Upload deepspeed_config.json
Browse files- deepspeed_config.json +28 -0
deepspeed_config.json
ADDED
|
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"train_micro_batch_size_per_gpu": 2,
|
| 3 |
+
"gradient_accumulation_steps": 1,
|
| 4 |
+
"steps_per_print": 100,
|
| 5 |
+
"gradient_clipping": 1.0,
|
| 6 |
+
"fp16": {
|
| 7 |
+
"enabled": true,
|
| 8 |
+
"loss_scale": 0,
|
| 9 |
+
"loss_scale_window": 2000,
|
| 10 |
+
"hysteresis": 2,
|
| 11 |
+
"min_loss_scale": 0.0
|
| 12 |
+
},
|
| 13 |
+
"zero_optimization": {
|
| 14 |
+
"stage": 2,
|
| 15 |
+
"reduce_bucket_size": 50000000,
|
| 16 |
+
"overlap_comm": true
|
| 17 |
+
},
|
| 18 |
+
"sparse_attention": {
|
| 19 |
+
"mode": "fixed",
|
| 20 |
+
"block": 16,
|
| 21 |
+
"different_layout_per_head": true,
|
| 22 |
+
"num_local_blocks": 8,
|
| 23 |
+
"num_global_blocks": 1,
|
| 24 |
+
"attention": "unidirectional",
|
| 25 |
+
"horizontal_global_attention": false,
|
| 26 |
+
"num_different_global_patterns": 8
|
| 27 |
+
}
|
| 28 |
+
}
|