.024757742881774902, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.024474194273352623, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.017516080290079117, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.019318953156471252, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.016905441880226135, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.015332142822444439, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.053012192249298096, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.053012192249298096, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.17.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.17920807003974915, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.16894814372062683, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.1654949188232422, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.15042522549629211, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.08480190485715866, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.08118552714586258, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09394297748804092, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.0870596170425415, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.08552305400371552, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07610448449850082, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07268858700990677, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.0478590652346611, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.041722215712070465, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.040701594203710556, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04046439751982689, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.023913301527500153, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.021064655855298042, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.020821254700422287, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.019311532378196716, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.01916024088859558, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.012572027742862701, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.012895447202026844, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.012242146767675877, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.008798995986580849, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.0478590652346611, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.0478590652346611, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.17.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2347826361656189, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.22153432667255402, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.21713292598724365, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.19754256308078766, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11094895750284195, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10628801584243774, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12268248200416565, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11371606588363647, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11188702285289764, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09962248057126999, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09500861912965775, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.062280960381031036, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05429321527481079, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.053030986338853836, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.052740681916475296, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.031077632680535316, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.026953788474202156, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.026630571112036705, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.024595588445663452, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.024408681318163872, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.015988247469067574, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01573449745774269, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01556260697543621, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009756099432706833, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05429321527481079, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05429321527481079, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.17.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.24803459644317627, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.22053183615207672, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.21003517508506775, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.18690438568592072, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.1137990802526474, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.1038358137011528, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.1335625797510147, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12320761382579803, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11782516539096832, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09799497574567795, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09174426645040512, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06827274709939957, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.05916014313697815, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.054834961891174316, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05377964675426483, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03440406545996666, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.02876039780676365, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.028556598350405693, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.02565905638039112, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.024858813732862473, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.018706945702433586, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.01862308569252491, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.017378242686390877, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.012996941804885864, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.054834961891174316, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05377964675426483, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.18.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.12573249638080597, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.11831645667552948, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11539861559867859, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.10478005558252335, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05918480083346367, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.05639307573437691, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06743433326482773, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.06159477308392525, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05971737578511238, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05309103801846504, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.05121232569217682, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03432724252343178, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.02947038598358631, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.028405314311385155, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.028154194355010986, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.017234429717063904, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.01458627451211214, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.014364547096192837, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.013350412249565125, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.0131929786875844, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.009025697596371174, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.008807371370494366, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.008685405366122723, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0057244617491960526, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05309103801846504, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05309103801846504, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.18.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10868475586175919, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.10219308733940125, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.09930822253227234, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.09011692553758621, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.051037780940532684, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.048463162034749985, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05789477378129959, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05359835922718048, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05152016133069992, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.045723702758550644, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04368000477552414, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.029333552345633507, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.025564268231391907, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.024449961259961128, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.024181026965379715, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.014652390033006668, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.012546653859317303, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.012325349263846874, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.011460062116384506, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01128623727709055, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007636739872395992, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007621710188686848, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007277813274413347, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.004949438851326704, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05359835922718048, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05359835922718048, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.18.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.25414636731147766, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.23971526324748993, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.23457348346710205, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.21325039863586426, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.12015889585018158, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.1148720532655716, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.13425025343894958, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.12390565872192383, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.12119385600090027, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.10781047493219376, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.10309696197509766, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06822380423545837, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05918441340327263, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05743676424026489, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.057018399238586426, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03405112773180008, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02915506437420845, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.028753990307450294, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.026605430990457535, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.026334160938858986, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.017496692016720772, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.017064308747649193, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.016838885843753815, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.010472538881003857, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03405112773180008, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03405112773180008, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.18.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.22510766983032227, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.19675549864768982, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.1858450174331665, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.15696381032466888, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10556783527135849, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09475957602262497, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.12356973439455032, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11338913440704346, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10801465064287186, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.08433875441551208, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.07625599205493927, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.0635780617594719, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05494004860520363, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05144390091300011, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05057760328054428, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03201630711555481, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02770957350730896, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.026960501447319984, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02372227981686592, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.02316700480878353, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.017697762697935104, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.0188076701015234, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.016629232093691826, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.014272736385464668, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05494004860520363, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05494004860520363, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.18.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.18669772148132324, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.17607085406780243, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.1725565642118454, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.15704184770584106, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.08835092931985855, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.08457572013139725, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09780536592006683, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09062650054693222, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.08909532427787781, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.07934518158435822, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07578343898057938, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.0498187392950058, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04342050105333328, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.042395494878292084, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04214651882648468, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.02488519996404648, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02194100059568882, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.021686352789402008, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02013634331524372, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.019981902092695236, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.013087954372167587, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01344526931643486, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01274678111076355, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009184243157505989, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.0498187392950058, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.0498187392950058, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.18.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.24078072607517242, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.2272462397813797, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.222821906208992, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.20277045667171478, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.1137470230460167, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10901272296905518, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.1257009506225586, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11657256633043289, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11471147835254669, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10224594175815582, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09760653972625732, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06383872777223587, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05562656745314598, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.054361045360565186, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05405759811401367, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.031844232231378555, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.027620848268270493, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.027297310531139374, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.025233808904886246, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.025040971115231514, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01638742722570896, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01611393503844738, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.015962377190589905, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009983822703361511, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05562656745314598, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05405759811401367, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.18.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.24773521721363068, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.21931077539920807, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.20810416340827942, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.18527738749980927, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11343701183795929, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10289786756038666, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.1345621794462204, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12378192692995071, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11772167682647705, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09745801240205765, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09138821065425873, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06873684376478195, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.05943546071648598, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05463419854640961, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.053459733724594116, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.034672390669584274, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.02861306443810463, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.0283672958612442, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.025466056540608406, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.024591835215687752, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.01891372725367546, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.018553858622908592, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.017415205016732216, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.01281932182610035, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05463419854640961, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05463419854640961, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.19.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.12865589559078217, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.12097233533859253, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11794671416282654, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.10717602074146271, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.06046721339225769, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.057549212127923965, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06911508738994598, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.06299655884504318, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.06101032346487045, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.054235342890024185, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.052464041858911514, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03524404391646385, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.03018767386674881, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.029029719531536102, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.028762949630618095, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.017675768584012985, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.01488480344414711, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.01464757602661848, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.013623871840536594, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.013448241166770458, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.009231875650584698, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.008974751457571983, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.008868617936968803, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005801368970423937, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.054235342890024185, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.054235342890024185, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.19.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.10610787570476532, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.09965331107378006, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.09665215760469437, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.08765887469053268, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.04969891160726547, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.04706355184316635, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.05666752904653549, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05245157703757286, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05015062168240547, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.044472139328718185, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.042490534484386444, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.0286762285977602, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.025006627663969994, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.023786863312125206, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.023493150249123573, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.014328550547361374, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.012207959778606892, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.011961027979850769, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.011142086237668991, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01095510646700859, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007475107908248901, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007428009063005447, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007081033196300268, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.004785921890288591, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05245157703757286, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05245157703757286, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.19.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.2526075839996338, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.23799097537994385, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.23284322023391724, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.21164965629577637, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.11912423372268677, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.11387841403484344, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.13305433094501495, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.12278294563293457, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.12019495666027069, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.10690423101186752, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.10215602815151215, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06759096682071686, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05860161781311035, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.056933410465717316, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05652796849608421, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.033712174743413925, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02892223559319973, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.028540082275867462, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.026378890499472618, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.026128660887479782, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.017337966710329056, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.01694769226014614, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.016709137707948685, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.010465324856340885, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.033712174743413925, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.033712174743413925, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.19.