update ds_config
Browse files- ds_config.json +19 -0
ds_config.json
CHANGED
|
@@ -10,6 +10,15 @@
|
|
| 10 |
"hysteresis": 2,
|
| 11 |
"min_loss_scale": 1
|
| 12 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
"scheduler": {
|
| 14 |
"type": "OneCycle",
|
| 15 |
"params": {
|
|
@@ -19,12 +28,22 @@
|
|
| 19 |
},
|
| 20 |
"zero_optimization": {
|
| 21 |
"stage": 2,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 22 |
"overlap_comm": true,
|
| 23 |
"allgather_partitions": true,
|
| 24 |
"allgather_bucket_size": 5e8,
|
| 25 |
"contiguous_gradients": true,
|
| 26 |
"reduce_bucket_size": "auto",
|
| 27 |
"reduce_scatter": true,
|
|
|
|
|
|
|
| 28 |
"stage3_gather_16bit_weights_on_model_save": true
|
| 29 |
},
|
| 30 |
"gradient_accumulation_steps": "auto",
|
|
|
|
| 10 |
"hysteresis": 2,
|
| 11 |
"min_loss_scale": 1
|
| 12 |
},
|
| 13 |
+
"optimizer": {
|
| 14 |
+
"type": "Adam",
|
| 15 |
+
"params": {
|
| 16 |
+
"lr": "auto",
|
| 17 |
+
"betas": "auto",
|
| 18 |
+
"eps": "auto",
|
| 19 |
+
"weight_decay": "auto"
|
| 20 |
+
}
|
| 21 |
+
},
|
| 22 |
"scheduler": {
|
| 23 |
"type": "OneCycle",
|
| 24 |
"params": {
|
|
|
|
| 28 |
},
|
| 29 |
"zero_optimization": {
|
| 30 |
"stage": 2,
|
| 31 |
+
"offload_optimizer": {
|
| 32 |
+
"device": "cpu",
|
| 33 |
+
"pin_memory": true
|
| 34 |
+
},
|
| 35 |
+
"offload_param": {
|
| 36 |
+
"device": "cpu",
|
| 37 |
+
"pin_memory": true
|
| 38 |
+
},
|
| 39 |
"overlap_comm": true,
|
| 40 |
"allgather_partitions": true,
|
| 41 |
"allgather_bucket_size": 5e8,
|
| 42 |
"contiguous_gradients": true,
|
| 43 |
"reduce_bucket_size": "auto",
|
| 44 |
"reduce_scatter": true,
|
| 45 |
+
"stage3_max_live_parameters": 0,
|
| 46 |
+
"stage3_max_reuse_distance": 0,
|
| 47 |
"stage3_gather_16bit_weights_on_model_save": true
|
| 48 |
},
|
| 49 |
"gradient_accumulation_steps": "auto",
|