!553 [bug_fix]experience_count

Merge pull request !553 from zhoubeirong/grpo_32b
This commit is contained in:
zhoubeirong
2025-08-25 06:33:24 +00:00
committed by i-robot
parent 870967acc1
commit d75a6148df

View File

@ -215,8 +215,16 @@ class ActorHybridWorkerBase(BaseWorker):
if is_multimodal():
experience_columns.extend(['attention_mask', 'position_ids'])
experience_count = self.rl_config.actor_update_dispatch_size
experience_count = (
self.megatron_config.global_batch_size // self.parallel_state.get_data_parallel_world_size()
)
if self.rl_config.filter_groups_enable:
experience_count = (
self.rl_config.filter_groups_train_batch_size * self.rl_config.n_samples_per_prompt //
self.parallel_state.get_data_parallel_world_size()
)
if skip_actor_log_prob:
experience_columns.remove('old_log_prob')