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.22536911070346832, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.19546762108802795, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.1853315681219101, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.15866680443286896, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10575375705957413, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.0949752926826477, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.12187464535236359, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.1115664541721344, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10788227617740631, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.08400039374828339, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.07611096650362015, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06273064017295837, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.054317817091941833, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05171440914273262, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05106789991259575, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03175849840044975, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.028131810948252678, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.027503041550517082, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02407749928534031, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.023698633536696434, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.017811927944421768, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.0192661564797163, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.01701580546796322, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.015032398514449596, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.054317817091941833, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.054317817091941833, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.19.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.19655397534370422, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.1854104995727539, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.18174795806407928, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.1654091626405716, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.09296967089176178, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.0890861228108406, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10290973633527756, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09535378217697144, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.09378258138895035, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.0835360437631607, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07984631508588791, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05240233242511749, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.045664019882678986, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04457980394363403, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04432924836874008, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.02617432177066803, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.023005113005638123, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02274375408887863, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.021090054884552956, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.0209332387894392, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.013724149204790592, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01397770270705223, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.013366847299039364, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009419785812497139, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05240233242511749, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05240233242511749, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.19.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.24241283535957336, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.22882017493247986, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.2243860960006714, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.20426255464553833, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11456773430109024, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10977515578269958, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12667840719223022, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11734557896852493, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11551059037446976, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10294977575540543, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09826786071062088, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06432472914457321, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.056021638214588165, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05474841222167015, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05444183200597763, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.032116834074258804, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.027819950133562088, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02749878540635109, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02540988102555275, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02521936036646366, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01657404564321041, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.016240503638982773, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.016141844913363457, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010077339597046375, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05474841222167015, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05444183200597763, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.19.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.250781387090683, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.222235307097435, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.21081338822841644, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.18772654235363007, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11472588777542114, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10427171736955643, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.13660675287246704, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.1255098581314087, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11912214010953903, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09883954375982285, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.0927816852927208, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06969510018825531, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.0602794848382473, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05535200238227844, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.054149094969034195, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03518933802843094, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.029132898896932602, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.028886649757623672, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.026026947423815727, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.025155991315841675, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.019198033958673477, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.019075745716691017, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.01763223670423031, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.013413256034255028, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05535200238227844, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.054149094969034195, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.20.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.12915150821208954, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.12147461622953415, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11840514838695526, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.10749921202659607, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.06072058156132698, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.057783909142017365, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06962110102176666, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.06330258399248123, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.06126369908452034, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05447010323405266, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.05259517952799797, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.0354139506816864, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.030351625755429268, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.029197299852967262, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.028925422579050064, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.017831487581133842, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.01507392805069685, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.014830879867076874, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.013812580145895481, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.013640740886330605, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.009453367441892624, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.00920812040567398, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.009089510887861252, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.006104159168899059, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05447010323405266, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05447010323405266, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.20.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.1091337576508522, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.10255641490221024, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.09950520098209381, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.09021362662315369, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05119180306792259, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.048491787165403366, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.058337997645139694, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05392671376466751, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05164627730846405, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.045788057148456573, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04376347362995148, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.029522860422730446, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.025727594271302223, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.024499230086803436, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.024202773347496986, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.01475599966943264, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.0125694889575243, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.012324796989560127, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.01147802546620369, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01128850132226944, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007708135060966015, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007641404401510954, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007312479428946972, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.004928771872073412, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05392671376466751, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05392671376466751, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.20.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.2625221014022827, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.24763016402721405, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.24245485663414001, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.2205376923084259, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.12405066937208176, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.11871691793203354, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.1382632851600647, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.12760403752326965, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.12511830031871796, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.1113571897149086, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.10647068917751312, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.0702514722943306, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.060884080827236176, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05926799029111862, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05888897553086281, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03502791374921799, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.030048897489905357, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.029654353857040405, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.027400115504860878, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.027158530429005623, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.017964452505111694, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.017473232001066208, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.017345363274216652, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.010623985901474953, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03502791374921799, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03502791374921799, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.20.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.18831877410411835, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.16578511893749237, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.15808378159999847, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.13293752074241638, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.08758462965488434, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.07888194918632507, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.10141944140195847, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.09294569492340088, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.0890139490365982, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.06970137357711792, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.06350491940975189, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05225534737110138, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.045474447309970856, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.04312147572636604, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.04254807159304619, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.026541627943515778, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02385532669723034, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.023313000798225403, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02061566151678562, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.020240647718310356, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.015149924904108047, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.01677580364048481, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.014439703896641731, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.013426905497908592, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05225534737110138, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05225534737110138, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.20.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.19163985550403595, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.18086539208889008, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.1772618144750595, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.16131292283535004, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.09061132371425629, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.08681251853704453, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10020676255226135, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09287451207637787, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.09137959778308868, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.08139237761497498, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07772212475538254, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05098940059542656, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04444565251469612, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.043419983237981796, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04317118972539902, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.025456486269831657, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.022350305691361427, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.022094851359725, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02048560604453087, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02033083513379097, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01329858135432005, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.013506721705198288, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.012965069152414799, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009007272310554981, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05098940059542656, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05098940059542656, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.20.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.23590771853923798, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.22270269691944122, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.2183779776096344, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.1988334357738495, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11142033338546753, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10679684579372406, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12308630347251892, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11412108689546585, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11236079037189484, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10014018416404724, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.0956864133477211, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.062515027821064, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05447191372513771, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05324029549956322, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05294356867671013, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03118317760527134, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.027047179639339447, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.026730511337518692, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02470950223505497, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02452080138027668, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01604663021862507, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01576600782573223, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.015635011717677116, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009744023904204369, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05447191372513771, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05447191372513771, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.20.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.23436783254146576, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.20726993680000305, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.19606371223926544, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.17520590126514435, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.10701203346252441, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.09692549705505371, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.1286272257566452, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.11786571145057678, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.1113593652844429, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.0924076959490776, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.08695124834775925, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06539037078619003, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.05666380375623703, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.051702938973903656, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05050637573003769, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03310725465416908, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.027386663481593132, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.027106313034892082, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.024559248238801956, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.023673120886087418, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.018338697031140327, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.018193142488598824, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.016843624413013458, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.012953891418874264, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.051702938973903656, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.051702938973903656, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.21.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.13189035654067993, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.1242438480257988, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.12118007242679596, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.11015181243419647, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.06206908077001572, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.059178367257118225, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.07147566229104996, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.06458587199449539, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.06262264400720596, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.055772483348846436, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0541280098259449, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03643978014588356, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.03098875842988491, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.02985946461558342, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.029598016291856766, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.018327273428440094, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.015340745449066162, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.015108730643987656, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.014057431370019913, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.013891678303480148, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.009634619578719139, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.009241515770554543, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.009286582469940186, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005997848697006702, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.055772483348846436, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0541280098259449, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } } }, { "key": "model.layers.21.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11253061890602112, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.10576672852039337, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.10282167792320251, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.09326667338609695, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.052803803235292435, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.05010507255792618, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.060033850371837616, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05548953264951706, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05327237769961357, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.047306258231401443, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04522722214460373, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.030397867783904076, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.026470793411135674, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.025282995775341988, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.02500210702419281, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.015202480368316174, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.012972027063369751, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.012731408700346947, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.011854098178446293, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01166684739291668, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007940065115690231, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007876653224229813, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007543481886386871, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005086150486022234, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05548953264951706, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05548953264951706, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.21.