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

use cfg.max_completion_length, not sequence_len

Browse Source
This commit is contained in:
Wing Lian
2025-02-05 13:20:17 -05:00
parent bdb0f97082
commit 3659d812f7
1 changed files with 7 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
src/axolotl/core/trainers/grpo/__init__.py
View File

@@ -38,14 +38,12 @@ class GRPOStrategy:
else: else:
grpo_args_kwargs["vllm_device"] = "auto" grpo_args_kwargs["vllm_device"] = "auto"
if cfg.grpo_vllm_gpu_memory_utilization: if cfg.grpo_vllm_gpu_memory_utilization:
grpo_args_kwargs["vllm_gpu_memory_utilization"] = ( grpo_args_kwargs[
cfg.grpo_vllm_gpu_memory_utilization "vllm_gpu_memory_utilization"
) ] = cfg.grpo_vllm_gpu_memory_utilization
if cfg.grpo_num_generations: if cfg.grpo_num_generations:
grpo_args_kwargs["num_generations"] = cfg.grpo_num_generations grpo_args_kwargs["num_generations"] = cfg.grpo_num_generations
grpo_args_kwargs["max_completion_length"] = ( grpo_args_kwargs["max_completion_length"] = cfg.max_completion_length
cfg.max_completion_length or cfg.sequence_len
)
return grpo_args_kwargs return grpo_args_kwargs
@classmethod @classmethod
@@ -57,9 +55,9 @@ class GRPOStrategy:
reward_funcs.append(cls.get_reward_func(reward_func_fqn)) reward_funcs.append(cls.get_reward_func(reward_func_fqn))
trainer_kwargs["reward_funcs"] = reward_funcs trainer_kwargs["reward_funcs"] = reward_funcs
if cfg.grpo_reward_processing_classes: if cfg.grpo_reward_processing_classes:
trainer_kwargs["reward_processing_classes"] = ( trainer_kwargs[
cfg.grpo_reward_processing_classes "reward_processing_classes"
) ] = cfg.grpo_reward_processing_classes
return trainer_kwargs return trainer_kwargs
@classmethod @classmethod