[PR #201] [MERGED] remove the dependency of gdrcopy #463

New issue

Closed

opened 2026-10-26 10:01:02 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 10:01:02 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/201
Author: @youkaichao
Created: 6/10/2025
Status: ✅ Merged
Merged: 6/10/2025
Merged by: @sphish

Base: main ← Head: no_gdrcopy

📝 Commits (2)

97be5a3 update the patch
b9b7ce3 update readme

📊 Changes

2 files changed (+65 additions, -70 deletions)

View changed files

📝 third-party/README.md (+28 -70)
📝 third-party/nvshmem.patch (+37 -0)

📄 Description

Right now we use either nvlink or ibgda, gdrcopy is not needed anymore.

I find that modifying the nvshmemt_ibgda_progress function can make nvshmem compile without gdrcopy. This can greatly simplify the process of installing DeepEP. We just need to compile nvshmem, with one additional driver flag options nvidia NVreg_EnableStreamMemOPs=1 NVreg_RegistryDwords="PeerMappingOverride=1;" .

NOTE: theoretically ibgda has a cpu-assisted mode that needs gdrcopy, however, the default mode is "auto", and if we compile without gdrcopy, the cpu-assisted mode will be turned off automatically. therefore, we don't need to do anything special in deepep.

_{🔄 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/201 **Author:** [@youkaichao](https://github.com/youkaichao) **Created:** 6/10/2025 **Status:** ✅ Merged **Merged:** 6/10/2025 **Merged by:** [@sphish](https://github.com/sphish) **Base:** `main` ← **Head:** `no_gdrcopy` --- ### 📝 Commits (2) - [`97be5a3`](https://github.com/deepseek-ai/DeepEP/commit/97be5a38737399c487aa4786f44f7475728ce35a) update the patch - [`b9b7ce3`](https://github.com/deepseek-ai/DeepEP/commit/b9b7ce348bc797e040fa7f7fd7932f91f02b5f8e) update readme ### 📊 Changes **2 files changed** (+65 additions, -70 deletions) <details> <summary>View changed files</summary> 📝 `third-party/README.md` (+28 -70) 📝 `third-party/nvshmem.patch` (+37 -0) </details> ### 📄 Description Right now we use either nvlink or ibgda, gdrcopy is not needed anymore. I find that modifying the `nvshmemt_ibgda_progress` function can make nvshmem compile without gdrcopy. This can greatly simplify the process of installing DeepEP. We just need to compile nvshmem, with one additional driver flag `options nvidia NVreg_EnableStreamMemOPs=1 NVreg_RegistryDwords="PeerMappingOverride=1;"` . NOTE: theoretically ibgda has a cpu-assisted mode that needs gdrcopy, however, the default mode is "auto", and if we compile without gdrcopy, the cpu-assisted mode will be turned off automatically. therefore, we don't need to do anything special in deepep. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 10:01:02 +00:00