[PR #775] fix(build): Discover toolkit NVTX libraries and CCCL headers #774

New issue

Open

opened 2026-10-27 21:06:13 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-27 21:06:13 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/775
Author: @Chlience
Created: 10/9/2026
Status: 🔄 Open

Base: hybrid-ep ← Head: fix/cuda-toolkit-build-compat

📝 Commits (1)

d1c05b4 fix(build): Discover toolkit NVTX libraries and CCCL headers

📊 Changes

2 files changed (+116 additions, -7 deletions)

View changed files

📝 setup.py (+26 -7)
➕ tests/test_setup_cuda_toolkit.py (+90 -0)

📄 Description

The hybrid-ep build hardcodes nvtx3interop and omits toolkit CCCL headers from the DeepEP extension. CUDA toolkits providing only legacy NVTX cannot satisfy that fixed library requirement.

Prefer available nvtx3interop, fall back to installed nvToolsExt, and discover existing CCCL headers from CUDA_HOME. Preserve the existing linker requirement when neither library is found. Target hybrid-ep, whose setup still contains these paths; main uses a different build layout. Kernels, JIT and NVSHMEM settings are unchanged.

Twelve CPU extension-configuration checks and the existing C++-standard regression passed. On the latest hybrid-ep base, the ten-case library/header subset has five expected failures. isort, new-test Black and staged whitespace checks passed. NVCC compilation, HybridEP JIT and device communication were not run.

Related integration: https://github.com/flagos-ai/Megatron-LM-FL/pull/104. This PR owns the implementation and its regressions; the integration tracks merge/version dependencies.

This carries the build compatibility part of a downstream CUDA 12.8 patch, revised to preserve newer toolkit behavior. It does not claim complete CUDA 12.8 support for HybridEP JIT or replace an external CCCL header setup.

_{🔄 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/775 **Author:** [@Chlience](https://github.com/Chlience) **Created:** 10/9/2026 **Status:** 🔄 Open **Base:** `hybrid-ep` ← **Head:** `fix/cuda-toolkit-build-compat` --- ### 📝 Commits (1) - [`d1c05b4`](https://github.com/deepseek-ai/DeepEP/commit/d1c05b412a5f87d4865819bba10db2a5057cd0cb) fix(build): Discover toolkit NVTX libraries and CCCL headers ### 📊 Changes **2 files changed** (+116 additions, -7 deletions) <details> <summary>View changed files</summary> 📝 `setup.py` (+26 -7) ➕ `tests/test_setup_cuda_toolkit.py` (+90 -0) </details> ### 📄 Description The hybrid-ep build hardcodes nvtx3interop and omits toolkit CCCL headers from the DeepEP extension. CUDA toolkits providing only legacy NVTX cannot satisfy that fixed library requirement. Prefer available nvtx3interop, fall back to installed nvToolsExt, and discover existing CCCL headers from CUDA_HOME. Preserve the existing linker requirement when neither library is found. Target hybrid-ep, whose setup still contains these paths; main uses a different build layout. Kernels, JIT and NVSHMEM settings are unchanged. Twelve CPU extension-configuration checks and the existing C++-standard regression passed. On the latest hybrid-ep base, the ten-case library/header subset has five expected failures. isort, new-test Black and staged whitespace checks passed. NVCC compilation, HybridEP JIT and device communication were not run. Related integration: https://github.com/flagos-ai/Megatron-LM-FL/pull/104. This PR owns the implementation and its regressions; the integration tracks merge/version dependencies. This carries the build compatibility part of a downstream CUDA 12.8 patch, revised to preserve newer toolkit behavior. It does not claim complete CUDA 12.8 support for HybridEP JIT or replace an external CCCL header setup. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi added the

pull-request

label

2026-10-27 21:06:13 +00:00

No labels

pull-request

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

deepseek-ai/DeepEP#774

No description provided.

Rows
Columns