[PR #672] [CLOSED] 修复 DeepEPv2 反向 MoE 场景的路由权重梯度传播问题 #695

Closed
opened 2026-10-27 21:05:21 +00:00 by bsakaguchi · 0 comments
Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/672
Author: @WANGWEI011
Created: 6/26/2026
Status: ❌ Closed

Base: main ← Head: wangwei-deepepv2-0626


📝 Commits (1)

  • fb73fe9 修复 DeepEPv2 反向 MoE 场景的路由权重梯度传播问题

📊 Changes

8 files changed (+242 additions, -45 deletions)

View changed files

📝 csrc/elastic/buffer.hpp (+88 -12)
📝 csrc/kernels/elastic/dispatch.hpp (+18 -5)
📝 deep_ep/buffers/elastic.py (+36 -10)
📝 deep_ep/include/deep_ep/common/layout.cuh (+11 -1)
📝 deep_ep/include/deep_ep/impls/combine.cuh (+15 -3)
📝 deep_ep/include/deep_ep/impls/combine_reduce_epilogue.cuh (+28 -10)
📝 deep_ep/include/deep_ep/impls/dispatch.cuh (+3 -0)
📝 deep_ep/include/deep_ep/impls/dispatch_copy_epilogue.cuh (+43 -4)

📄 Description

(cherry picked from commit 7a9715dffbb47a33a7aebbb337d507b17d3408ec)


🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.

## 📋 Pull Request Information **Original PR:** https://github.com/deepseek-ai/DeepEP/pull/672 **Author:** [@WANGWEI011](https://github.com/WANGWEI011) **Created:** 6/26/2026 **Status:** ❌ Closed **Base:** `main` ← **Head:** `wangwei-deepepv2-0626` --- ### 📝 Commits (1) - [`fb73fe9`](https://github.com/deepseek-ai/DeepEP/commit/fb73fe94d0bd0b11e97924644265797754b7f608) 修复 DeepEPv2 反向 MoE 场景的路由权重梯度传播问题 ### 📊 Changes **8 files changed** (+242 additions, -45 deletions) <details> <summary>View changed files</summary> 📝 `csrc/elastic/buffer.hpp` (+88 -12) 📝 `csrc/kernels/elastic/dispatch.hpp` (+18 -5) 📝 `deep_ep/buffers/elastic.py` (+36 -10) 📝 `deep_ep/include/deep_ep/common/layout.cuh` (+11 -1) 📝 `deep_ep/include/deep_ep/impls/combine.cuh` (+15 -3) 📝 `deep_ep/include/deep_ep/impls/combine_reduce_epilogue.cuh` (+28 -10) 📝 `deep_ep/include/deep_ep/impls/dispatch.cuh` (+3 -0) 📝 `deep_ep/include/deep_ep/impls/dispatch_copy_epilogue.cuh` (+43 -4) </details> ### 📄 Description (cherry picked from commit 7a9715dffbb47a33a7aebbb337d507b17d3408ec) --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>
bsakaguchi 2026-10-27 21:05:21 +00:00
Sign in to join this conversation.
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/DeepEP#695
No description provided.