Prepare cuda-bindings 12.9.8 release

[mdboom]

Regenerate files and add release notes for cuda-bindings 12.9.8

[rwgk]

I'm still looking with codex as well.

[rwgk]

Suggested change

	A new version of the ``nvrtc`` API is available as ``cuda.bindings._v2.nvrtc``. The
	This release introduces an experimental, alternative ``cuda.bindings._v2.nvrtc`` API. The

The main idea is to insert the word "experimental" early.

[rwgk]

Correction after checking the branch and the PR-built wheel: it seems this whole section should be removed. _v2.nvrtc is a 13.4-only preview API and is not packaged by 12.9.x.

[leofang]

xref: #2836

Suggested change

	Prerelease feature
	------------------
	Preview feature
	---------------

[leofang]

The formatting here seems off? PR XXXX comes before the bullet point

[leofang]

also it doesn't seem that the style is consistent with past release notes (we don't link to PRs), although this is just a minor nit

[mdboom]

The PR is at the end of each bullet point.

Precedent for adding PR numbers is all throughout the 13.3 release notes. https://raw.githubusercontent.com/NVIDIA/cuda-python/refs/heads/main/cuda_bindings/docs/source/release/13.3.0-notes.rst

[leofang]

Q: We probably should also update cuda_python/docs in this branch?

[rwgk]

The cibuildwheel 4.2.0 bump means the cp315/cp315t wheels are now built with CPython 3.15.0rc1, but the post-build/test environments still force 3.15.0-beta.2 in:

• .github/workflows/build-wheel.yml:233
• .github/workflows/test-wheel-linux.yml:220
• .github/workflows/test-wheel-windows.yml:210

This crosses a prerelease ABI change and is now causing all six 3.15/3.15t wheel-test jobs to crash during extension import—segmentation faults on Linux and access violations on Windows. Could we update or remove these three beta.2 overrides so the build and test interpreters match?

[rwgk]

This appears to have been based on the 13.4.1 notes rather than the changes actually shipped on 12.9.x. The #2581, #2551, and #2417 changes and the _v2.nvrtc preview API (#2398) are not present on this branch; #2690 and #2530 are present. Conversely, actual changes since v12.9.7 such as #2145 and #2382 are not mentioned. Could we rebuild these notes from the v12.9.7..12.9.x history and remove the _v2 section?

Also, I think this new page needs to be linked from cuda_bindings/docs/source/release.rst.

[mdboom]

Also, I think this new page needs to be linked from cuda_bindings/docs/source/release.rst.

The docs are not actually shipped from this branch, so that file hasn't been updated in a long time. I instead just think we put the new release notes for 12.9.8 here (with the corrections you outlined) and then file another PR to copy it to main.

[rwgk]

Fresh verdict: the workflow blocker is fixed, but I’d hold approval for one more release-note correction.

The fresh PR head is 8fe688e4287dbb26d5e5a3d95e70113010313bca, matching local HEAD.

Remaining substantive finding

The Bugfixes section still differs from the actual release delta in three ways:

1. It accidentally dropped the valid cuda.bindings: Fix API status handling #2530 and nvbug 6602863: system_event_set_wait fails on events: resize() on non-owning SystemEventData_v1._data view #2690 entries.

   • #2530 is present through 12.9.x: Update bindings to latest generator release #2604.
   • #2690 is present in this PR’s NVML refresh and was removed by the latest notes rewrite.

2. The #2382 entry says runtime, but that change affected NVVM, not the runtime bindings. This should be runtime → NVVM.

3. The #2206 entries overstate the delta:

   • The exception-declaration bullet should be removed; v12.9.7 already used the except ?CUDA_ERROR_NOT_FOUND driver signature.
   • FunctionNotFoundError became new for missing driver functions only. NVRTC already raised it in v12.9.7.

I’d post those as one consolidated follow-up on your existing release-note thread:

The rewritten Bugfixes section still has three mismatches with v12.9.7..HEAD:

• Please restore cuda.bindings: Fix API status handling #2530 and nvbug 6602863: system_event_set_wait fails on events: resize() on non-owning SystemEventData_v1._data view #2690; both fixes are present, and nvbug 6602863: system_event_set_wait fails on events: resize() on non-owning SystemEventData_v1._data view #2690 was dropped by the latest rewrite.
• For Fix potential data race in initialization (12.9.x backport) #2382, runtime should be NVVM.
• The first Backport #1972 to 12.9.x #2206 bullet does not describe a new release delta and should be removed. FunctionNotFoundError became new for missing driver functions only; NVRTC already used it in v12.9.7, so the second bullet should say only driver.

Could we correct those before merging?

Nonblocking cleanup

• matrix.python-verison is misspelled, so the free-threaded verification step is skipped even for 3.15t. This predates the PR; I’d treat it as an optional inline cleanup.
• cuda_core should be cuda.core or cuda-core. I wouldn’t post this separately.

Non-generated files reviewed

• build-wheel.yml — beta.2 pin fixed; NumPy workaround removal works; optional typo above.
• test-wheel-linux.yml — no remaining finding.
• test-wheel-windows.yml — no remaining finding.
• 12.9.8-notes.rst — findings above.

The _v2 section and false #2581/#2551/#2417 claims are gone. I would not repeat the release.rst point given Michael’s explanation.

Of 32 changed files, 28 matched the generated-file marker and were excluded. git diff --check is clean. At my final CI refresh, 60 checks passed, one unrelated Windows 3.11 job remained in progress, and none had failed; importantly, all twelve Python 3.15/3.15t build and test jobs now pass.

I did not post any comments.

[mdboom]

The only actionable things from that review were putting #2530 and #2690 back in.