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.261903315782547, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.24707607924938202, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.2419055998325348, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.22008660435676575, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.1236434057354927, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.11834736913442612, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.13770456612110138, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.12711651623249054, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.1247112825512886, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.11097096651792526, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.10619127750396729, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06988205015659332, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.06065396964550018, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05907595530152321, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05870475992560387, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.0348757840692997, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.029974181205034256, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.029599374160170555, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.027352964505553246, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.027107521891593933, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01791357435286045, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.017470693215727806, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.01731695793569088, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.010720979422330856, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.0348757840692997, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.0348757840692997, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.21.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.20765376091003418, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.18094220757484436, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.17001014947891235, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.14534631371498108, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.09645006060600281, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.08566100150346756, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11460978537797928, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.10514989495277405, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.0991160050034523, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.07733805477619171, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.07250460982322693, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.059038255363702774, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.051458343863487244, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.04753463715314865, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.046554241329431534, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.029985804110765457, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.026401570066809654, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.025547675788402557, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02296093851327896, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.022369757294654846, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.017041457816958427, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.01889895461499691, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.01583831012248993, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.015116025693714619, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.051458343863487244, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.051458343863487244, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.21.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.19986137747764587, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.18862442672252655, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.18493035435676575, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.16832804679870605, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.0945638045668602, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.09062931686639786, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10461486130952835, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09691014885902405, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.09537114948034286, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.08496975898742676, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.08119820058345795, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05324811860918999, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.046391744166612625, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.0453234426677227, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04507011920213699, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.026595668867230415, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02336707152426243, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.023103710263967514, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.021425997838377953, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02126706764101982, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.013935129158198833, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.014160528779029846, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01358514092862606, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.00950260367244482, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05324811860918999, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05324811860918999, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.21.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.24063590168952942, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.2272176444530487, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.22283944487571716, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.20278143882751465, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11371743679046631, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10902249068021774, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12567946314811707, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11645779013633728, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.1146707683801651, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10221944749355316, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09763570129871368, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.0638810247182846, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05559808760881424, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05435105040669441, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05404900759458542, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03184347227215767, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.027591217309236526, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.027272511273622513, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02520718425512314, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02501513622701168, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01640884205698967, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01607891358435154, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.015987521037459373, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009924628771841526, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05559808760881424, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05404900759458542, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.21.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.24121218919754028, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.21409177780151367, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.20324786007404327, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.1814775913953781, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.1103367879986763, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10031205415725708, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.13094578683376312, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12067366391420364, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11459866911172867, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09535209089517593, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.08954010903835297, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06659344583749771, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.057854101061820984, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05312088876962662, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.051971305161714554, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03340702876448631, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.02774079516530037, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.027493929490447044, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.024782568216323853, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.02391503006219864, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.017846567556262016, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.01789182424545288, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.016353584825992584, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.012207618914544582, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05312088876962662, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05312088876962662, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.22.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.1308702677488327, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.12322817742824554, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.1201099082827568, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.10917460173368454, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.061562445014715195, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.0586450956761837, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.0708051323890686, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.0641799122095108, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.062130920588970184, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05523481220006943, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.053452637046575546, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.035992562770843506, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.03074152022600174, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.02957887202501297, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.029302898794412613, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.018116092309355736, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.015253237448632717, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.015010742470622063, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.013970205560326576, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.013796398416161537, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.009593408554792404, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.009271099232137203, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.009229009039700031, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.006097664125263691, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05523481220006943, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05523481220006943, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.22.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11298790574073792, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.10623736679553986, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.1032816544175148, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.09375951439142227, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05310096591711044, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.05039208009839058, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06031421571969986, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.055757127702236176, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05359272286295891, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.047535043209791183, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04545976594090462, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03057616762816906, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.026602450758218765, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.025405369699001312, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.025118181481957436, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.015277794562280178, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.012999771162867546, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.012765910476446152, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.011861935257911682, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.011678625829517841, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.007940229028463364, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007843352854251862, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007547175511717796, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005002976395189762, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.055757127702236176, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.055757127702236176, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.22.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.2558346390724182, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.2412317395210266, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.23619218170642853, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.2148790955543518, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.12078085541725159, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.11549890041351318, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.13451004028320312, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.12418103218078613, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.1218704879283905, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.10833409428596497, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.10356030613183975, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06831011921167374, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.0592619962990284, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05768894404172897, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05730506032705307, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03406046703457832, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.029238680377602577, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.028857555240392685, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02664264105260372, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.026406463235616684, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.017463289201259613, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.016999538987874985, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.01686554215848446, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.010329910553991795, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03406046703457832, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03406046703457832, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.22.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.20453417301177979, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.17652222514152527, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.1658717691898346, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.1391828954219818, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.09191088378429413, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.0796973779797554, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.10993953049182892, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.10036877542734146, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.09510983526706696, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.07258443534374237, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.06851871311664581, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05648098140954971, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04938296228647232, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.04558711126446724, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.04466882348060608, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.029146285727620125, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.025846371427178383, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.025014858692884445, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02248534746468067, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.02194349095225334, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.0173436738550663, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.019007842987775803, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.01625676266849041, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.015629731118679047, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04938296228647232, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04938296228647232, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.22.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.20416580140590668, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.1926327645778656, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.18880799412727356, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.17191638052463531, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.09656139463186264, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.09250795096158981, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10688724368810654, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09898065030574799, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.09740372002124786, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.08671775460243225, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.08288512378931046, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05438847839832306, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.047366514801979065, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.046258024871349335, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.046010907739400864, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.027170643210411072, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.023811303079128265, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02353883907198906, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.021806390956044197, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.021642642095685005, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.014231803826987743, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.014366557821631432, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.013869830407202244, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009550370275974274, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05438847839832306, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05438847839832306, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.22.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.24344757199287415, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.2297212928533554, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.22522728145122528, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.20503798127174377, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11503471434116364, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.1102415919303894, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12711882591247559, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11780798435211182, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11600243300199509, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10334617644548416, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09871062636375427, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06459168344736099, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05624864250421524, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05498179420828819, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.054680272936820984, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.032229673117399216, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.027927087619900703, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.027596773579716682, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02549736574292183, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.025303563103079796, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.016617365181446075, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.016284000128507614, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.016195164993405342, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010063494555652142, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05498179420828819, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.054680272936820984, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.22.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.24110737442970276, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.2140939086675644, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.20322883129119873, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.1816653162240982, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11019710451364517, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.1002369076013565, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.13045021891593933, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12053056061267853, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11442498862743378, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.0952753871679306, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.08948920667171478, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06637133657932281, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.057597558945417404, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.052904557436704636, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05174733325839043, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.033198289573192596, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.027326079085469246, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.02707737870514393, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.024324925616383553, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.02345210127532482, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.017472486943006516, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.01723596453666687, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.015941372141242027, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.01124655269086361, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.052904557436704636, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.052904557436704636, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.23.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.14300541579723358, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.1347164362668991, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.13157781958580017, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.11959321796894073, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.06742677837610245, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.06429222971200943, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.07706887274980545, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.06999016553163528, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.06801740825176239, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.060510050505399704, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.05860625207424164, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03927180543541908, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.033522818237543106, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.03239015117287636, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.032115042209625244, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.019768767058849335, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.016619402915239334, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.016370728611946106, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.01520607341080904, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.015035922639071941, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.010338438674807549, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.009931999258697033, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.009975946508347988, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0063564106822013855, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03927180543541908, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03927180543541908, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.23.