[PR #758] Fix pinned host count handling in unpermute #762

New issue

Open

opened 2026-10-27 21:06:02 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-27 21:06:02 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/758
Author: @Autumn1998
Created: 9/21/2026
Status: 🔄 Open

Base: hybrid-ep ← Head: fix/issue756-unpermute-scalar

📝 Commits (1)

15e38df Prefer host token count values in standalone unpermute

📊 Changes

5 files changed (+9 additions, -1 deletions)

View changed files

📝 csrc/hybrid_ep/executor/executor.cu (+1 -0)
📝 csrc/hybrid_ep/executor/executor.cuh (+1 -0)
📝 csrc/hybrid_ep/extension/permute.cu (+4 -1)
📝 csrc/hybrid_ep/extension/permute.cuh (+1 -0)
📝 csrc/hybrid_ep/hybrid_ep.cu (+2 -0)

📄 Description

Pass CPU token counts to unpermute by value to avoid asynchronous reads of pinned host memory. Preserve device-pointer reads for CUDA counts.

Validated with tests/test_hybrid_ep.py --num-processes 4 on 4 GB200 GPUs: all six BF16/FP8 correctness groups passed.

_{🔄 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/758 **Author:** [@Autumn1998](https://github.com/Autumn1998) **Created:** 9/21/2026 **Status:** 🔄 Open **Base:** `hybrid-ep` ← **Head:** `fix/issue756-unpermute-scalar` --- ### 📝 Commits (1) - [`15e38df`](https://github.com/deepseek-ai/DeepEP/commit/15e38df2d78aaec6bf27beaa511c9ab7044859c3) Prefer host token count values in standalone unpermute ### 📊 Changes **5 files changed** (+9 additions, -1 deletions) <details> <summary>View changed files</summary> 📝 `csrc/hybrid_ep/executor/executor.cu` (+1 -0) 📝 `csrc/hybrid_ep/executor/executor.cuh` (+1 -0) 📝 `csrc/hybrid_ep/extension/permute.cu` (+4 -1) 📝 `csrc/hybrid_ep/extension/permute.cuh` (+1 -0) 📝 `csrc/hybrid_ep/hybrid_ep.cu` (+2 -0) </details> ### 📄 Description Pass CPU token counts to unpermute by value to avoid asynchronous reads of pinned host memory. Preserve device-pointer reads for CUDA counts. Related issue: https://github.com/deepseek-ai/DeepEP/issues/756 Validated with tests/test_hybrid_ep.py --num-processes 4 on 4 GB200 GPUs: all six BF16/FP8 correctness groups passed. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi added the

pull-request

label

2026-10-27 21:06:02 +00:00

bsakaguchi referenced this issue

2026-10-27 21:06:03 +00:00

[PR #762] Fix hybrid combine top-k weight gradients with single reduction #764