[PR #765] [CLOSED] [Bugfix][NVSHMEM] Support v2 state in legacy IBGDA kernels #766

New issue

Closed

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

bsakaguchi commented

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

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/765
Author: @XFDG
Created: 9/25/2026
Status: ❌ Closed

Base: main ← Head: fix/deepep-764-nvshmem-v2

📝 Commits (1)

4dbea47 [Bugfix][NVSHMEM] Support v2 state in legacy IBGDA kernels

📊 Changes

2 files changed (+22 additions, -3 deletions)

View changed files

📝 csrc/kernels/legacy/ibgda_device.cuh (+15 -3)
📝 deep_ep/include/deep_ep/common/compiled.cuh (+7 -0)

📄 Description

Summary

Port NVSHMEM v2 IBGDA RC indexing support into DeepEP's legacy kernels.

Fix

Preserve the existing RC mapping for NVSHMEM v1 state.
Use the v2 PE/NPES indexing scheme for the newer state layout.
Define __CUDACC_RDC__ when NVSHMEM is enabled so device symbols use the correct extern declarations.

Testing

Full SM100 extension build, NVSHMEM device-link, and shared-library link passed with CUDA 13.1, NVSHMEM 3.5.21, and isolated NCCL 2.32.3.
csrc/kernels/legacy/internode.cu and internode_ll.cu also compiled for SM100 with NVSHMEM 3.4.5.
An extension import smoke test failed in the available Torch environment with ImportError: unsupported scalarType; therefore runtime behavior is not claimed as verified.
Cross-node IBGDA runtime testing was not available on the single 8-GPU node.

Fixes #764

_{🔄 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/765 **Author:** [@XFDG](https://github.com/XFDG) **Created:** 9/25/2026 **Status:** ❌ Closed **Base:** `main` ← **Head:** `fix/deepep-764-nvshmem-v2` --- ### 📝 Commits (1) - [`4dbea47`](https://github.com/deepseek-ai/DeepEP/commit/4dbea471a5c703fe8affbecce39d420543ecdf7d) [Bugfix][NVSHMEM] Support v2 state in legacy IBGDA kernels ### 📊 Changes **2 files changed** (+22 additions, -3 deletions) <details> <summary>View changed files</summary> 📝 `csrc/kernels/legacy/ibgda_device.cuh` (+15 -3) 📝 `deep_ep/include/deep_ep/common/compiled.cuh` (+7 -0) </details> ### 📄 Description ## Summary Port NVSHMEM v2 IBGDA RC indexing support into DeepEP's legacy kernels. ## Fix - Preserve the existing RC mapping for NVSHMEM v1 state. - Use the v2 PE/NPES indexing scheme for the newer state layout. - Define `__CUDACC_RDC__` when NVSHMEM is enabled so device symbols use the correct extern declarations. ## Testing - Full SM100 extension build, NVSHMEM device-link, and shared-library link passed with CUDA 13.1, NVSHMEM 3.5.21, and isolated NCCL 2.32.3. - `csrc/kernels/legacy/internode.cu` and `internode_ll.cu` also compiled for SM100 with NVSHMEM 3.4.5. - An extension import smoke test failed in the available Torch environment with `ImportError: unsupported scalarType`; therefore runtime behavior is not claimed as verified. - Cross-node IBGDA runtime testing was not available on the single 8-GPU node. Fixes #764 --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

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