tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

fix: set rollout in GRPO training_kwargs (#3392)

Browse Source
This commit is contained in:
VED
2026-02-10 16:36:15 +05:30
committed by GitHub
parent c67cbcb0f5
commit a4ee56c315
1 changed files with 2 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
src/axolotl/core/trainers/grpo/__init__.py
View File

@@ -126,9 +126,6 @@ class GRPOStrategy:
if trl.use_liger_loss is not None: if trl.use_liger_loss is not None:
grpo_args_kwargs["use_liger_loss"] = trl.use_liger_loss grpo_args_kwargs["use_liger_loss"] = trl.use_liger_loss
if trl.rollout_func:
grpo_args_kwargs["rollout_func"] = cls.get_rollout_func(trl.rollout_func)
if trl.multi_objective_aggregation is not None: if trl.multi_objective_aggregation is not None:
grpo_args_kwargs["multi_objective_aggregation"] = ( grpo_args_kwargs["multi_objective_aggregation"] = (
trl.multi_objective_aggregation trl.multi_objective_aggregation
@@ -154,6 +151,8 @@ class GRPOStrategy:
trainer_kwargs["reward_processing_classes"] = ( trainer_kwargs["reward_processing_classes"] = (
cfg.trl.reward_processing_classes cfg.trl.reward_processing_classes
) )
if cfg.trl and cfg.trl.rollout_func:
trainer_kwargs["rollout_func"] = cls.get_rollout_func(cfg.trl.rollout_func)
return trainer_kwargs return trainer_kwargs