[PR #474] [CLOSED] Add RDMA support for hybrid-EP #591

New issue

Closed

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

bsakaguchi commented

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

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/474
Author: @Autumn1998
Created: 10/31/2025
Status: ❌ Closed

Base: hybrid-ep ← Head: tongliu/inter_node

📝 Commits (10+)

8fcaea2 init 3rd party
507f358 rdma dependency
8b960f4 inter-node
2871bb8 The left part of the inter-node
2e20529 Merge remote-tracking branch 'origin/hybrid-ep' into tongliu/inter_node
2313535 make compiler work
aa1fcd4 fix bug
5c11072 fix bugs, make it runnable on the intra-node case
aa3740b fix some bug
217b5b1 fix jit

📊 Changes

19 files changed (+2091 additions, -300 deletions)

View changed files

➕ .gitmodules (+3 -0)
📝 Hybrid-EP_Implementation.md (+5 -6)
📝 csrc/hybrid_ep/backend/hybrid_ep_backend.cuh (+671 -49)
📝 csrc/hybrid_ep/backend/utils.cuh (+85 -30)
📝 csrc/hybrid_ep/config.cuh (+2 -0)
📝 csrc/hybrid_ep/executor/executor.cu (+40 -11)
📝 csrc/hybrid_ep/executor/executor.cuh (+1 -1)
📝 csrc/hybrid_ep/hybrid_ep.cu (+97 -110)
📝 csrc/hybrid_ep/hybrid_ep.cuh (+13 -2)
➕ csrc/hybrid_ep/internode.cu (+717 -0)
➕ csrc/hybrid_ep/internode.cuh (+196 -0)
📝 csrc/hybrid_ep/jit/compiler.cu (+45 -24)
📝 csrc/hybrid_ep/jit/compiler.cuh (+4 -2)
📝 csrc/hybrid_ep/pybind_hybrid_ep.cu (+24 -9)
📝 deep_ep/hybrid_ep_buffer.py (+36 -13)
📝 setup.py (+79 -10)
📝 tests/test_hybrid_ep.py (+50 -27)
📝 tests/utils.py (+22 -6)
➕ third-party/nccl (+1 -0)

📄 Description

Add RDMA support for hybrid-EP, WIP

_{🔄 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/474 **Author:** [@Autumn1998](https://github.com/Autumn1998) **Created:** 10/31/2025 **Status:** ❌ Closed **Base:** `hybrid-ep` ← **Head:** `tongliu/inter_node` --- ### 📝 Commits (10+) - [`8fcaea2`](https://github.com/deepseek-ai/DeepEP/commit/8fcaea2067135909eccf398b0141f90e865034fe) init 3rd party - [`507f358`](https://github.com/deepseek-ai/DeepEP/commit/507f358ef736db82af370700663b2b60c1cd5570) rdma dependency - [`8b960f4`](https://github.com/deepseek-ai/DeepEP/commit/8b960f428fd4edfeb3081ebe225215582edc39cc) inter-node - [`2871bb8`](https://github.com/deepseek-ai/DeepEP/commit/2871bb8022bdc8ec8082463b09e96ef2356ad1cd) The left part of the inter-node - [`2e20529`](https://github.com/deepseek-ai/DeepEP/commit/2e205290b1c2828bf62bc2993f234d06b193236a) Merge remote-tracking branch 'origin/hybrid-ep' into tongliu/inter_node - [`2313535`](https://github.com/deepseek-ai/DeepEP/commit/2313535d73a65ac478b3fbae794d9659d34200d8) make compiler work - [`aa1fcd4`](https://github.com/deepseek-ai/DeepEP/commit/aa1fcd410d13aac6afb937894890a5e600998cc0) fix bug - [`5c11072`](https://github.com/deepseek-ai/DeepEP/commit/5c110724e5c5945e999237263711a793e1380847) fix bugs, make it runnable on the intra-node case - [`aa3740b`](https://github.com/deepseek-ai/DeepEP/commit/aa3740bc876ad2e57f8701db7be60a50524aaeed) fix some bug - [`217b5b1`](https://github.com/deepseek-ai/DeepEP/commit/217b5b10cfc61907b1a81758d2b79092b36e4f22) fix jit ### 📊 Changes **19 files changed** (+2091 additions, -300 deletions) <details> <summary>View changed files</summary> ➕ `.gitmodules` (+3 -0) 📝 `Hybrid-EP_Implementation.md` (+5 -6) 📝 `csrc/hybrid_ep/backend/hybrid_ep_backend.cuh` (+671 -49) 📝 `csrc/hybrid_ep/backend/utils.cuh` (+85 -30) 📝 `csrc/hybrid_ep/config.cuh` (+2 -0) 📝 `csrc/hybrid_ep/executor/executor.cu` (+40 -11) 📝 `csrc/hybrid_ep/executor/executor.cuh` (+1 -1) 📝 `csrc/hybrid_ep/hybrid_ep.cu` (+97 -110) 📝 `csrc/hybrid_ep/hybrid_ep.cuh` (+13 -2) ➕ `csrc/hybrid_ep/internode.cu` (+717 -0) ➕ `csrc/hybrid_ep/internode.cuh` (+196 -0) 📝 `csrc/hybrid_ep/jit/compiler.cu` (+45 -24) 📝 `csrc/hybrid_ep/jit/compiler.cuh` (+4 -2) 📝 `csrc/hybrid_ep/pybind_hybrid_ep.cu` (+24 -9) 📝 `deep_ep/hybrid_ep_buffer.py` (+36 -13) 📝 `setup.py` (+79 -10) 📝 `tests/test_hybrid_ep.py` (+50 -27) 📝 `tests/utils.py` (+22 -6) ➕ `third-party/nccl` (+1 -0) </details> ### 📄 Description Add RDMA support for hybrid-EP, WIP --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

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