Exclude broken cuda-toolkit wheels on Windows

[rwgk]

Summary

• exclude cuda-toolkit 12.9.2 and 13.0.3 only on Windows in cuda_pathfinder, cuda_core, and cuda_bindings
• keep the existing floating 12.* and 13.* behavior on Linux, since the Linux matrix did not show the same breakage
• preserve flexibility to pick up later good patch releases while avoiding the two known-bad Windows resolutions

Background

During CI for PR #1817, a group of Windows test jobs started failing in cuda.pathfinder strict mode after new cuda-toolkit patch releases became available on the package index.

Compared with the last successful CI run against main, the affected Windows jobs changed from resolving:

• cuda-toolkit 12.9.1 to 12.9.2
• cuda-toolkit 13.0.2 to 13.0.3

Those failing Windows jobs then stopped installing the full set of CTK DLL-providing packages such as cublas, cufft, curand, cusolver, cusparse, npp, and nvjpeg, and cuda.pathfinder strict checks failed as a result.

The strongest current hypothesis is that these two newly published cuda-toolkit patch releases have broken Windows dependency metadata for extras resolution. Linux did not show the same regression.

What This PR Changes

• cuda_pathfinder

  • split the cu12 and cu13 toolkit requirements by platform
  • exclude 12.9.2 and 13.0.3 only when sys_platform == 'win32'
  • leave non-Windows resolution on the original 12.* / 13.* ranges

• cuda_core

  • apply the same Windows-only exclusions to test-cu12, test-cu13, test-cu12-ft, and test-cu13-ft

• cuda_bindings

  • apply the same Windows-only exclusion to the all extra for the toolkit packages it installs

Why Windows-Only

The exclusion-based workaround was first applied uniformly (Linux & Windows) under #1817. That is a significantly simpler change, but strictly speaking the observed breakage is Windows-specific:

• the failing cuda.pathfinder jobs were Windows jobs
• Linux jobs using floating 12.* / 13.* specs continued to pass

Restricting the exclusions to Windows makes the workaround narrower and keeps Linux on the normal floating-major behavior.

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-1884/
https://nvidia.github.io/cuda-python/pr-preview/pr-1884/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-1884/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-1884/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[mdboom]

This looks good to me if the expectation is that 13.0.3 and 12.9.2 are "permanently broken" and upstream will fix this by making a 13.0.4 and 12.9.3 release.

I also have no objection to the simpler approach that would also apply to Linux, especially since I expect the CTK team would always make a new release for all platforms simultaneously.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[rwgk]

Back to draft mode: it looks like we will not need this.

[rwgk]

Closing after seeing that cuda-toolkit 12.9.2 and 13.0.3 were yanked on PyPI:

https://pypi.org/project/cuda-toolkit/#history