[PR #724] [MERGED] [HybridEP] Support configurable RoCE settings in DOCA backend #732

New issue

Closed

opened 2026-10-27 21:05:42 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-27 21:05:42 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/724
Author: @juntaowww
Created: 8/14/2026
Status: ✅ Merged
Merged: 8/20/2026
Merged by: @jershi425

Base: hybrid-ep ← Head: fix/hybridep-roce-config

📝 Commits (2)

5d8e760 [HybridEP] Support configurable RoCE settings in DOCA backend
c1e9599 Fix RoCE configuration edge cases

📊 Changes

3 files changed (+97 additions, -7 deletions)

View changed files

📝 csrc/hybrid_ep/backend/ibvcore.h (+53 -3)
📝 csrc/hybrid_ep/buffer/internode_doca.cu (+35 -3)
📝 docs/Hybrid-EP_Implementation.md (+9 -1)

📄 Description

Summary

Add support for standard NCCL RoCE settings in the HybridEP DOCA backend.
Configure GID selection, address family, and traffic class through NCCL_IB_GID_INDEX, NCCL_IB_ADDR_FAMILY, and NCCL_IB_TC.
Preserve the existing native InfiniBand FLID selection behavior.
Document the supported network settings.

Why

HybridEP previously used default RoCE settings, including automatic GID selection, IPv4 addressing, and traffic class 0. This can create invalid DOCA queue pairs on fabrics requiring a specific GID, IPv6, or non-default traffic class, resulting in requester CQ errors and device-side assertions, Assertion `status >= 0` failed.

Using standard NCCL environment variables makes HybridEP compatible with these RoCE configurations without introducing backend-specific settings.

_{🔄 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/724 **Author:** [@juntaowww](https://github.com/juntaowww) **Created:** 8/14/2026 **Status:** ✅ Merged **Merged:** 8/20/2026 **Merged by:** [@jershi425](https://github.com/jershi425) **Base:** `hybrid-ep` ← **Head:** `fix/hybridep-roce-config` --- ### 📝 Commits (2) - [`5d8e760`](https://github.com/deepseek-ai/DeepEP/commit/5d8e76056135c76fc9376110e3f38c661fb670da) [HybridEP] Support configurable RoCE settings in DOCA backend - [`c1e9599`](https://github.com/deepseek-ai/DeepEP/commit/c1e9599b433d3ea912e8b47abba3abd4b70f9c39) Fix RoCE configuration edge cases ### 📊 Changes **3 files changed** (+97 additions, -7 deletions) <details> <summary>View changed files</summary> 📝 `csrc/hybrid_ep/backend/ibvcore.h` (+53 -3) 📝 `csrc/hybrid_ep/buffer/internode_doca.cu` (+35 -3) 📝 `docs/Hybrid-EP_Implementation.md` (+9 -1) </details> ### 📄 Description ## Summary Add support for standard NCCL RoCE settings in the HybridEP DOCA backend. Configure GID selection, address family, and traffic class through `NCCL_IB_GID_INDEX`, `NCCL_IB_ADDR_FAMILY`, and `NCCL_IB_TC`. Preserve the existing native InfiniBand FLID selection behavior. Document the supported network settings. ## Why HybridEP previously used default RoCE settings, including automatic GID selection, IPv4 addressing, and traffic class 0. This can create invalid DOCA queue pairs on fabrics requiring a specific GID, IPv6, or non-default traffic class, resulting in requester CQ errors and device-side assertions, ```Assertion `status >= 0` failed.``` Using standard NCCL environment variables makes HybridEP compatible with these RoCE configurations without introducing backend-specific settings. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-27 21:05:42 +00:00