tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 3 Packages Projects Releases Wiki Activity

fix: ignore max_length for grpo

Browse Source
This commit is contained in:
NanoCode012
2025-02-21 13:55:34 +07:00
parent edaec9fe98
commit 8aa722a140
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
src/axolotl/core/trainers/grpo/__init__.py
View File

@@ -131,8 +131,8 @@ class GRPOStrategy:
return None return None
@classmethod @classmethod
def get_blocklist_args_kwargs(cls) -> list[str]: def get_blocklist_args_kwargs(cls):
return ["dataset_num_proc"] return ["dataset_num_proc", "max_length"]
@classmethod @classmethod
def get_reward_func(cls, reward_func_fqn: str) -> RewardFunc: def get_reward_func(cls, reward_func_fqn: str) -> RewardFunc:
@@ -167,4 +167,4 @@ class GRPOStrategy:
LOG.info( LOG.info(
f"Reward function {reward_func_fqn} is a pre-trained model path - if this is unexpected, please check the reward function path." f"Reward function {reward_func_fqn} is a pre-trained model path - if this is unexpected, please check the reward function path."
) )
return reward_func return reward_func_fqn