[PR #761] [CLOSED] docs: document mlx5 requirement for legacy IBGDA #763

New issue

Closed

opened 2026-10-27 21:06:02 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-27 21:06:02 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/761
Author: @dvd233
Created: 9/22/2026
Status: ❌ Closed

Base: main ← Head: fix/docs-mlx5-ibgda-759

📝 Commits (1)

6188bb4 docs: document mlx5 requirement for legacy IBGDA

📊 Changes

1 file changed (+17 additions, -0 deletions)

View changed files

📝 docs/nvshmem.md (+17 -0)

📄 Description

Summary

Clarify the RDMA adapter prerequisites for DeepEP V1's NVSHMEM-backed internode and low-latency paths.

Why

DeepEP V1 enables NVSHMEM_IB_ENABLE_IBGDA=1 for these paths. Issue #759 reports that on Alibaba Cloud eRDMA (erdma_*), NVSHMEM cannot load libmlx5dv; the log says that directNIC features are disabled, but the low-latency path is not usable. The existing guide only says “RDMA devices” and “IBGDA”, which can make this look like a recoverable warning.

Change

Document the Mellanox/CX-4+ and mlx5/OFED requirements for the NVSHMEM IBGDA path.
Call out non-mlx5 eRDMA as unsupported for V1 internode and low-latency examples.
Point readers to the V2 NCCL Gin path without claiming it supports every NIC.

Validation

git diff --check
Checked the ../README.md relative link.
Documentation-only change; no runtime code changed.

Fixes #759

_{🔄 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/761 **Author:** [@dvd233](https://github.com/dvd233) **Created:** 9/22/2026 **Status:** ❌ Closed **Base:** `main` ← **Head:** `fix/docs-mlx5-ibgda-759` --- ### 📝 Commits (1) - [`6188bb4`](https://github.com/deepseek-ai/DeepEP/commit/6188bb474bcee928dac99bef0e912392289b5690) docs: document mlx5 requirement for legacy IBGDA ### 📊 Changes **1 file changed** (+17 additions, -0 deletions) <details> <summary>View changed files</summary> 📝 `docs/nvshmem.md` (+17 -0) </details> ### 📄 Description ## Summary Clarify the RDMA adapter prerequisites for DeepEP V1's NVSHMEM-backed internode and low-latency paths. ## Why DeepEP V1 enables `NVSHMEM_IB_ENABLE_IBGDA=1` for these paths. Issue #759 reports that on Alibaba Cloud eRDMA (`erdma_*`), NVSHMEM cannot load `libmlx5dv`; the log says that directNIC features are disabled, but the low-latency path is not usable. The existing guide only says “RDMA devices” and “IBGDA”, which can make this look like a recoverable warning. ## Change - Document the Mellanox/CX-4+ and mlx5/OFED requirements for the NVSHMEM IBGDA path. - Call out non-`mlx5` eRDMA as unsupported for V1 internode and low-latency examples. - Point readers to the V2 NCCL Gin path without claiming it supports every NIC. ## Validation - `git diff --check` - Checked the `../README.md` relative link. - Documentation-only change; no runtime code changed. Fixes #759 --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-27 21:06:02 +00:00