fix: set gradient_checkpointing=False (FSDP handles it)
This commit is contained in:
2
train.py
2
train.py
@@ -168,7 +168,7 @@ def train(config_path):
|
||||
eval_strategy=config.get("eval_strategy", "steps"),
|
||||
eval_steps=config.get("eval_steps", 100),
|
||||
bf16=True,
|
||||
gradient_checkpointing=True,
|
||||
gradient_checkpointing=False, # FSDP activation_checkpointing handles it
|
||||
fsdp=True,
|
||||
fsdp_config={
|
||||
"sharding_strategy": "SHARD_GRAD_OP",
|
||||
|
||||
Reference in New Issue
Block a user