Create new file
Browse files- 4-A100-config.json +66 -0
4-A100-config.json
ADDED
@@ -0,0 +1,66 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
|
3 |
+
|
4 |
+
"pipe-parallel-size": 1,
|
5 |
+
"model-parallel-size": 1,
|
6 |
+
|
7 |
+
|
8 |
+
"scaled-upper-triang-masked-softmax-fusion": false,
|
9 |
+
"bias-gelu-fusion": false,
|
10 |
+
|
11 |
+
"fp16": {
|
12 |
+
"enabled": "auto",
|
13 |
+
"auto_cast": false,
|
14 |
+
"loss_scale": 0,
|
15 |
+
"initial_scale_power": 32,
|
16 |
+
"loss_scale_window": 1000,
|
17 |
+
"hysteresis": 2,
|
18 |
+
"min_loss_scale": 1
|
19 |
+
},
|
20 |
+
|
21 |
+
"tensorboard": {
|
22 |
+
"enabled": true,
|
23 |
+
"output_path": "output/ds_logs/",
|
24 |
+
"job_name": "train_bert"
|
25 |
+
},
|
26 |
+
|
27 |
+
"optimizer": {
|
28 |
+
"type": "AdamW",
|
29 |
+
"params": {
|
30 |
+
"lr": "auto",
|
31 |
+
"betas": "auto",
|
32 |
+
"eps": "auto",
|
33 |
+
"weight_decay": "auto"
|
34 |
+
}
|
35 |
+
},
|
36 |
+
|
37 |
+
"scheduler": {
|
38 |
+
"type": "WarmupLR",
|
39 |
+
"params": {
|
40 |
+
"warmup_min_lr": "auto",
|
41 |
+
"warmup_max_lr": "auto",
|
42 |
+
"warmup_num_steps": "auto"
|
43 |
+
}
|
44 |
+
},
|
45 |
+
|
46 |
+
"zero_optimization": {
|
47 |
+
"stage": 2,
|
48 |
+
"offload_optimizer": {
|
49 |
+
"device": "cpu",
|
50 |
+
"pin_memory": true
|
51 |
+
},
|
52 |
+
"allgather_partitions": true,
|
53 |
+
"allgather_bucket_size": 2e8,
|
54 |
+
"overlap_comm": true,
|
55 |
+
"reduce_scatter": true,
|
56 |
+
"reduce_bucket_size": 2e8,
|
57 |
+
"contiguous_gradients": true
|
58 |
+
},
|
59 |
+
|
60 |
+
"gradient_accumulation_steps": "auto",
|
61 |
+
"gradient_clipping": "auto",
|
62 |
+
"steps_per_print": 2000,
|
63 |
+
"train_batch_size": "auto",
|
64 |
+
"train_micro_batch_size_per_gpu": "auto",
|
65 |
+
"wall_clock_breakdown": false
|
66 |
+
}
|