CI doesn't test the latest release of cuda_core against cuda_bindings main

[mdboom]

This means it's possible to find ways in which cuda_bindings breaks when installed alongside the latest release of cuda_core.

We should add this to our CI to guard against it.

[atiaomar1978-hub]

Hi @mdboom — looking at this today, it appears the “released cuda_core × main cuda_bindings” quadrant may already be covered by the nightly-cuda-core mode added in #2283:

• ci-nightly.yml jobs test-cuda-core-linux-64 / test-cuda-core-windows
• matrix rows with ENV.MODE: nightly-cuda-core in ci/test-matrix.yml
• install path in test-wheel-{linux,windows}.yml: main pathfinder/bindings + released cuda-core from PyPI, then the released tag’s test suite

That runs on the nightly schedule, but not as part of PR CI (ci.yml).

Could you clarify the intended scope for this issue?

1. Close as addressed by CI: nightly-cuda-core + nightly-numba-cuda-mlir + PTDS fix #2283 (nightly coverage is enough), or
2. Additionally wire the same (or a reduced) nightly-cuda-core check into PR CI when cuda_bindings / related paths change?

Happy to help with (2) if that’s what’s needed.

[rparolin]

@leofang for vis. Thoughts on this?

[mdboom]

Thanks for confirming it's covered in the nightlies, @atiaomar1978-hub. IMHO, this is an important enough case that we should cover it in at least one PR-running configuration. EDIT: But we should get @leofang's thoughts before proceeding with an implementation.

[atiaomar1978-hub]

Thanks @mdboom — that makes sense. Nightly coverage is good, but a PR-running check would catch regressions before merge.

I'm happy to work on this once @leofang weighs in. My thinking is to reuse the existing nightly-cuda-core path (released cuda-core from PyPI + main pathfinder/bindings) and wire a reduced matrix into PR CI when cuda_bindings / related paths change — similar to what #2283 added for nightlies.

Happy to draft a PR once direction is confirmed.

[rparolin]

Hi @atiaomar1978-hub. Thanks for offering to tackle this issue. I've assigned you as the owner of this ticket.
Let us know if you have any questions.

[atiaomar1978-hub]

Opened #2638 to wire the existing nightly-cuda-core test mode into PR CI when bindings-related paths change.

What it does

• Adds test-cuda-core-linux-64 to ci.yml (linux-only, matching the "at least one PR-running configuration" ask)
• Reuses released cuda-core from PyPI + PR-built pathfinder/bindings wheels from the current run
• Gates on detect-changes.outputs.test_bindings; updates the checks aggregator accordingly
• Adds guardrail tests in ci/tools/tests/test_ci_pr_cuda_core_compat.py

Local verification

• pytest -v --noconftest ci/tools/tests/test_ci_pr_cuda_core_compat.py — 6 passed

No RunPod needed — GPU tests run on existing self-hosted CI runners.

[atiaomar1978-hub]

Update on #2638:

• Addressed Bugbot/Grok review: kept pathfinder/bindings wheel downloads enabled (only test-core: false / test-python: false) so nightly-cuda-core can install PR-built wheels
• Fixed pre-commit.ci ruff format failure on ci/tools/tests/test_ci_pr_cuda_core_compat.py (run https://results.pre-commit.ci/run/github/381173759/1786751697.7FNpRYiJSZCR8Xfq95KQYw)

Local guardrail tests still pass (pytest -v --noconftest ci/tools/tests/test_ci_pr_cuda_core_compat.py).