[PR #630] [MERGED] Fix V2 initialization for pure single-node intra EP #673

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/630
Author: @DanielDanyang
Created: 5/10/2026
Status: ✅ Merged
Merged: 5/22/2026
Merged by: @sphish

Base: main ← Head: single-node-b300

📝 Commits (1)

fa8a9b1 fix V2 initialization for pure single-node intra EP

📊 Changes

1 file changed (+11 additions, -8 deletions)

View changed files

📝 csrc/kernels/backend/nccl.cu (+11 -8)

📄 Description

Summary

This PR fixes V2/elastic initialization for pure single-node intra-node EP runs where NCCL reports no Gin capability. (#629)

This keeps the existing EP_DISABLE_GIN policy: when the variable is set, DeepEP skips Gin context requests instead of forcing Gin availability during preflight initialization.

Validation

Validated on 8 x NVIDIA B300 GPUs, CUDA 13.2, PyTorch 2.13.0.dev20260502+cu130, and NCCL 2.30.4.

export EP_DISABLE_GIN=1

python tests/elastic/test_ep.py \
  --num-processes 8 \
  --num-tokens 8192 \
  --hidden 7168 \
  --num-topk 8 \
  --num-experts 256 \
  --allow-hybrid-mode 0 \
  --test-first-only

Average across 8 ranks:

Metric	Result	README SM100 EP8 reference
Dispatch	731.9 GB/s SU	726 GB/s NVLink
Combine	741.0 GB/s SU	740 GB/s NVLink

_{🔄 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/630 **Author:** [@DanielDanyang](https://github.com/DanielDanyang) **Created:** 5/10/2026 **Status:** ✅ Merged **Merged:** 5/22/2026 **Merged by:** [@sphish](https://github.com/sphish) **Base:** `main` ← **Head:** `single-node-b300` --- ### 📝 Commits (1) - [`fa8a9b1`](https://github.com/deepseek-ai/DeepEP/commit/fa8a9b16898204afd347c663b89e65ef87dc6ce6) fix V2 initialization for pure single-node intra EP ### 📊 Changes **1 file changed** (+11 additions, -8 deletions) <details> <summary>View changed files</summary> 📝 `csrc/kernels/backend/nccl.cu` (+11 -8) </details> ### 📄 Description ## Summary This PR fixes V2/elastic initialization for pure single-node intra-node EP runs where NCCL reports no Gin capability. (#629) This keeps the existing `EP_DISABLE_GIN` policy: when the variable is set, DeepEP skips Gin context requests instead of forcing Gin availability during preflight initialization. ## Validation Validated on 8 x NVIDIA B300 GPUs, CUDA 13.2, PyTorch `2.13.0.dev20260502+cu130`, and NCCL `2.30.4`. ```bash export EP_DISABLE_GIN=1 python tests/elastic/test_ep.py \ --num-processes 8 \ --num-tokens 8192 \ --hidden 7168 \ --num-topk 8 \ --num-experts 256 \ --allow-hybrid-mode 0 \ --test-first-only ``` Average across 8 ranks: | Metric | Result | README SM100 EP8 reference | | --- | ---: | ---: | | Dispatch | 731.9 GB/s SU | 726 GB/s NVLink | | Combine | 741.0 GB/s SU | 740 GB/s NVLink | --- <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