fix: remove gradient_checkpointing (FSDP activation_checkpointing handles it)
This commit is contained in:
2
train.py
2
train.py
@@ -166,7 +166,7 @@ def train(config_path):
|
||||
eval_strategy=config.get("eval_strategy", "steps"),
|
||||
eval_steps=config.get("eval_steps", 100),
|
||||
bf16=True,
|
||||
gradient_checkpointing=config.get("gradient_checkpointing", True),
|
||||
# gradient_checkpointing removed - FSDP activation_checkpointing handles it
|
||||
fsdp=True,
|
||||
fsdp_config={
|
||||
"backward_prefetch": "backward_pre",
|
||||
|
||||
Reference in New Issue
Block a user