cmedina/agenx-lora-training
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: use SHARD_GRAD_OP instead of FULL_SHARD for 2-GPU setup

Browse Source
This commit is contained in:
Christian Medina
2026-07-02 00:58:41 -04:00
parent e3ea45fe96
commit dbc97138e9
1 changed files with 4 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
train.py
View File

@@ -169,12 +169,14 @@ def train(config_path):
# gradient_checkpointing removed - FSDP activation_checkpointing handles it # gradient_checkpointing removed - FSDP activation_checkpointing handles it
fsdp=True, fsdp=True,
fsdp_config={ fsdp_config={
"sharding_strategy": "SHARD_GRAD_OP",
"cpu_offload": False,
"activation_checkpointing": True,
"limit_all_gathers": True,
"backward_prefetch": "backward_pre", "backward_prefetch": "backward_pre",
"forward_prefetch": "true", "forward_prefetch": "true",
"cpu_offload": "false",
"auto_wrap_policy": "TRANSFORMER_BASED_WRAP", "auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
"transformer_layer_cls_to_wrap": "Qwen3_5MoeDecoderLayer", "transformer_layer_cls_to_wrap": "Qwen3_5MoeDecoderLayer",
"activation_checkpointing": True,
}, },
) )