[PR #151] [MERGED] Feat: enhance nvidia peer memory detection #443

New issue

Closed

opened 2026-10-26 10:00:48 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 10:00:48 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/151
Author: @vicoooo26
Created: 5/9/2025
Status: ✅ Merged
Merged: 5/12/2025
Merged by: @sphish

Base: main ← Head: feat/nvidia-peer-mem-detection

📝 Commits (1)

c6051f3 Feat: enhance nvidia peer memory detection

📊 Changes

1 file changed (+14 additions, -0 deletions)

View changed files

📝 third-party/nvshmem.patch (+14 -0)

📄 Description

In my H800 environment which NVIDIA Driver is 550.54.15 and nvidia_peermem module has been loaded successfully, but NVSHMEM detections of nvidia_peermem existence never pass, then I will get some error messages like neither nv_peer_mem, or nvidia_peermem detected. Skipping transport.. That's because there is not neither /sys/kernel/mm/memory_peers/nv_mem/version nor /sys/kernel/mm/memory_peers/nv_mem_nc/version file in my machine.
According to NCCL's implementation I use /sys/module/nvidia_peermem/version file as an indication of the existence of nvidia_peermem, then I can run DeepDP internode、intronode and low_latency over multiple nodes successfully after patching it.

_{🔄 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/151 **Author:** [@vicoooo26](https://github.com/vicoooo26) **Created:** 5/9/2025 **Status:** ✅ Merged **Merged:** 5/12/2025 **Merged by:** [@sphish](https://github.com/sphish) **Base:** `main` ← **Head:** `feat/nvidia-peer-mem-detection` --- ### 📝 Commits (1) - [`c6051f3`](https://github.com/deepseek-ai/DeepEP/commit/c6051f388003f8d9fa60b43fced1475ad5d099af) Feat: enhance nvidia peer memory detection ### 📊 Changes **1 file changed** (+14 additions, -0 deletions) <details> <summary>View changed files</summary> 📝 `third-party/nvshmem.patch` (+14 -0) </details> ### 📄 Description In my H800 environment which NVIDIA Driver is `550.54.15` and `nvidia_peermem` module has been loaded successfully, but NVSHMEM detections of `nvidia_peermem` existence never pass, then I will get some error messages like `neither nv_peer_mem, or nvidia_peermem detected. Skipping transport.`. That's because there is not neither `/sys/kernel/mm/memory_peers/nv_mem/version` nor `/sys/kernel/mm/memory_peers/nv_mem_nc/version` file in my machine. According to [NCCL's implementation](https://github.com/NVIDIA/nccl/blob/3000e3c797b4b236221188c07aa09c1f3a0170d4/src/transport/net_ib.cc#L711) I use `/sys/module/nvidia_peermem/version` file as an indication of the existence of `nvidia_peermem`, then I can run DeepDP `internode`、`intronode` and `low_latency` over multiple nodes successfully after patching it. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 10:00:48 +00:00