Merge changes for cuda-bindings 13.1.0 release

[rwgk]

• Add support for new APIs introduced in CUDA 13.1

  • https://developer.download.nvidia.com/compute/cuda/redist/redistrib_13.1.0.json

• Prepare for cuda-pathfinder 1.3.3 release

  • Purge support for CTK 11 (those became unsupported with the CTK 13.0 release in August 2025).
  • Remove cudart64_65.dll and cudart64_101.dll from supported_nvidia_libs.py; these outdated DLLs were included accidentally.

[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]

/ok to test

[copy-pr-bot]

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

Contributors can view more details about this message here.

[rwgk]

User prompt

We're at the final stage of making the cuda-bindings v13.1.0 release.

The PR is under merge-for-13.1 here.

CI passes. One teammate approved already.

merge-for-13.1 is in sync with the main branch here.

Could you please make yourself aware of the diff between main and merge-for-13.1, then scan the selected CI logs here for anything suspicious?

~/cuda-python_ci_run_19941271439

This is only for sanity checking.

cuda-python_ci_run_19941271439/
├── build-linux-64.txt  [2025-12-04 12:34]  (1,321,111 bytes)
├── build-linux-aarch64.txt  [2025-12-04 12:34]  (1,320,103 bytes)
├── build-win-64.txt  [2025-12-04 12:35]  (2,194,293 bytes)
├── test-linux-64.txt  [2025-12-04 12:31]  (642,109 bytes)
├── test-linux-aarch64.txt  [2025-12-04 12:32]  (644,421 bytes)
└── test-win-64.txt  [2025-12-04 12:30]  (498,781 bytes)
    ____________________________________________________________
    Number of subdirs: 0
    Sum of file sizes: (6,620,818 bytes)

Assistant response (Diff sanity check)

Diff sanity check (main ↔ merge-for-13.1)

• Scope of changes:
  • CI/config: ci/test-matrix.yml adds coverage for CUDA 13.1.0 across Linux/Windows, x86_64/arm64, multiple Python versions; ci/versions.yml bumps the CUDA build version from 13.0.2 → 13.1.0.
  • Bindings: Large updates in cuda_bindings/cuda/bindings/* (driver/runtime/nvrtc/cufile/etc.) to track CTK 13.1.0, including new APIs (e.g. multicast *_v2 variants) and adjusted types/flags; _version.py updated accordingly.
  • Docs/release notes: New/renamed release notes files for 12.9.5 and 13.1.0, and module docs (driver.rst, runtime.rst, nvrtc.rst) updated for new APIs and behavior.
  • Support matrix pruning: supported_nvidia_libs.py and related pathfinder data drop older 11.x-era libs and some very old DLLs, keeping 12/13-era sonames and DLLs.
  • cuda-pathfinder: Version bumped 1.3.3a0 → 1.3.3, with new release notes and an extra cu13 dependency (nvidia-cufftmp-cu13), plus updated nv-versions JSON.
  • Minor infra: .gitattributes extended (e.g. *.cmd EOL, *.svg binary), .spdx-ignore extended for a Draw.io SVG.

Nothing in the diff itself looks out-of-character for a CTK 13.1.0 cut: it’s the expected combination of new APIs, doc updates, CI matrix expansion, and support-matrix cleanup.

CI logs: build jobs

From build-linux-64.txt, build-linux-aarch64.txt, build-win-64.txt:

• Errors / failures:

  • Grep for FAILED, ERROR, Traceback, E across all build logs returned no matches indicative of build failures or exceptions.
  • The only hits on that pattern were:
    • except OSError: around a symlink helper (which then falls back to shutil.copy2), and
    • C/C++ compiler warnings in generated cufile.cpp (see below), not errors.

• Notable warnings (non-fatal, but worth being aware of):

  • cufile stream type warnings (Linux x86_64 & aarch64):
    • Several warning: invalid conversion from 'void*' to 'CUstream' in generated cuda/bindings/cufile.cpp for:
      • cuFileReadAsync
      • cuFileWriteAsync
      • cuFileStreamRegister
      • cuFileStreamDeregister
    • These are compile-time warnings; builds still complete successfully and produce the .so modules.
    • Semantically, they reflect passing a void*-typed stream handle to functions expecting CUstream; if that’s intentional in your Cython layer, this is noisy but benign. If not, it might be something to tighten later, but it’s not blocking CI.
  • Windows loader warnings in build-win-64.txt:
    • MSVC C4996 for wcstombs and strncat in cuda/bindings/_bindings/loader.cpp (“unsafe” CRT functions). Typical for older-style C APIs; compilation still succeeds.
    • C4551 warnings in cynvrtc.cpp about “function call missing argument list” — also warnings only; module still links and is used by tests.
  • Setuptools packaging warnings:
    • _Warning: Package 'cuda.bindings' is absent from the packages configuration (and similarly for cuda.bindings._bindings, _internal, _lib) during wheel build on Windows. This pattern is consistent with using ext_modules and non-standard package layouts; since the build completes and tests run against installed wheels, these appear to be expected noise in your existing configuration.
  • Node.js deprecation warnings:
    • Several [DEP0040] DeprecationWarning: The 'punycode' module is deprecated messages, coming from GitHub Actions tooling, not from your project.

Nothing in the build logs suggests incomplete artifacts, missing symbols, or link errors for the new 13.1.0 APIs.

CI logs: test jobs

From test-linux-64.txt, test-linux-aarch64.txt, test-win-64.txt:

• Errors / failures:

  • Grep for FAILED, ERROR, Traceback, and typical pytest failure-summary regex across all three test logs found:
    • No failures or errors in the two Linux test logs.
    • In test-win-64.txt, only a clean pytest summary with skips:
      • 190 passed, 4 skipped in 5.49s
      • Skips were:
        • tests/test_cufile.py: could not import cuda.bindings.cufile (no module in that environment).
        • tests/test_cudart.py: “cudart dynamic lib not available”.
        • tests/test_graphics_apis.py: GL context/texture could not be created (AttributeError: opengl32).
      • These match the usual pattern of environment-dependent tests (file I/O libs, cudart dynlib, OpenGL) being skipped when prerequisites aren’t present, not new regressions.

• Warnings:

  • No additional suspicious warnings (beyond the Node deprecations already mentioned for the build jobs) surfaced in the test logs.
  • Pytest’s short-test-summary is clean other than the above skips; no failed/errors entries.

Overall sanity verdict

• Diff review: Changes between main and merge-for-13.1 align with a normal CTK bump and binding refresh (new 13.1.0 APIs, CI matrix updates, docs and version bumps, and cleanup of old libs); nothing stands out as an obviously risky, off-topic, or accidental change.
• CI logs:
  • No build or test failures appear in the captured logs.
  • The only noteworthy items are:
    • Expected compiler warnings in generated code (cufile stream casts, MSVC “unsafe” CRT calls).
    • Standard packaging and Node.js deprecation warnings.
    • Expected environment-related pytest skips on Windows.
• Sanity conclusion: From the perspective of these diffs and the provided CI logs, merge-for-13.1 looks clean for the cuda-bindings v13.1.0 release, with no suspicious red flags that would block merging.

[leofang]

Let one question on release note / version selector, otherwise LGTM!

[github-actions]

Doc Preview CI
Preview removed because the pull request was closed or merged.