tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

feat: add deepspeed 3 with cpuoffload (#1466)

Browse Source 
* feat: add deepspeed 3 with cpuoffload

* make bf16 explicit, add param only offload variant

---------

Co-authored-by: Wing Lian <wing.lian@gmail.com>
This commit is contained in:
NanoCode012
2024-04-01 21:42:52 +09:00
committed by GitHub
parent 0ddfb24fcf
commit 946b497c3f
2 changed files with 74 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
deepspeed_configs/zero3_bf16_cpuoffload_all.json Normal file
View File

@@ -0,0 +1,39 @@
{
"zero_optimization": {
"stage": 3,
"offload_optimizer": {
"device": "cpu",
"pin_memory": true
},
"offload_param": {
"device": "cpu",
"pin_memory": true
},
"overlap_comm": true,
"contiguous_gradients": true,
"sub_group_size": 0,
"reduce_bucket_size": "auto",
"stage3_prefetch_bucket_size": "auto",
"stage3_param_persistence_threshold": "auto",
"stage3_max_live_parameters": 0,
"stage3_max_reuse_distance": 0,
"stage3_gather_16bit_weights_on_model_save": true
},
"bf16": {
"enabled": true
},
"fp16": {
"enabled": "auto",
"auto_cast": false,
"loss_scale": 0,
"initial_scale_power": 32,
"loss_scale_window": 1000,
"hysteresis": 2,
"min_loss_scale": 1
},
"gradient_accumulation_steps": "auto",
"gradient_clipping": "auto",
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto",
"wall_clock_breakdown": false
}

35
deepspeed_configs/zero3_bf16_cpuoffload_params.json Normal file
View File

@@ -0,0 +1,35 @@
{
"zero_optimization": {
"stage": 3,
"offload_param": {
"device": "cpu",
"pin_memory": true
},
"overlap_comm": true,
"contiguous_gradients": true,
"sub_group_size": 0,
"reduce_bucket_size": "auto",
"stage3_prefetch_bucket_size": "auto",
"stage3_param_persistence_threshold": "auto",
"stage3_max_live_parameters": 0,
"stage3_max_reuse_distance": 0,
"stage3_gather_16bit_weights_on_model_save": true
},
"bf16": {
"enabled": true
},
"fp16": {
"enabled": "auto",
"auto_cast": false,
"loss_scale": 0,
"initial_scale_power": 32,
"loss_scale_window": 1000,
"hysteresis": 2,
"min_loss_scale": 1
},
"gradient_accumulation_steps": "auto",
"gradient_clipping": "auto",
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto",
"wall_clock_breakdown": false
}