Update ds_config.json
Browse files- ds_config.json +54 -0
ds_config.json
CHANGED
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"zero_optimization": {
|
3 |
+
"stage": 3,
|
4 |
+
"overlap_comm": true,
|
5 |
+
"contiguous_gradients": true,
|
6 |
+
"sub_group_size": 0,
|
7 |
+
"reduce_bucket_size": "auto",
|
8 |
+
"stage3_prefetch_bucket_size": "auto",
|
9 |
+
"stage3_param_persistence_threshold": "auto",
|
10 |
+
"stage3_max_live_parameters": 0,
|
11 |
+
"stage3_max_reuse_distance": 0,
|
12 |
+
"stage3_gather_16bit_weights_on_model_save": true
|
13 |
+
},
|
14 |
+
"bf16": {
|
15 |
+
"enabled": true,
|
16 |
+
"auto_cast": false,
|
17 |
+
"loss_scale": 0,
|
18 |
+
"initial_scale_power": 32,
|
19 |
+
"loss_scale_window": 1000,
|
20 |
+
"hysteresis": 2,
|
21 |
+
"min_loss_scale": 1
|
22 |
+
},
|
23 |
+
"optimizer": {
|
24 |
+
"type": "AdamW",
|
25 |
+
"params": {
|
26 |
+
"lr": "auto",
|
27 |
+
"betas": [
|
28 |
+
0.9,
|
29 |
+
0.999
|
30 |
+
],
|
31 |
+
"eps": 1e-8,
|
32 |
+
"weight_decay": 0
|
33 |
+
}
|
34 |
+
},
|
35 |
+
"scheduler": {
|
36 |
+
"type": "OneCycle",
|
37 |
+
"params": {
|
38 |
+
"cycle_first_step_size": 1000,
|
39 |
+
"cycle_first_stair_count": 500,
|
40 |
+
"cycle_second_step_size": 1000,
|
41 |
+
"cycle_second_stair_count": 500,
|
42 |
+
"decay_step_size": 1000,
|
43 |
+
"cycle_min_lr": 0,
|
44 |
+
"cycle_max_lr": 1e-5,
|
45 |
+
"decay_lr_rate": 0.001,
|
46 |
+
"cycle_min_mom": 0.85,
|
47 |
+
"cycle_max_mom": 0.99,
|
48 |
+
"decay_mom_rate": 0.0
|
49 |
+
}
|
50 |
+
},
|
51 |
+
"train_batch_size": "auto",
|
52 |
+
"train_micro_batch_size_per_gpu": "auto",
|
53 |
+
"wall_clock_breakdown": false
|
54 |
+
}
|