已合并
remove dapo_ray_trainer patch #848
yezhibei创建于 2025年12月16日
remove dapo_ray_trainer patch #848
已合并
yezhibei创建于 2025年12月16日
2 个文件变更+1-28
@@ -1,22 +0,0 @@
1-diff --git a/recipe/dapo/dapo_ray_trainer.py b/recipe/dapo/dapo_ray_trainer.py
2-index 18995f67..db209482 100644
3---- a/recipe/dapo/dapo_ray_trainer.py
4-+++ b/recipe/dapo/dapo_ray_trainer.py
5-@@ -156,7 +156,7 @@ class RayDAPOTrainer(RayPPOTrainer):
6- non_tensor_batch_keys=["raw_prompt_ids"],
7- )
8- gen_batch_output = gen_batch.repeat(
9-- repeat_times=self.config.actor_rollout_ref.rollout.n, interleave=True
10-+ repeat_times=self.config.actor_rollout_ref.rollout.n, interleave=False
11- )
12-
13- is_last_step = self.global_steps >= self.total_training_steps
14-@@ -196,7 +196,7 @@ class RayDAPOTrainer(RayPPOTrainer):
15- [str(uuid.uuid4()) for _ in range(len(new_batch.batch))], dtype=object
16- )
17- # repeat to align with repeated responses in rollout
18-- new_batch = new_batch.repeat(repeat_times=self.config.actor_rollout_ref.rollout.n, interleave=True)
19-+ new_batch = new_batch.repeat(repeat_times=self.config.actor_rollout_ref.rollout.n, interleave=False)
20- new_batch = new_batch.union(gen_batch_output)
21-
22- if self.config.algorithm.use_kl_in_reward:
@@ -41,9 +41,4 @@ patches:
41 name: __init__41 name: __init__
42 - action: replaced # [added, replaced, deleted]42 - action: replaced # [added, replaced, deleted]
43 kind: method # [class, method, attribute]43 kind: method # [class, method, attribute]
44- name: _init_dp_env44+ name: _init_dp_env
45- - name: dapo_ray_trainer
46- diff:
47- class_changes: # [action, kind, name]
48- - action: updated # [added, replaced, deleted, updated]
49- name: RayDAPOTrainer