feat: add Blackwell GPU (sm_120) CUDA support

[mvanhorn]

Adds TORCH_CUDA_ARCH_LIST to the CUDA build step in release.yml to include 12.0+PTX for Blackwell GPU forward compatibility. 5 reports (#386 #395 #396 #399 #400) from RTX 50-series users hitting 'no kernel image' because pre-built PyTorch cu128 doesn't include sm_120. Fixes #386. This contribution was developed with AI assistance (Claude Code).

Summary by CodeRabbit

• Chores
  • Updated build configuration to support additional NVIDIA GPU architectures, expanding GPU compatibility and compatibility with newer GPU models.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 34205f6d-9ff6-4b78-94ce-4b62e53a6ac2

📥 Commits

Reviewing files that changed from the base of the PR and between 75abbb0 and f9835d2.

📒 Files selected for processing (1)

• .github/workflows/release.yml

---

📝 Walkthrough

Walkthrough

The GitHub Actions release workflow for Windows now explicitly specifies CUDA GPU architectures (8.0;8.6;8.9;9.0;12.0+PTX) via the TORCH_CUDA_ARCH_LIST environment variable during the CUDA build step, enabling Blackwell GPU support through PTX forward compatibility.

Changes

Cohort / File(s)	Summary
CUDA Architecture Configuration .github/workflows/release.yml	Added TORCH_CUDA_ARCH_LIST environment variable to Windows CUDA build step, explicitly targeting compute capabilities 8.0, 8.6, 8.9, 9.0, and 12.0+PTX (Blackwell).

Estimated code review effort

🎯 1 (Trivial) | ⏱️ ~3 minutes

Possibly related issues

• #362: The change directly addresses the missing sm_120 kernel image by explicitly targeting Blackwell GPU architecture (12.0+PTX) during Windows CUDA compilation.

Poem

🐰 A rabbit hops through cuda streams,
Adding Blackwell to our dreams,
Twelve-point-oh with PTX in sight,
Windows builds now compile just right! ✨

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly summarizes the main change: adding Blackwell GPU (sm_120) CUDA support via the TORCH_CUDA_ARCH_LIST environment variable in the release workflow.
Linked Issues check	✅ Passed	The PR successfully addresses the linked issue #386 by adding Blackwell GPU (sm_120) support through the TORCH_CUDA_ARCH_LIST environment variable (12.0+PTX) in the CUDA build configuration.
Out of Scope Changes check	✅ Passed	The PR contains only the necessary change to the release.yml workflow for adding Blackwell CUDA support; no unrelated or out-of-scope modifications are present.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}