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.12394481152296066, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.11665883660316467, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11368909478187561, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.10321568697690964, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.058369606733322144, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.05554945394396782, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06587744504213333, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.06082286685705185, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05888955295085907, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05228541046380997, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.05001688003540039, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.033419907093048096, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.029048366472125053, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.027937930077314377, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.02767045423388481, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.01668931543827057, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.014269701205193996, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.014033595100045204, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.013014818541705608, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.012848839163780212, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.008662324398756027, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.008530938997864723, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.008287268690764904, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.00540607376024127, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05228541046380997, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05228541046380997, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.23.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.25993645191192627, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.2449338734149933, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.2399405539035797, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.21827204525470734, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.12278774380683899, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.11749269813299179, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.13634321093559265, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.12608815729618073, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.12385006994009018, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.11012716591358185, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.10523828864097595, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06922565400600433, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.06017725169658661, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05862746015191078, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.058281492441892624, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03454384207725525, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.029733553528785706, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.029354874044656754, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.027098460122942924, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.02686755731701851, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01771269179880619, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.017271989956498146, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.017145588994026184, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.010520373471081257, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03454384207725525, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03454384207725525, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.23.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.23128919303417206, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.20083512365818024, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.18972471356391907, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.15674623847007751, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10752445459365845, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09603380411863327, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.12670813500881195, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11610922962427139, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.1102980226278305, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.08505843579769135, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.07793879508972168, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06538365036249161, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05662380903959274, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.052703794091939926, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05173761397600174, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03307012841105461, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02902202121913433, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.028236567974090576, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02493816427886486, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.024332983419299126, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01863379217684269, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.020411979407072067, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.017423417419195175, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.016088509932160378, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.052703794091939926, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.052703794091939926, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.23.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2149004340171814, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.20268096029758453, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.19866256415843964, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.18072736263275146, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.10169267654418945, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.09742513298988342, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11261537671089172, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.10427721589803696, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.10260184854269028, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09128057956695557, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.08715379238128662, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05727846547961235, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.049869343638420105, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04870525747537613, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.048436809331178665, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.028600389137864113, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.024983488023281097, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02469528838992119, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02285112626850605, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.022677460685372353, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.014928535558283329, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01496635191142559, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.014541917480528355, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009794418700039387, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.049869343638420105, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.049869343638420105, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.23.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.25044485926628113, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.23619809746742249, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.23166489601135254, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.21075519919395447, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11847111582756042, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.11349357664585114, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.13101451098918915, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.12138961255550385, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.1194869801402092, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10633746534585953, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.10161568224430084, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06665405631065369, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.057973578572273254, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.056654538959264755, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05634798854589462, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.033267829567193985, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02883436158299446, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.028498921543359756, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.026318728923797607, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.026122065261006355, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01723092421889305, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01689930260181427, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.016790417954325676, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010572131723165512, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.033267829567193985, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.033267829567193985, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.23.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.2470356822013855, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.2205185890197754, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.20922811329364777, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.18693362176418304, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11316970735788345, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10317009687423706, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.13518445193767548, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12425489723682404, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11731353402137756, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09827831387519836, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09229253977537155, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06853914260864258, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.05946818366646767, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05445189028978348, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.053198423236608505, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.034514736384153366, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.02834455482661724, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.028046168386936188, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.02537633292376995, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.024463575333356857, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.0184644665569067, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.018189318478107452, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.016842501237988472, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.012239824049174786, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05445189028978348, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05445189028978348, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.24.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.140705868601799, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.1323518455028534, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.12899036705493927, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.11713144183158875, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.06633664667606354, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.06314628571271896, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.07622062414884567, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.06910473108291626, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.0669841542840004, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05939541757106781, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.05746594816446304, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03878997266292572, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.03313213214278221, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.03188861161470413, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.03159864991903305, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.019508134573698044, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.016420328989624977, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.016159944236278534, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.015010584145784378, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01482543908059597, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.010285228490829468, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.009956582449376583, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.009896349161863327, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.006511008832603693, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03878997266292572, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03878997266292572, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.24.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11780524253845215, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.11066719889640808, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.1074923723936081, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.09739948064088821, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05540752038359642, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.05251430347561836, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06315751373767853, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05828304961323738, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05593585968017578, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.04951333627104759, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04731175675988197, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03199557587504387, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.027819432318210602, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.026534464210271835, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.02622465416789055, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.015994925051927567, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.013588249683380127, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.01333386916667223, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.012386070564389229, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01218663714826107, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.008331798948347569, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.008219985291361809, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007897033356130123, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005256431642919779, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05540752038359642, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05540752038359642, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.24.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.2638879418373108, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.248426154255867, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.24322305619716644, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.22074070572853088, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.1247963011264801, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.1192190945148468, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.1390109807252884, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.1283544898033142, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.12589047849178314, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.11166999489068985, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.10664783418178558, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.07063240557909012, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.061307456344366074, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.059640880674123764, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.059244945645332336, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03521491214632988, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.030255792662501335, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.029854390770196915, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.027525566518306732, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.027271347120404243, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.018082251772284508, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.017642371356487274, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.017470603808760643, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.010785437189042568, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03521491214632988, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03521491214632988, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.24.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.1953389197587967, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.1734912246465683, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.16613806784152985, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.13552062213420868, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.09063125401735306, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.08307541161775589, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.10342860966920853, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.09487856179475784, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.09219109266996384, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.07242228090763092, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.06586612015962601, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.053056392818689346, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04681256413459778, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.04492460563778877, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.0444694347679615, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.027097027748823166, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02518109418451786, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.024740606546401978, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.021887052804231644, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.021605880931019783, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01563250832259655, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.017957651987671852, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.015088235028088093, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.014725912362337112, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.053056392818689346, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.053056392818689346, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.24.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2172894924879074, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.20474378764629364, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.2006513625383377, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.182536318898201, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.10290338844060898, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.09849581122398376, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.1139189824461937, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.1055021733045578, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.10377683490514755, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09221406280994415, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.08807449787855148, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05796757712960243, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05047766864299774, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.049289532005786896, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04900715872645378, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.028962206095457077, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02530789189040661, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.025011209771037102, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02312256395816803, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.022946646437048912, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.015128456056118011, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.015177265740931034, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01473461277782917, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009966996498405933, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05047766864299774, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05047766864299774, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.24.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2540113925933838, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.23946908116340637, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.23474539816379547, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.21347418427467346, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.12026273459196091, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.11514388769865036, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.13296255469322205, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.12321378290653229, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.12126189470291138, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10778343677520752, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.10286581516265869, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06757785379886627, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.058821409940719604, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.0574699342250824, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.057151418179273605, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03371727466583252, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.029185475781559944, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02883942238986492, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02659253031015396, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.026390956714749336, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.017383575439453125, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01699763350188732, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.016933955252170563, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010481434874236584, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03371727466583252, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03371727466583252, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.24.