Skip to content

Conversation

@gzy19990617
Copy link
Collaborator

@gzy19990617 gzy19990617 commented Sep 5, 2025

  1. 支持 EP Rollout 模型初始化。
  2. 告别 Hard Code:num_nvl_bytes 和 num_rdma_bytes 由模型参数动态计算,单机 8 卡下每卡节省 5G+ 显存。
  3. 引入 DeepEPBufferManager:实现DeeepEP buffer的动态申请与释放,按需分配,进一步节约显存。
  4. 重构部分ep.py,抽象出 DeepEPBuffer 类,代码结构更清晰、简洁(未来还需进一步优化)。

@paddle-bot
Copy link

paddle-bot bot commented Sep 5, 2025

Thanks for your contribution!

@gzy19990617 gzy19990617 force-pushed the my-feature branch 5 times, most recently from 7ccfad0 to 1b44e52 Compare September 10, 2025 14:36
@gzy19990617 gzy19990617 changed the title [NewFeture]add ep rollout model init [NewFeture]add ep rollout model init and update/clear ep buffer Sep 10, 2025
@gzy19990617 gzy19990617 changed the base branch from release/2.2 to feature/experimental_feature_20250908 September 10, 2025 14:50
@yuanlehome
Copy link
Collaborator

yuanlehome commented Sep 11, 2025

此PR与 #4051 有冲突,后续提到develop时需注意

@gzy19990617
Copy link
Collaborator Author

此PR与 #4051 有冲突,后续提到develop时需注意

好的

@Jiang-Jia-Jun Jiang-Jia-Jun merged commit 10768a4 into PaddlePaddle:feature/experimental_feature_20250908 Sep 12, 2025
14 of 15 checks passed
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.

3 participants