Skip to content

actor rollout hidden states Transferqueue implement - #48

Open
zjuKirito wants to merge 5 commits into
verl-project:mainfrom
zjuKirito:transferqueue
Open

actor rollout hidden states Transferqueue implement#48
zjuKirito wants to merge 5 commits into
verl-project:mainfrom
zjuKirito:transferqueue

Conversation

@zjuKirito

@zjuKirito zjuKirito commented Aug 3, 2026

Copy link
Copy Markdown

TQ_experiment_report.md
SpeCo 自有的 hidden state 传输已全部覆盖。剩余的标准 PPO 数据面是 verl 上游 TQ 的范畴,需切 main_ppo_sync,属更大工程,且仅多机有实质收益。单机场景下 SpeCo 的 TQ 能做的已完成。
具体实验参考数据参考markdown,和基线对比实验,数据无损(acceptance_length 匹配)、无报错、drafter 训练正常。在timing_s/old_log_prob ,用tq实现hidden state,每两步收集一次,平均劣化3秒,端到端偏差较大,原因是波动本身较大(跑两次基线实验验证,端到端波动较大)。
劣化解释:
hidden state本就用 ray.put ray.get绕过 driver,TQ 只是换了传输方式(且 GET 更慢)。多机有(跨节点 DataProto 传输),单机有限(本地 driver 无瓶颈)
transfer_queue_verify_20260802_030633_eurus_rloo_drafter_online_0801_czy_trasnferqueue.log
transfer_queue_verify_20260802_074255_eurus_rloo_drafter_online_no_tq_0801_main.log

zjuKirito and others added 5 commits July 27, 2026 16:36
Conflict was in _resolve_hidden_state_chunks call site:
- transferqueue branch added cache=self._tq_chunk_cache (cross-sample get cache)
- upstream/main reformatted to multi-line (pre-commit/black)
Resolution: keep cache param + adopt upstream multi-line formatting

Co-Authored-By: Claude <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant