feat: add deepspeed 3 with cpuoffload (#1466)
Browse files* feat: add deepspeed 3 with cpuoffload
* make bf16 explicit, add param only offload variant
---------
Co-authored-by: Wing Lian <wing.lian@gmail.com>
deepspeed_configs/zero3_bf16_cpuoffload_all.json
ADDED
|
@@ -0,0 +1,39 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"zero_optimization": {
|
| 3 |
+
"stage": 3,
|
| 4 |
+
"offload_optimizer": {
|
| 5 |
+
"device": "cpu",
|
| 6 |
+
"pin_memory": true
|
| 7 |
+
},
|
| 8 |
+
"offload_param": {
|
| 9 |
+
"device": "cpu",
|
| 10 |
+
"pin_memory": true
|
| 11 |
+
},
|
| 12 |
+
"overlap_comm": true,
|
| 13 |
+
"contiguous_gradients": true,
|
| 14 |
+
"sub_group_size": 0,
|
| 15 |
+
"reduce_bucket_size": "auto",
|
| 16 |
+
"stage3_prefetch_bucket_size": "auto",
|
| 17 |
+
"stage3_param_persistence_threshold": "auto",
|
| 18 |
+
"stage3_max_live_parameters": 0,
|
| 19 |
+
"stage3_max_reuse_distance": 0,
|
| 20 |
+
"stage3_gather_16bit_weights_on_model_save": true
|
| 21 |
+
},
|
| 22 |
+
"bf16": {
|
| 23 |
+
"enabled": true
|
| 24 |
+
},
|
| 25 |
+
"fp16": {
|
| 26 |
+
"enabled": "auto",
|
| 27 |
+
"auto_cast": false,
|
| 28 |
+
"loss_scale": 0,
|
| 29 |
+
"initial_scale_power": 32,
|
| 30 |
+
"loss_scale_window": 1000,
|
| 31 |
+
"hysteresis": 2,
|
| 32 |
+
"min_loss_scale": 1
|
| 33 |
+
},
|
| 34 |
+
"gradient_accumulation_steps": "auto",
|
| 35 |
+
"gradient_clipping": "auto",
|
| 36 |
+
"train_batch_size": "auto",
|
| 37 |
+
"train_micro_batch_size_per_gpu": "auto",
|
| 38 |
+
"wall_clock_breakdown": false
|
| 39 |
+
}
|
deepspeed_configs/zero3_bf16_cpuoffload_params.json
ADDED
|
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"zero_optimization": {
|
| 3 |
+
"stage": 3,
|
| 4 |
+
"offload_param": {
|
| 5 |
+
"device": "cpu",
|
| 6 |
+
"pin_memory": true
|
| 7 |
+
},
|
| 8 |
+
"overlap_comm": true,
|
| 9 |
+
"contiguous_gradients": true,
|
| 10 |
+
"sub_group_size": 0,
|
| 11 |
+
"reduce_bucket_size": "auto",
|
| 12 |
+
"stage3_prefetch_bucket_size": "auto",
|
| 13 |
+
"stage3_param_persistence_threshold": "auto",
|
| 14 |
+
"stage3_max_live_parameters": 0,
|
| 15 |
+
"stage3_max_reuse_distance": 0,
|
| 16 |
+
"stage3_gather_16bit_weights_on_model_save": true
|
| 17 |
+
},
|
| 18 |
+
"bf16": {
|
| 19 |
+
"enabled": true
|
| 20 |
+
},
|
| 21 |
+
"fp16": {
|
| 22 |
+
"enabled": "auto",
|
| 23 |
+
"auto_cast": false,
|
| 24 |
+
"loss_scale": 0,
|
| 25 |
+
"initial_scale_power": 32,
|
| 26 |
+
"loss_scale_window": 1000,
|
| 27 |
+
"hysteresis": 2,
|
| 28 |
+
"min_loss_scale": 1
|
| 29 |
+
},
|
| 30 |
+
"gradient_accumulation_steps": "auto",
|
| 31 |
+
"gradient_clipping": "auto",
|
| 32 |
+
"train_batch_size": "auto",
|
| 33 |
+
"train_micro_batch_size_per_gpu": "auto",
|
| 34 |
+
"wall_clock_breakdown": false
|
| 35 |
+
}
|