[PR #556] [MERGED] Add custom AG , remove the usage of some environments #635

New issue

Closed

opened 2026-10-27 21:04:43 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-27 21:04:43 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/556
Author: @Autumn1998
Created: 1/6/2026
Status: ✅ Merged
Merged: 1/20/2026
Merged by: @jershi425

Base: hybrid-ep ← Head: tongliu_dev

📝 Commits (10+)

13d8379 Adapt for cuda graph
efdc76a add hybrid ep ut
0eb433c [BUG fix] add system fenc
f4a8fd8 fix token_per_exp in overflow check
9d115d7 [Fix] use less threads in scan kernel in EP64, and misc
20c7cb9 rm ENV: NUM_HYBRIE_EP_RANKS_PER_NVLINK_DOMAIN + USE_MNNVL
fb3ff5e fix test
5eb0722 fiix test on gb
415787b fix on B200
6fdd17b fix doc, code clean

📊 Changes

24 files changed (+1043 additions, -244 deletions)

View changed files

📝 .gitmodules (+1 -0)
📝 Hybrid-EP_Implementation.md (+1 -9)
📝 csrc/hybrid_ep/allocator/allocator.cu (+118 -31)
📝 csrc/hybrid_ep/allocator/allocator.cuh (+22 -6)
📝 csrc/hybrid_ep/backend/hybrid_ep_backend.cuh (+2 -2)
📝 csrc/hybrid_ep/config.cuh (+8 -0)
📝 csrc/hybrid_ep/executor/executor.cu (+56 -21)
📝 csrc/hybrid_ep/executor/executor.cuh (+13 -3)
➕ csrc/hybrid_ep/extension/allgather.cu (+244 -0)
➕ csrc/hybrid_ep/extension/allgather.cuh (+37 -0)
📝 csrc/hybrid_ep/extension/permute.cu (+73 -29)
📝 csrc/hybrid_ep/extension/permute.cuh (+5 -1)
📝 csrc/hybrid_ep/hybrid_ep.cu (+76 -14)
📝 csrc/hybrid_ep/hybrid_ep.cuh (+6 -6)
📝 csrc/hybrid_ep/internode.cu (+25 -21)
📝 csrc/hybrid_ep/internode.cuh (+3 -2)
📝 csrc/hybrid_ep/jit/compiler.cu (+25 -13)
📝 csrc/hybrid_ep/jit/compiler.cuh (+3 -2)
📝 csrc/hybrid_ep/pybind_hybrid_ep.cu (+7 -2)
📝 deep_ep/hybrid_ep_buffer.py (+44 -54)

...and 4 more files

📄 Description

Add custom AG on the intra-node case
Remove the ENV
Several small fix

_{🔄 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/556 **Author:** [@Autumn1998](https://github.com/Autumn1998) **Created:** 1/6/2026 **Status:** ✅ Merged **Merged:** 1/20/2026 **Merged by:** [@jershi425](https://github.com/jershi425) **Base:** `hybrid-ep` ← **Head:** `tongliu_dev` --- ### 📝 Commits (10+) - [`13d8379`](https://github.com/deepseek-ai/DeepEP/commit/13d837996db915094b01cb604e5748ceb09c82c0) Adapt for cuda graph - [`efdc76a`](https://github.com/deepseek-ai/DeepEP/commit/efdc76a5fa553def620ec062671983123e756c03) add hybrid ep ut - [`0eb433c`](https://github.com/deepseek-ai/DeepEP/commit/0eb433ccaf99faed8c77ae190195e1535b825e59) [BUG fix] add system fenc - [`f4a8fd8`](https://github.com/deepseek-ai/DeepEP/commit/f4a8fd812248b54158a99b28d10e69bc9a650a36) fix token_per_exp in overflow check - [`9d115d7`](https://github.com/deepseek-ai/DeepEP/commit/9d115d72567bf791d91acaa0ffe29d6ca8d7d1d6) [Fix] use less threads in scan kernel in EP64, and misc - [`20c7cb9`](https://github.com/deepseek-ai/DeepEP/commit/20c7cb978f7f93d53a0bd55ad8e4cdf5ad0b4d82) rm ENV: NUM_HYBRIE_EP_RANKS_PER_NVLINK_DOMAIN + USE_MNNVL - [`fb3ff5e`](https://github.com/deepseek-ai/DeepEP/commit/fb3ff5e6376f0d1b835a99ee6b48ee6de81040ac) fix test - [`5eb0722`](https://github.com/deepseek-ai/DeepEP/commit/5eb0722d205ecca498ae3e32fb14ba3c3aa00fb7) fiix test on gb - [`415787b`](https://github.com/deepseek-ai/DeepEP/commit/415787b08d7b90dedd1f57a1aff13e3ae688f841) fix on B200 - [`6fdd17b`](https://github.com/deepseek-ai/DeepEP/commit/6fdd17b23addc02638e81d73650f6725dfc649d9) fix doc, code clean ### 📊 Changes **24 files changed** (+1043 additions, -244 deletions) <details> <summary>View changed files</summary> 📝 `.gitmodules` (+1 -0) 📝 `Hybrid-EP_Implementation.md` (+1 -9) 📝 `csrc/hybrid_ep/allocator/allocator.cu` (+118 -31) 📝 `csrc/hybrid_ep/allocator/allocator.cuh` (+22 -6) 📝 `csrc/hybrid_ep/backend/hybrid_ep_backend.cuh` (+2 -2) 📝 `csrc/hybrid_ep/config.cuh` (+8 -0) 📝 `csrc/hybrid_ep/executor/executor.cu` (+56 -21) 📝 `csrc/hybrid_ep/executor/executor.cuh` (+13 -3) ➕ `csrc/hybrid_ep/extension/allgather.cu` (+244 -0) ➕ `csrc/hybrid_ep/extension/allgather.cuh` (+37 -0) 📝 `csrc/hybrid_ep/extension/permute.cu` (+73 -29) 📝 `csrc/hybrid_ep/extension/permute.cuh` (+5 -1) 📝 `csrc/hybrid_ep/hybrid_ep.cu` (+76 -14) 📝 `csrc/hybrid_ep/hybrid_ep.cuh` (+6 -6) 📝 `csrc/hybrid_ep/internode.cu` (+25 -21) 📝 `csrc/hybrid_ep/internode.cuh` (+3 -2) 📝 `csrc/hybrid_ep/jit/compiler.cu` (+25 -13) 📝 `csrc/hybrid_ep/jit/compiler.cuh` (+3 -2) 📝 `csrc/hybrid_ep/pybind_hybrid_ep.cu` (+7 -2) 📝 `deep_ep/hybrid_ep_buffer.py` (+44 -54) _...and 4 more files_ </details> ### 📄 Description 1. Add custom AG on the intra-node case 2. Remove the ENV 3. Several small fix --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-27 21:04:43 +00:00