[PR #625] [MERGED] Optimization of the standalone permute path #672

New issue

Closed

opened 2026-10-27 21:05:08 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-27 21:05:08 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/625
Author: @Autumn1998
Created: 5/7/2026
Status: ✅ Merged
Merged: 5/13/2026
Merged by: @jershi425

Base: hybrid-ep ← Head: tongliu_new_permute

📝 Commits (5)

33b29d4 add new permute
512cc09 code clean
aa5f03a unblock sm count
0dbaaf0 add policy and fix
892a1c4 Fix DOCA build/path regressions from NIXL integration (#606) (#616)

📊 Changes

14 files changed (+520 additions, -762 deletions)

View changed files

➖ .use_nixl (+0 -0)
📝 csrc/hybrid_ep/backend/hybrid_ep_backend.cuh (+1 -1)
📝 csrc/hybrid_ep/buffer/internode_doca.cuh (+8 -0)
📝 csrc/hybrid_ep/buffer/intranode.cu (+1 -2)
📝 csrc/hybrid_ep/config.cuh (+1 -1)
📝 csrc/hybrid_ep/executor/executor.cu (+15 -33)
📝 csrc/hybrid_ep/executor/executor.cuh (+3 -20)
📝 csrc/hybrid_ep/extension/permute.cu (+413 -525)
📝 csrc/hybrid_ep/extension/permute.cuh (+5 -50)
📝 csrc/hybrid_ep/hybrid_ep.cu (+6 -17)
📝 csrc/hybrid_ep/pybind_hybrid_ep.cu (+1 -2)
📝 deep_ep/hybrid_ep_buffer.py (+44 -91)
📝 docs/README_Hybrid-EP.md (+19 -9)
📝 setup.py (+3 -11)

📄 Description

Add new permute kernel, optimize the perf at the overlap case
unify the preprocess part of the fuse-path and non-fuse-path
Fix out-of-bound error on scan kernel

_{🔄 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/625 **Author:** [@Autumn1998](https://github.com/Autumn1998) **Created:** 5/7/2026 **Status:** ✅ Merged **Merged:** 5/13/2026 **Merged by:** [@jershi425](https://github.com/jershi425) **Base:** `hybrid-ep` ← **Head:** `tongliu_new_permute` --- ### 📝 Commits (5) - [`33b29d4`](https://github.com/deepseek-ai/DeepEP/commit/33b29d475bb57933bd4c5e497ed6f37bbe0ab600) add new permute - [`512cc09`](https://github.com/deepseek-ai/DeepEP/commit/512cc093e158b9de007e416054cb8d07bfe86d6e) code clean - [`aa5f03a`](https://github.com/deepseek-ai/DeepEP/commit/aa5f03a895cd6c3388868755f2f8f2f4c748bd0e) unblock sm count - [`0dbaaf0`](https://github.com/deepseek-ai/DeepEP/commit/0dbaaf039c647b7288cfe3900339d5587853b3ac) add policy and fix - [`892a1c4`](https://github.com/deepseek-ai/DeepEP/commit/892a1c40a9f3d1ba52e0f66b0d2e6e5b4f2eb3fd) Fix DOCA build/path regressions from NIXL integration (#606) (#616) ### 📊 Changes **14 files changed** (+520 additions, -762 deletions) <details> <summary>View changed files</summary> ➖ `.use_nixl` (+0 -0) 📝 `csrc/hybrid_ep/backend/hybrid_ep_backend.cuh` (+1 -1) 📝 `csrc/hybrid_ep/buffer/internode_doca.cuh` (+8 -0) 📝 `csrc/hybrid_ep/buffer/intranode.cu` (+1 -2) 📝 `csrc/hybrid_ep/config.cuh` (+1 -1) 📝 `csrc/hybrid_ep/executor/executor.cu` (+15 -33) 📝 `csrc/hybrid_ep/executor/executor.cuh` (+3 -20) 📝 `csrc/hybrid_ep/extension/permute.cu` (+413 -525) 📝 `csrc/hybrid_ep/extension/permute.cuh` (+5 -50) 📝 `csrc/hybrid_ep/hybrid_ep.cu` (+6 -17) 📝 `csrc/hybrid_ep/pybind_hybrid_ep.cu` (+1 -2) 📝 `deep_ep/hybrid_ep_buffer.py` (+44 -91) 📝 `docs/README_Hybrid-EP.md` (+19 -9) 📝 `setup.py` (+3 -11) </details> ### 📄 Description 1. Add new permute kernel, optimize the perf at the overlap case 2. unify the preprocess part of the fuse-path and non-fuse-path 3. Fix out-of-bound error on scan kernel --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-27 21:05:08 +00:00