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.2541363835334778, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.22718188166618347, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.21641325950622559, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.19315838813781738, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11661815643310547, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10669258236885071, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.13747048377990723, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12686923146247864, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12071214616298676, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10116695612668991, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09485932439565659, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.06981314718723297, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.06064160540699959, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.055983707308769226, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05484231561422348, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03502523899078369, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.028926685452461243, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.02867327816784382, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.02582540363073349, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.024957453832030296, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.018586868420243263, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.01820095255970955, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.01710210181772709, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.011925329454243183, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05484231561422348, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05484231561422348, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.25.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.15262410044670105, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.14351466298103333, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.14010877907276154, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.12726913392543793, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.07213567942380905, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.06880693137645721, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.08228421956300735, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.0748036652803421, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.07281918823719025, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.06465406715869904, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0624811016023159, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.04196249693632126, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.03586465120315552, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.03466937318444252, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.03438710793852806, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.02103016898036003, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.01774660497903824, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.017488915473222733, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.01620652712881565, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01603061333298683, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.010982934385538101, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.010603168979287148, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.010587552562355995, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.006779585033655167, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.04196249693632126, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.04196249693632126, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.25.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.1324664056301117, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.12455034255981445, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.12148912250995636, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.1101895347237587, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.06253352016210556, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.05949399247765541, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.07054928690195084, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.06507173925638199, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.06309974193572998, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05590054392814636, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0534297376871109, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.035815928131341934, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.031095270067453384, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.029949430376291275, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.02967151626944542, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.01789374276995659, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.015302423387765884, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.015049896202981472, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.01393555011600256, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.013759125955402851, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.009297927841544151, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.009144064038991928, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.00890494603663683, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005808712914586067, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0534297376871109, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, "best_option": { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0534297376871109, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } } }, { "key": "model.layers.25.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.26622751355171204, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.25064048171043396, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.24538542330265045, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.22287191450595856, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.1260506510734558, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.12044212222099304, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.14002391695976257, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.12950453162193298, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.1271764636039734, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.11279453337192535, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.1076071709394455, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.07110863924026489, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.06179831922054291, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.06023082509636879, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.0598541684448719, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03547130525112152, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.03053511492908001, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.030135994777083397, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.027764607220888138, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.027518505230545998, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01819603703916073, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.017730534076690674, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.01761651784181595, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.010788029059767723, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03547130525112152, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03547130525112152, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.25.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.17659273743629456, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.15381108224391937, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.1469661146402359, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.12339131534099579, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.0829804390668869, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.07508585602045059, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.09384608268737793, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.08610698580741882, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.0843789279460907, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.06504599004983902, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.05941269174218178, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.048208050429821014, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.041535358875989914, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.0401446595788002, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.039810992777347565, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.024311671033501625, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.021199755370616913, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.020762067288160324, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.017722908407449722, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.017498407512903214, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.013491123914718628, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.013647996820509434, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.013070196844637394, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.009955653920769691, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.048208050429821014, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.048208050429821014, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.25.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.21745766699314117, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.20491155982017517, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.2007986605167389, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.18243998289108276, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.10318083316087723, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.09873530268669128, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11420965194702148, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.10580181330442429, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.10408852249383926, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09240476787090302, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.0881691724061966, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05826719477772713, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0506737045943737, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04946853965520859, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.049178626388311386, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.029099661856889725, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02543475292623043, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02512979507446289, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.023234844207763672, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.023054203018546104, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.015247536823153496, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01531121414154768, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01484950352460146, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010125952772796154, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0506737045943737, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0506737045943737, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.25.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2555176317691803, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.24087877571582794, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.2360382378101349, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.2145046442747116, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.12103641033172607, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.11586093157529831, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.1339658796787262, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.1240973025560379, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.12212088704109192, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10843075811862946, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.10347147285938263, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06812767684459686, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05930614843964577, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05792738497257233, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05760182440280914, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03400009125471115, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.029468361288309097, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.029113005846738815, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.026845546439290047, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.026637455448508263, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.017583008855581284, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.017252063378691673, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01712534762918949, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010758671909570694, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03400009125471115, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03400009125471115, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.25.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.252880722284317, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.2265290766954422, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.21564172208309174, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.1925303041934967, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11628405004739761, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10634353011846542, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.13788577914237976, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.1269301176071167, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12031421810388565, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.1011466532945633, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09478671848773956, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.07013829052448273, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.060854505747556686, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05592896416783333, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05472717806696892, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03540275990962982, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.02906825579702854, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.028759129345417023, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.02604028955101967, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.025128763169050217, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.01905767060816288, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.01854470930993557, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.01746956817805767, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.01239550206810236, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05472717806696892, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05472717806696892, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.26.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.14452672004699707, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.13585714995861053, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.13239768147468567, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.1201387271285057, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.06845054030418396, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.06511163711547852, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.07882272452116013, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.07124143093824387, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.06909491121768951, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.06121459975838661, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.05941333249211311, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.04029582440853119, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.03421945124864578, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.03296615183353424, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.03265265002846718, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.020255587995052338, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.016923444345593452, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.01666027121245861, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.015459122136235237, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01526228804141283, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.010642902925610542, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.010212714783847332, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.010234716348350048, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.006635936442762613, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.04029582440853119, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.04029582440853119, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.26.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.12351682782173157, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.11592871695756912, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11271816492080688, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.10209384560585022, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05828811228275299, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.05526662990450859, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06621098518371582, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.06112572178244591, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05884062498807907, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.051968663930892944, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04965226352214813, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03359244018793106, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.02919614501297474, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.02792910858988762, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.027620753273367882, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.016802232712507248, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.014299298636615276, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.014040766283869743, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.013002512976527214, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01280676294118166, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.008743451908230782, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.008618882857263088, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.008311830461025238, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0055073583498597145, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.051968663930892944, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.051968663930892944, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.26.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.2712506949901581, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.25514042377471924, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.24967044591903687, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.22651679813861847, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.1286047399044037, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.12279170006513596, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.1430523693561554, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.13220356404781342, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.1297347992658615, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.11486953496932983, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.10947592556476593, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.07275502383708954, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.06316248327493668, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.061513327062129974, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.06111610308289528, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.036277517676353455, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.03129778802394867, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.030894333496689796, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02845091000199318, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.02819201909005642, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.018690887838602066, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.018380451947450638, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.018076010048389435, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.011467917822301388, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.036277517676353455, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.036277517676353455, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.26.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.17746339738368988, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.1466691493988037, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.13579639792442322, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.11768276989459991, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.08130185306072235, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.06873133033514023, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.09730909019708633, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.08909489214420319, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.08418645709753036, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.06365420669317245, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.05960472673177719, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05003005638718605, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04375143721699715, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.04024425148963928, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.039369020611047745, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.025460993871092796, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.022722044959664345, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.02185634896159172, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.019634943455457687, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.019106969237327576, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.014633896760642529, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.016641540452837944, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.013620277866721153, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.013571696355938911, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05003005638718605, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.05003005638718605, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.26.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.21033689379692078, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.19809553027153015, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.1940799504518509, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.17622695863246918, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.0998624786734581, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.09548890590667725, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11048716306686401, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.10243675857782364, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.10072393715381622, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.08933552354574203, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.08516944944858551, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.056302912533283234, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0490628182888031, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.047877728939056396, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04760704189538956, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.028136610984802246, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02464480511844158, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.024349449202418327, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.0225027147680521, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02232225611805916, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.014716394245624542, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.014894775114953518, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01432567834854126, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.00990319810807705, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0490628182888031, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0490628182888031, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.26.