fix(ci): fail cleanly on a malformed cuda.build.version instead of tracebacking

[LeSingh1]

Problem

ci/tools/check_pixi_cuda_version.py returns a deliberate diagnostic exit code for every problem it anticipates — missing versions.yml, missing cuda.build.version, missing pixi.toml, missing feature key — and then unpacks the version with no checking at all:

try:
    build_version = yaml.safe_load(...)["cuda"]["build"]["version"]
except (KeyError, TypeError):
    print(f"error: cuda.build.version not found in {VERSIONS_FILE_PATH}", file=sys.stderr)
    return 2

major, minor, *_ = build_version.split(".")

YAML makes that easy to break. ci/versions.yml quotes the value today:

cuda:
  build:
    version: "13.3.0"

Drop the quotes and a two-component value loads as a float, a bare number as an int — neither has .split. And a quoted-but-single-component value survives to the tuple unpacking and dies there instead.

Driving main() against the real pixi.toml files with only ci/versions.yml varied:

version: value	loaded as	result on main
"13.3.0"	str	exit 0 (correct)
13.3	float	AttributeError: 'float' object has no attribute 'split'
13	int	AttributeError: 'int' object has no attribute 'split'
(empty)	None	AttributeError: 'NoneType' object has no attribute 'split'
[13, 3]	list	AttributeError: 'list' object has no attribute 'split'
"13"	str	ValueError: not enough values to unpack (expected at least 2, got 1)

All five escape as an uncaught traceback out of a pre-commit hook, pointing the contributor at this script rather than at the line they edited. The except (KeyError, TypeError) above shows the intent was already to fail cleanly here.

Fix

Add parse_build_version, which returns (major, minor) only for a <major>.<minor>[.<patch>] string of digits and None otherwise, and have main report the bad value with the same return 2 shape as its neighbours:

error: cuda.build.version=13.3 in .../ci/versions.yml is not a '<major>.<minor>[.<patch>]'
version string. Quote the value in YAML so it is not loaded as a number (13.3 becomes a
float, 13 becomes an int).

The message names the quoting trap because that is how the value actually goes wrong.

No behavior change for well-formed input: the current "13.3.0" still yields expected="13.3.*" and cuda_feature="cu13", and running the tool against the real repo still exits 0.

Tests

This script had no tests. Added ci/tools/tests/test_check_pixi_cuda_version.py:

• parse_build_version over the accepted shapes (13.3.0, 12.9.1, 13.3, 13.3.0.1) and every rejected one (float, int, None, list, "13", "", "13.", ".3", "cuda.13").
• End-to-end tests that build a temporary repo layout (ci/versions.yml + both pixi.toml files), monkeypatch the module's path constants, and assert main() returns 2 with the diagnostic — and returns 0 for a well-formed version.

The module imports PyYAML at module scope (the pre-commit hook supplies it via additional_dependencies), so the test module opens with pytest.importorskip("yaml") rather than failing collection where PyYAML is absent.

Verification

Executed in full (pure Python, no GPU):

• The six-row table above was produced by driving the upstream/main main() directly with monkeypatched path constants; five of the six raise uncaught.
• With the fix, all of those return exit 2 with the new message, and pytest ci/tools/tests is 73 passed.
• Running the tool against the real repo checkout still prints OK: ... and exits 0.
• Note on the usual "restore the file from upstream/main and watch the tests fail" check: because the fix introduces a new function, the test module cannot even import against main (ImportError: cannot import name 'parse_build_version'). The evidence that matters is the table above, taken against unmodified upstream/main code.
• ruff check / ruff format --check clean on both files; toolshed/check_spdx.py clean on the new file; python -m py_compile clean. Index verified clean before committing.

Related

Independent of #2538, #2543 (both check_mempool_hygiene.py) and #2496 (check_release_notes.py) — different files, no overlap.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.