[PR #709] [MERGED] zero-initialize overflow_flag to prevent stale garbage value #723

New issue

Closed

opened 2026-10-27 21:05:37 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-27 21:05:37 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/709
Author: @Autumn1998
Created: 7/30/2026
Status: ✅ Merged
Merged: 8/4/2026
Merged by: @jershi425

Base: hybrid-ep ← Head: fix-token-drop-uninit

📝 Commits (2)

0757253 Fix uninitialized overflow_flag by using torch::zeros instead of torch::empty
6df62af fix: zero-initialize overflow_flag to prevent stale garbage value

📊 Changes

1 file changed (+1 additions, -0 deletions)

View changed files

📝 csrc/hybrid_ep/executor/executor.cu (+1 -0)

📄 Description

After #625 removed the standalone permute_preprocessing that unconditionally wrote *overflow_flag = 0, the non-fused path leaves overflow_flag
uninitialized (torch::empty garbage), causing callers to see a false-positive token drop on every dispatch.

_{🔄 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/709 **Author:** [@Autumn1998](https://github.com/Autumn1998) **Created:** 7/30/2026 **Status:** ✅ Merged **Merged:** 8/4/2026 **Merged by:** [@jershi425](https://github.com/jershi425) **Base:** `hybrid-ep` ← **Head:** `fix-token-drop-uninit` --- ### 📝 Commits (2) - [`0757253`](https://github.com/deepseek-ai/DeepEP/commit/0757253ea6c5b7ba23a38c30110bcadcfb65b9db) Fix uninitialized overflow_flag by using torch::zeros instead of torch::empty - [`6df62af`](https://github.com/deepseek-ai/DeepEP/commit/6df62af90381eee6f67715ad88238260c4126404) fix: zero-initialize overflow_flag to prevent stale garbage value ### 📊 Changes **1 file changed** (+1 additions, -0 deletions) <details> <summary>View changed files</summary> 📝 `csrc/hybrid_ep/executor/executor.cu` (+1 -0) </details> ### 📄 Description After #625 removed the standalone permute_preprocessing that unconditionally wrote *overflow_flag = 0, the non-fused path leaves overflow_flag uninitialized (torch::empty garbage), causing callers to see a false-positive token drop on every dispatch. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-27 21:05:37 +00:00