Skip to content

Restrict XQA CUDA compilation to SM80 and newer - #32547

Draft
Tianlei Wu (tianleiwu) with Copilot wants to merge 3 commits into
mainfrom
copilot/fix-link-error-cuda-133
Draft

Tianlei Wu (tianleiwu) with Copilot wants to merge 3 commits into
mainfrom
copilot/fix-link-error-cuda-133

Conversation

Copilot AI commented Sep 10, 2026

Copy link
Copy Markdown
Contributor

Description

  • Extract XQA CUDA sources into dedicated object libraries.
  • Filter those targets to SM80+ architectures.
  • Apply identical routing to the standard and plugin CUDA providers.
  • Preserve existing fallback behavior for configurations without SM80+ targets.

Motivation and Context

CUDA 13.3 can emit unresolved XQA kernel_mha references when SM75 and SM80+ architectures are compiled together, causing the final shared-library link to fail. XQA requires SM80+, so excluding its sources from the SM75 compilation pass avoids the inconsistent kernel emission.

@azure-pipelines

Copy link
Copy Markdown
Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

Co-authored-by: tianleiwu <30328909+tianleiwu@users.noreply.github.com>
Copilot AI changed the title [WIP] Fix link error with multiple architectures and CUDA 13.3 Restrict XQA CUDA compilation to SM80 and newer Sep 10, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

[Build] Link error at final link with CUDA 13.3 and multiple architectures.

2 participants