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

fix: disable sync_module_states + add mixed precision for FSDP

Browse Source
This commit is contained in:
Christian Medina
2026-07-02 09:14:18 -04:00
parent 4eb06a625e
commit 42d61c15b0
1 changed files with 3 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
train.py
View File

@@ -169,17 +169,18 @@ def train(config_path):
eval_steps=config.get("eval_steps", 100), eval_steps=config.get("eval_steps", 100),
bf16=True, bf16=True,
# gradient_checkpointing removed - FSDP activation_checkpointing handles it # gradient_checkpointing removed - FSDP activation_checkpointing handles it
fsdp="full_shard", fsdp=True,
fsdp_config={ fsdp_config={
"sharding_strategy": "SHARD_GRAD_OP", "sharding_strategy": "SHARD_GRAD_OP",
"cpu_offload": False, "cpu_offload": False,
"activation_checkpointing": True, "activation_checkpointing": True,
"limit_all_gathers": True, "limit_all_gathers": True,
"sync_module_states": True, "sync_module_states": False, # Don't sync from CPU, FSDP will handle placement
"backward_prefetch": "backward_pre", "backward_prefetch": "backward_pre",
"forward_prefetch": "true", "forward_prefetch": "true",
"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",
"mixed_precision": {"param_dtype": torch.bfloat16, "reduce_dtype": torch.float32, "buffer_dtype": torch.float32},
}, },
) )