[PR #660] [CLOSED] Fix PP TMA stage reuse race #693

New issue

Closed

opened 2026-10-27 21:05:19 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-27 21:05:19 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/660
Author: @qescccczmr
Created: 6/4/2026
Status: ❌ Closed

Base: main ← Head: fix-pp-tma-store-wait

📝 Commits (1)

2dfbdb9 Fix PP TMA stage reuse race

📊 Changes

1 file changed (+1 additions, -1 deletions)

View changed files

📝 deep_ep/include/deep_ep/impls/pp_send_recv.cuh (+1 -1)

📄 Description

Fix a PP send/recv TMA copy race when reusing a shared-memory pipeline stage.
This changes the reuse wait to ptx::tma_store_wait() / wait_group 0, ensuring all outstanding TMA
stores complete before the same stage buffer is reused.
Fixes #657.

_{🔄 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/660 **Author:** [@qescccczmr](https://github.com/qescccczmr) **Created:** 6/4/2026 **Status:** ❌ Closed **Base:** `main` ← **Head:** `fix-pp-tma-store-wait` --- ### 📝 Commits (1) - [`2dfbdb9`](https://github.com/deepseek-ai/DeepEP/commit/2dfbdb9e61e288add39db2f06c6333fcffd92bf0) Fix PP TMA stage reuse race ### 📊 Changes **1 file changed** (+1 additions, -1 deletions) <details> <summary>View changed files</summary> 📝 `deep_ep/include/deep_ep/impls/pp_send_recv.cuh` (+1 -1) </details> ### 📄 Description Fix a PP send/recv TMA copy race when reusing a shared-memory pipeline stage. This changes the reuse wait to `ptx::tma_store_wait()` / `wait_group 0`, ensuring all outstanding TMA stores complete before the same stage buffer is reused. Fixes #657. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-27 21:05:19 +00:00