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.2511913776397705, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.23669186234474182, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.23188622295856476, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.21066223084926605, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11909718066453934, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.11394430696964264, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.13186763226985931, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.12212034314870834, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.12012806534767151, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10656309127807617, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.10161078721284866, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06706134229898453, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05834176763892174, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05697138234972954, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05664971098303795, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.0334642231464386, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.028965242207050323, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.028611307963728905, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.026367686688899994, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.026152122765779495, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.017300812527537346, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.016933482140302658, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.016837608069181442, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010508554987609386, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.0334642231464386, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.0334642231464386, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.26.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.259419322013855, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.23244504630565643, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.2215418964624405, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.1977955549955368, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11929469555616379, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10932204127311707, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.1410561203956604, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.1297542005777359, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12341699004173279, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10368351638317108, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09718535840511322, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.0716170221567154, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.062082741409540176, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05736657604575157, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.0562143512070179, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.035907577723264694, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.029766758903861046, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.029516205191612244, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.02664240449666977, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.02576860971748829, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.019095629453659058, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.018878718838095665, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.01756914332509041, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.012612799182534218, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.035907577723264694, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.035907577723264694, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.27.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.1503482162952423, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.14141198992729187, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.13809603452682495, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.1255243569612503, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.0713210254907608, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.06804714351892471, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.08195190876722336, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.07382010668516159, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.07196202874183655, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.06386135518550873, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.06202295795083046, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.04191228002309799, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.03542828932404518, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.034316834062337875, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.034063197672367096, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.021044837310910225, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.017540108412504196, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.017290517687797546, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.016012949869036674, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01584441587328911, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.010946464724838734, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.01040306594222784, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.010580448433756828, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.006570441648364067, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.04191228002309799, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.04191228002309799, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.27.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.13000521063804626, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.12223958224058151, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11925756931304932, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.10809911787509918, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.06150400638580322, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.05856495350599289, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06908194720745087, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.06381735950708389, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.062096256762742996, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05488702654838562, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.05237786844372749, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03508133068680763, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.030487066134810448, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.02944747917354107, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.02919997274875641, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.0175189059227705, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.015010863542556763, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.014781446196138859, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.01364900078624487, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.013490150682628155, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.0090719573199749, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.008905930444598198, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.008707568049430847, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005599315743893385, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05488702654838562, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05488702654838562, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.27.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.2675924599170685, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.2518564760684967, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.24663835763931274, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.22390179336071014, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.12688510119915009, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.12129063159227371, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.14076314866542816, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.13029077649116516, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.12802593410015106, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.11339695006608963, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.1080942302942276, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.07155348360538483, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.062211114913225174, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.060666024684906006, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.06028302013874054, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03568479046225548, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.030738458037376404, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.030351394787430763, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02791762351989746, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.027683746069669724, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01829427294433117, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.017852913588285446, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.017745718359947205, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0108677102252841, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03568479046225548, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03568479046225548, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.27.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.22498300671577454, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.19419528543949127, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.18247035145759583, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.15276825428009033, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.10466820746660233, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.09208057820796967, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.1238408014178276, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.11357395350933075, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.10798382759094238, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.08275431394577026, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.07679571211338043, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06383596360683441, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.055276911705732346, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.051200464367866516, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05020928755402565, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03224227577447891, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.027912672609090805, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.0269868616014719, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02385968714952469, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.023223428055644035, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.018009796738624573, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.019362444058060646, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.016753120347857475, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.014958227053284645, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.055276911705732346, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.055276911705732346, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.27.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.20343086123466492, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.19138842821121216, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.18747852742671967, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.17019250988960266, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.09662672877311707, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.09235145896673203, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10709221661090851, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09920401871204376, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.09750135987997055, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.08638391643762589, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.08234097063541412, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05459228903055191, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04755692183971405, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04639143496751785, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04611065238714218, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.027310175821185112, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02395210601389408, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.023671163246035576, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.021876851096749306, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.021703016012907028, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.014373204670846462, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.014587434940040112, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.01399331446737051, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.00984619464725256, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05459228903055191, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05459228903055191, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.27.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.24251142144203186, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.22825859487056732, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.2236490547657013, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.20301693677902222, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.11501646786928177, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10995404422283173, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.12737004458904266, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.11797488480806351, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.11603228002786636, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.10283762961626053, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09806112945079803, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.06473304331302643, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.0563664548099041, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05501530319452286, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05470141023397446, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.03231266885995865, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.027961788699030876, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.027615748345851898, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.025431184098124504, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02522582747042179, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01666492037475109, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.016342181712388992, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.016210408881306648, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010126362554728985, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05501530319452286, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.05470141023397446, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.27.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.2597598731517792, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.23178672790527344, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.22073784470558167, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.19704633951187134, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11932776868343353, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10915647447109222, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.1406010389328003, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12988759577274323, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12370869517326355, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10355240106582642, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.0969972088932991, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.07158666849136353, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.06223194673657417, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05743779242038727, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.056267544627189636, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03592411428689957, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.029925335198640823, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.029673557728528976, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.02678057737648487, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.025901801884174347, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.019075358286499977, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.019133897498250008, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.01750371791422367, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.012957953847944736, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03592411428689957, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03592411428689957, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.28.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.1501951366662979, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.14117321372032166, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.13786543905735016, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.12517642974853516, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.07141394168138504, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.06805884093046188, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.0826505646109581, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.07400581985712051, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.07205599546432495, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.06388213485479355, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.06235875189304352, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.042472973465919495, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.03559760004281998, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.03443032130599022, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.03413916006684303, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.021330665796995163, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.017646901309490204, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.017394838854670525, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.016117282211780548, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01594622991979122, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.011127046309411526, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.010524336248636246, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.010753428563475609, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.006718343123793602, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.042472973465919495, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.042472973465919495, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.28.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.13003452122211456, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.12216025590896606, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11903107166290283, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.10786503553390503, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.061511147767305374, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.058476485311985016, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.0692310780286789, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.06394821405410767, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.06210524961352348, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05483323708176613, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.05230026692152023, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.035169150680303574, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.030560405924916267, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.0294654443860054, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.029207760468125343, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.017589259892702103, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.015044128522276878, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.014805706217885017, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.013670330867171288, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.013505438342690468, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.009123843163251877, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.008975373581051826, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.008741109631955624, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.0056842295452952385, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05483323708176613, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05483323708176613, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.28.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.26801595091819763, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.2521716058254242, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.24678592383861542, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.2240181565284729, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.12722311913967133, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.12147247046232224, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.14117617905139923, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.13060800731182098, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.1283339411020279, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.11349321901798248, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.10820849984884262, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.07178804278373718, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.06241274252533913, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.06081666797399521, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.06045260280370712, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03580605611205101, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.03089131787419319, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.030491461977362633, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02805321477353573, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.027810698375105858, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01841035857796669, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.018035151064395905, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.017849789932370186, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.011098124086856842, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03580605611205101, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.03580605611205101, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.28.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.20609943568706512, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.17271748185157776, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.16026316583156586, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.13346615433692932, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.09563420712947845, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.08309106528759003, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.11449524760246277, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.10502898693084717, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.09959885478019714, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.07395123690366745, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.06747201085090637, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.059577494859695435, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05122661218047142, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.04697980731725693, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.04594220221042633, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.030316432937979698, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02572326734662056, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.024857917800545692, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.021746622398495674, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.021085325628519058, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.017293844372034073, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.018077779561281204, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.016011156141757965, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.014034167863428593, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05122661218047142, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.05122661218047142, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.28.