mirror of
https://gitee.com/ascend/MindSpeed-RL.git
synced 2025-10-20 16:23:45 +08:00
!553 [bug_fix]experience_count
Merge pull request !553 from zhoubeirong/grpo_32b
This commit is contained in:
@ -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')
|
||||
|
||||
|
Reference in New Issue
Block a user