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.19829823076725006, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.18654796481132507, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.18271447718143463, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.16587816178798676, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.09423922747373581, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.09006253629922867, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10449114441871643, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09679831564426422, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.09510841965675354, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.0842442736029625, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.08026401698589325, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05324939638376236, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04642120376229286, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04526088759303093, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04499194025993347, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.02661973610520363, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.023406416177749634, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02311735600233078, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.021376408636569977, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.021205272525548935, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.013986349105834961, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.014302566647529602, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.013605236075818539, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009718685410916805, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05324939638376236, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05324939638376236, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.28.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.22347231209278107, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.21031221747398376, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.20604442059993744, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.18704454600811005, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.10607802122831345, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.10137972235679626, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.11753629148006439, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.10881074517965317, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.1070355474948883, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.09478094428777695, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.09035833179950714, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05984243378043175, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05202794820070267, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.05075487121939659, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.0504552498459816, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.02986457571387291, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.025828195735812187, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.025508426129817963, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02348451502621174, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.023292040452361107, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01545234676450491, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.015133839100599289, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.015030207112431526, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.009446918033063412, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05202794820070267, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.05202794820070267, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.28.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.2569671869277954, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.22874073684215546, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.2177584320306778, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.19416989386081696, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.1179685965180397, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.107718825340271, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.13977406919002533, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.1284782886505127, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.12263083457946777, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.10231312364339828, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09577134996652603, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.07143382728099823, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.06187966465950012, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.056998997926712036, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05580848082900047, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03617429360747337, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.0300386231392622, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.029819052666425705, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.026980221271514893, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.026108404621481895, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.019887244328856468, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.01967829465866089, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.01836763322353363, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.01390916295349598, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05580848082900047, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05580848082900047, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.29.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.14447320997714996, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.1358359009027481, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.13245508074760437, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.12020433694124222, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.06879506260156631, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.06550005078315735, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.0801401361823082, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.07140260189771652, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.06940338015556335, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.06152524799108505, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.060317326337099075, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.041177406907081604, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.03438997641205788, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.03318992257118225, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.03290776163339615, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.020714962854981422, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.017024552449584007, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.01676160655915737, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.015546689741313457, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.015369923785328865, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.010793435387313366, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.010186009109020233, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.010412674397230148, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.006505606230348349, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.041177406907081604, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.041177406907081604, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.29.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.12201111763715744, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.11451736092567444, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11142194271087646, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.10081685334444046, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.057640738785266876, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.05466959998011589, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06527573615312576, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.060233935713768005, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05818766728043556, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05133402347564697, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.0489899180829525, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.03314772993326187, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.028794508427381516, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.027621852234005928, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.027342481538653374, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.016573438420891762, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.014138052240014076, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.013890959322452545, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.012842511758208275, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.012660312466323376, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.00861887726932764, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.008509920910000801, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.008210640400648117, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005440602079033852, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05133402347564697, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05133402347564697, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.29.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.2733360528945923, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.25707346200942993, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.2517838776111603, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.22835662961006165, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.129814013838768, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.12396321445703506, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.14420098066329956, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.1332879364490509, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.13092055916786194, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.11586786806583405, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.11044543981552124, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.07326741516590118, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.06367737054824829, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.06206781417131424, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.061687082052230835, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.036568839102983475, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.03147595748305321, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.03107074275612831, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.028569256886839867, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.028322216123342514, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01876893825829029, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.018301527947187424, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.018193401396274567, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.011181943118572235, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.036568839102983475, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.036568839102983475, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.29.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.17009495198726654, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.14912211894989014, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.14089471101760864, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.12281044572591782, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.07878069579601288, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.07103107869625092, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.09363411366939545, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.08606930822134018, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.08159352093935013, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.06534227728843689, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.061100587248802185, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.048314254730939865, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.04211268573999405, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.03887557610869408, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.03807985410094261, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.024592479690909386, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.021673373878002167, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.021142659708857536, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.01926802285015583, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.018777934834361076, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01410156674683094, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.015601026825606823, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.01317725982517004, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.012513265013694763, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.048314254730939865, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.048314254730939865, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.29.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.1886732578277588, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.17738650739192963, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.1736835092306137, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.15753203630447388, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.08986864984035492, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.08579381555318832, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.09996578097343445, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.09234903007745743, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.09071208536624908, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.08023874461650848, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.07662612944841385, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05132288858294487, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04458239674568176, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.04345346614718437, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04319804161787033, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.02573828026652336, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02297041192650795, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.022697703912854195, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02108636498451233, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.020923634991049767, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01401941105723381, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.014735084027051926, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.013662625104188919, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.010841702111065388, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05132288858294487, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.05132288858294487, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.29.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.20060788094997406, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.188735231757164, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.1847991943359375, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.167713463306427, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.09592235833406448, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.09169352799654007, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.10666665434837341, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.0985107570886612, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.09677537530660629, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.0857217013835907, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.08189980685710907, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.0549568310379982, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.04785337671637535, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.046695277094841, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.04642154648900032, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.027682172134518623, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.025230783969163895, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.024950256571173668, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02327834628522396, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.02312835305929184, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.01545379776507616, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.016862379387021065, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.015095075592398643, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.01310043316334486, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.0549568310379982, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.0549568310379982, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.29.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.2507997453212738, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.22222226858139038, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.21109889447689056, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.18807093799114227, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.11505317687988281, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.10469690710306168, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.13703858852386475, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.12554500997066498, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.11985083669424057, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.09942774474620819, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.09325665235519409, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.07014745473861694, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.060675881803035736, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05575506389141083, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05459590628743172, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.03566926345229149, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.02975570410490036, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.029547154903411865, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.026749346405267715, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.02588130161166191, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.01997038722038269, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.019926244392991066, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.018512649461627007, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.014552950859069824, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.05575506389141083, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, "best_option": { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.05459590628743172, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.30.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.14594706892967224, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.13745880126953125, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.13427361845970154, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.12210914492607117, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.06989551335573196, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.06676967442035675, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.08224070817232132, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.07226806879043579, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.07052115350961685, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.06276752054691315, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.062134649604558945, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.04263933002948761, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.034872233867645264, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.03381141647696495, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.0335567444562912, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.021406225860118866, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.01729978807270527, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.017062649130821228, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.01584966480731964, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01569005846977234, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.011147988960146904, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.010285463184118271, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.010793838649988174, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.006538465619087219, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.04263933002948761, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.04263933002948761, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.30.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.1275021731853485, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.11973381042480469, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11684240400791168, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.1058967262506485, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.060309089720249176, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.05742878094315529, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06775161623954773, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.06252273917198181, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.06089211627840996, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05378514155745506, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.05130627751350403, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.034436389803886414, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.029879840090870857, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.028885306790471077, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.02864847518503666, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.017212655395269394, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.014752442948520184, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.014534165151417255, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.01341604720801115, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01325822714716196, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.008937409147620201, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.00879247859120369, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.008578956127166748, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005586662795394659, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05378514155745506, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.05378514155745506, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.30.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.2620069682598114, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.24656641483306885, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.24148426949977875, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.21887868642807007, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.12437903136014938, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.11887617409229279, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.13816362619400024, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.12772385776042938, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.1255367547273636, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.11102019250392914, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.10576476901769638, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.07021176815032959, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.06105383485555649, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.05951700732111931, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05916190892457962, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.0350453220307827, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.030199332162737846, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.02981988526880741, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.0274201687425375, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.027173317968845367, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.018059080466628075, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.017594782635569572, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.01751936785876751, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.010777209885418415, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.0350453220307827, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.0350453220307827, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.30.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.191743865609169, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.15671806037425995, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.14374321699142456, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.12503968179225922, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.08877602219581604, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.07337486743927002, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.10565347224473953, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.09699049592018127, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.09202593564987183, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.06798963993787766, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.06342083215713501, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.054402243345975876, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.0474192351102829, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.04375455528497696, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.04286648705601692, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.027538936585187912, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02436051517724991, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.023403262719511986, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02065596729516983, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.020076528191566467, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01554702129215002, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.017474789172410965, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.014455059543251991, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.013971278443932533, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.054402243345975876, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.054402243345975876, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.30.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.1592331826686859, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.14944905042648315, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.14652937650680542, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.13301484286785126, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.0759345218539238, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.07247549295425415, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.08425768464803696, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.07801511883735657, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.07666859775781631, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.06800149381160736, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.0648951381444931, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.043442223221063614, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.03879326209425926, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.03784410282969475, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.03762964904308319, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.022160246968269348, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.02187679521739483, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.021682413294911385, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.02049766108393669, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.0203853752464056, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.012971136718988419, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.016183961182832718, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.012699753977358341, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.01386313233524561, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.043442223221063614, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.043442223221063614, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.30.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.15974390506744385, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.15004487335681915, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.14687950909137726, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.13338147103786469, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.07601819932460785, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.07268408685922623, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.08427012711763382, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.07809896767139435, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.07676264643669128, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.06809338927268982, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.06502202898263931, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04337390884757042, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.03901873156428337, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.03810220956802368, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.037901297211647034, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.022249026224017143, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.022260412573814392, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.02207547426223755, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.020924804732203484, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.020806048065423965, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.013150089420378208, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.01671028323471546, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.012894032523036003, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.01447177305817604, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04337390884757042, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.04337390884757042, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.30.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.1049116924405098, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.0941167101264, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.08264414966106415, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.07428630441427231, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.04951745271682739, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.04129604995250702, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.07718781381845474, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.06407857686281204, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.05113203078508377, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.043704960495233536, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.04277222976088524, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.03663746267557144, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.03208781033754349, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.02573540061712265, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.02360697090625763, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.020081810653209686, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.01561431959271431, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.014231707900762558, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.014703775756061077, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.013033432886004448, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.012910430319607258, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.0130729665979743, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.010488434694707394, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.010186615400016308, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.05113203078508377, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, "best_option": { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.05113203078508377, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.31.self_attn.q_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.1282641589641571, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.12040415406227112, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11709333956241608, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.10581760108470917, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.06061093509197235, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.057496052235364914, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.07033804804086685, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.0635581761598587, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.061226386576890945, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.053995221853256226, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.05202407389879227, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.035922128707170486, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.03042425960302353, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.029093552380800247, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.028781745582818985, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.018002424389123917, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.0149762611836195, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.01470910757780075, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.01362183690071106, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.01342293806374073, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.009519374929368496, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.009143862873315811, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.008994914591312408, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.006004243157804012, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.053995221853256226, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.053995221853256226, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.31.self_attn.k_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.11133863031864166, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.10455295443534851, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.10165022313594818, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.09187798947095871, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.05258603021502495, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.049858614802360535, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.06001829355955124, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05509420111775398, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.05309109017252922, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.04683037847280502, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.04475512355566025, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.030571673065423965, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.026370037347078323, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.025237148627638817, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.024965187534689903, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.015301867388188839, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.013004591688513756, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.01277538575232029, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.011839423328638077, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.011667846702039242, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.008044774644076824, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.007973119616508484, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.007647249381989241, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.005281067453324795, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05509420111775398, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.05509420111775398, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.31.self_attn.v_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.25890564918518066, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.24323846399784088, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.23801733553409576, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.21531671285629272, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.12289968132972717, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.11726237833499908, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.13751071691513062, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.1263437420129776, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.12402118742465973, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.10947252810001373, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.10421791672706604, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.06999333202838898, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.06047867238521576, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.058832138776779175, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.05844571441411972, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.0349733829498291, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.02995884045958519, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.0295668076723814, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.02718285471200943, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.02692689746618271, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.01807786151766777, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.01763761229813099, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.017428049817681313, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.011053148657083511, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.0349733829498291, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.0349733829498291, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.31.self_attn.o_proj", "numel": 16777216, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1878662109375, "total_bits": 36706304.0, "err": 0.13625667989253998, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3753662109375, "total_bits": 39852032.0, "err": 0.1201801598072052, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6253662109375, "total_bits": 44046336.0, "err": 0.11148351430892944, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7191162109375, "total_bits": 45619200.0, "err": 0.09863819926977158, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2191162109375, "total_bits": 54007808.0, "err": 0.0632852092385292, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6878662109375, "total_bits": 61872128.0, "err": 0.05573798343539238, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 50857472.0, "err": 0.0813639909029007, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 52434944.0, "err": 0.07256967574357986, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1878662109375, "total_bits": 53483520.0, "err": 0.0643613338470459, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.053873054683208466, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.65643310546875, "total_bits": 61344768.0, "err": 0.051718275994062424, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 67634688.0, "err": 0.04201379418373108, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 69212160.0, "err": 0.03649080544710159, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2191162109375, "total_bits": 70785024.0, "err": 0.03242676705121994, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3128662109375, "total_bits": 72357888.0, "err": 0.031392164528369904, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 84411904.0, "err": 0.022515635937452316, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2191162109375, "total_bits": 87562240.0, "err": 0.01956104300916195, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3753662109375, "total_bits": 90183680.0, "err": 0.01892075315117836, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5316162109375, "total_bits": 92805120.0, "err": 0.017969707027077675, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.7191162109375, "total_bits": 95950848.0, "err": 0.017436733469367027, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 101189120.0, "err": 0.014388342387974262, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 102766592.0, "err": 0.015562283806502819, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218841552734375, "total_bits": 104334848.0, "err": 0.013319257646799088, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 136321024.0, "err": 0.013445070944726467, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.053873054683208466, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5316162109375, "total_bits": 59250688.0, "err": 0.053873054683208466, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.31.mlp.gate_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.13440778851509094, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.12632527947425842, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.12356027215719223, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.11200615763664246, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.06376270204782486, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.06084392964839935, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.07116390764713287, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.06561136245727539, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.06435111910104752, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.05688338726758957, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.054370880126953125, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.03639063611626625, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.03156168386340141, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.03071536123752594, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.03051697090268135, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.01821468025445938, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.016052410006523132, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.015854662284255028, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.014684950932860374, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.01456180214881897, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.009709029458463192, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.010070346295833588, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.00942945946007967, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.007129534147679806, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.054370880126953125, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, "best_option": { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.054370880126953125, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } } }, { "key": "model.layers.31.mlp.up_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1876362645348837, "total_bits": 98637824.0, "err": 0.12613776326179504, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.3751362645348837, "total_bits": 107091968.0, "err": 0.11853736639022827, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.6251362645348837, "total_bits": 118364160.0, "err": 0.11590805649757385, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.7188862645348837, "total_bits": 122591232.0, "err": 0.10505536198616028, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.2188862645348837, "total_bits": 145135616.0, "err": 0.06009263917803764, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.6876362645348837, "total_bits": 166270976.0, "err": 0.057371873408555984, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031284066133721, "total_bits": 136676864.0, "err": 0.0670788586139679, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1251362645348837, "total_bits": 140908544.0, "err": 0.06186089292168617, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1876362645348837, "total_bits": 143726592.0, "err": 0.06069101020693779, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.05367552861571312, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.656318132267442, "total_bits": 164858880.0, "err": 0.0513150617480278, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031284066133721, "total_bits": 181765632.0, "err": 0.034549787640571594, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125136264534884, "total_bits": 185997312.00000003, "err": 0.030049340799450874, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.218886264534884, "total_bits": 190224384.00000003, "err": 0.029259582981467247, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.312636264534884, "total_bits": 194451456.00000003, "err": 0.029067862778902054, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031284066133721, "total_bits": 226854400.0, "err": 0.017373444512486458, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.218886264534884, "total_bits": 235313152.00000003, "err": 0.01576969586312771, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375136264534884, "total_bits": 242358272.00000003, "err": 0.015592469833791256, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.531386264534884, "total_bits": 249403392.00000003, "err": 0.014544278383255005, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.718886264534884, "total_bits": 257857536.00000003, "err": 0.014438258484005928, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031284066133721, "total_bits": 271943168.0, "err": 0.009614422917366028, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125136264534884, "total_bits": 276174848.0, "err": 0.010525500401854515, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.218784066133721, "total_bits": 280397312.0, "err": 0.009364845231175423, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125136264534884, "total_bits": 366352384.0, "err": 0.00813465565443039, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.05367552861571312, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, "best_option": { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.5313862645348837, "total_bits": 159225856.0, "err": 0.05367552861571312, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.31.mlp.down_proj", "numel": 45088768, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.171877838844477, "total_bits": 97927296.0, "err": 0.09421050548553467, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.381180164425872, "total_bits": 107364480.0, "err": 0.08346568048000336, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.636994117914244, "total_bits": 118898816.0, "err": 0.07567330449819565, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.730017373728198, "total_bits": 123093120.00000001, "err": 0.06760416179895401, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.230017373728198, "total_bits": 145637504.0, "err": 0.04408443719148636, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7183894667514537, "total_bits": 167657600.0, "err": 0.03808237984776497, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.031341552734375, "total_bits": 136679456.0, "err": 0.059754885733127594, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.05320250988006592, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.171877838844477, "total_bits": 143016064.0, "err": 0.045577242970466614, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.520715048146802, "total_bits": 158744704.0, "err": 0.03828899934887886, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.667334268259448, "total_bits": 165355584.0, "err": 0.03685526177287102, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031341552734375, "total_bits": 181768224.0, "err": 0.031180215999484062, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1253662109375, "total_bits": 186007680.0, "err": 0.026662297546863556, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.230017373728198, "total_bits": 190726272.0, "err": 0.022472327575087547, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.334668536518896, "total_bits": 195444864.00000003, "err": 0.021421462297439575, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031341552734375, "total_bits": 226856992.0, "err": 0.016789203509688377, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.230017373728198, "total_bits": 235815040.0, "err": 0.013708803802728653, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.311412722565407, "total_bits": 239485056.0, "err": 0.013326524756848812, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.520715048146802, "total_bits": 248922240.0, "err": 0.012794583104550838, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.741645280704942, "total_bits": 258883712.0, "err": 0.012214189395308495, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031341552734375, "total_bits": 271945760.0, "err": 0.010880696587264538, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1253662109375, "total_bits": 276185216.0, "err": 0.011178802698850632, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.240643878315771, "total_bits": 281382944.0, "err": 0.009827500209212303, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.1253662109375, "total_bits": 366362752.0, "err": 0.009611038491129875, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.05320250988006592, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.1253662109375, "total_bits": 140918912.0, "err": 0.05320250988006592, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } } ], "last_module_idx": 66, "base_perplexity": 6.422811091992605 }