Fix pixi run test: CUDA 13.3 pin lag + cython-test .so placement#2180
Merged
Conversation
The bindings were regenerated against CUDA 13.3.0 (cc50515), adding NVRTC symbols (NVRTC_ERROR_BUSY, nvrtcBundledHeadersInfo, nvrtcGetBundledHeadersInfo), but the pixi cuda-version pins stayed at 13.2 in cuda_bindings/pixi.toml and cuda_core/pixi.toml. `pixi run test` then built 13.3-referencing Cython code against a 13.2 nvrtc.h and failed with "'nvrtcBundledHeadersInfo' was not declared in this scope". CI was unaffected because it builds wheels from ci/versions.yml (13.3.0) rather than via pixi run test. Bump the cuda-version pins (build-variants + feature.cu13) from 13.2.* to 13.3.* in both packages so the local toolkit matches the regenerated sources and ci/versions.yml. Re-solved pixi.lock files accordingly. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
tests/cython/build_tests.py runs `build_ext --inplace`, which writes the compiled .so relative to the current working directory. pixi runs the build-cython-tests task from the project root, so the .so landed in the package root instead of tests/cython/, where pytest imports it by bare module name. The test only passed previously because a correctly-placed .so from an earlier build persisted (gitignored); a clean checkout fails with ModuleNotFoundError. chdir to the script directory before build_ext --inplace so the .so lands next to its .pyx in both cuda_bindings and cuda_core (kept aligned per NVIDIA#1978). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
rparolin
added
bug
This comment has been minimized.
This comment has been minimized.
This was referenced Jun 9, 2026
Closed
rparolin
added a commit
to rparolin/cuda-python
that referenced
this pull request
Jun 9, 2026
Main CI tests prebuilt wheels and never exercises the pixi source build, so that developer path rots silently on CUDA-pin / generated-source / conda-forge / cython-build drift (NVIDIA#2182, NVIDIA#2183). Add a workflow that runs the pixi source build: - build-smoke (PRs touching the at-risk files): CPU-only. Source-builds bindings + core, imports them, builds the cython test extensions and checks placement. Catches the compile / ABI / .so-placement regressions without a GPU. - full-test (nightly + manual): GPU runner, full `pixi run test`. Shared pixi install factored into a composite action with an explicit, asserted version pin. Relates to NVIDIA#2183 (validate the source-build path over time); the regressions this guards against are NVIDIA#2182, fixed by NVIDIA#2180. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
leofang
approved these changes
Jun 9, 2026
|
rparolin
added a commit
that referenced
this pull request
Jun 9, 2026
…2185) * Fix cuda-version pin lag breaking local pixi run test The bindings were regenerated against CUDA 13.3.0 (cc50515), adding NVRTC symbols (NVRTC_ERROR_BUSY, nvrtcBundledHeadersInfo, nvrtcGetBundledHeadersInfo), but the pixi cuda-version pins stayed at 13.2 in cuda_bindings/pixi.toml and cuda_core/pixi.toml. `pixi run test` then built 13.3-referencing Cython code against a 13.2 nvrtc.h and failed with "'nvrtcBundledHeadersInfo' was not declared in this scope". CI was unaffected because it builds wheels from ci/versions.yml (13.3.0) rather than via pixi run test. Bump the cuda-version pins (build-variants + feature.cu13) from 13.2.* to 13.3.* in both packages so the local toolkit matches the regenerated sources and ci/versions.yml. Re-solved pixi.lock files accordingly. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * Place cython-test .so in tests/cython regardless of cwd tests/cython/build_tests.py runs `build_ext --inplace`, which writes the compiled .so relative to the current working directory. pixi runs the build-cython-tests task from the project root, so the .so landed in the package root instead of tests/cython/, where pytest imports it by bare module name. The test only passed previously because a correctly-placed .so from an earlier build persisted (gitignored); a clean checkout fails with ModuleNotFoundError. chdir to the script directory before build_ext --inplace so the .so lands next to its .pyx in both cuda_bindings and cuda_core (kept aligned per #1978). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * ci: validate pixi run test source-build (PR smoke + nightly GPU) Main CI tests prebuilt wheels and never exercises the pixi source build, so that developer path rots silently on CUDA-pin / generated-source / conda-forge / cython-build drift (#2182, #2183). Add a workflow that runs the pixi source build: - build-smoke (PRs touching the at-risk files): CPU-only. Source-builds bindings + core, imports them, builds the cython test extensions and checks placement. Catches the compile / ABI / .so-placement regressions without a GPU. - full-test (nightly + manual): GPU runner, full `pixi run test`. Shared pixi install factored into a composite action with an explicit, asserted version pin. Relates to #2183 (validate the source-build path over time); the regressions this guards against are #2182, fixed by #2180. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * ci: declare linux-amd64-gpu-l4-latest-1 self-hosted runner label for actionlint actionlint validates static runner labels against its known set; the new full-test job uses a literal GPU label (existing GPU jobs dodge this by building the label from a matrix expression). Declare it so pre-commit's actionlint hook passes. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * ci: checkout full history + tags so setuptools-scm derives the real version A shallow checkout has no tags, so the source-built packages get setuptools-scm's 0.1.dev1 fallback. cuda.core's import-time guard then rejects cuda.bindings ("12.x or 13.x must be installed"). Use fetch-depth: 0 in both jobs so the build resolves the real 13.x version. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * ci: use pinned prefix-dev/setup-pixi instead of curl|bash installer Addresses review (@mdboom): the composite action shelled out to `curl -fsSL https://pixi.sh/install.sh | bash`, an unverified installer (the codecov.io supply-chain failure mode). Replace it with prefix-dev/setup-pixi pinned to a commit SHA (v0.9.6) — its install logic is auditable and pinned — and delete the composite action file. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * ci: switch workflow to pinned prefix-dev/setup-pixi (fixup) The prior commit only removed the composite action file; this commits the workflow change that actually uses prefix-dev/setup-pixi@<sha> in both jobs (and drops the now-unneeded curl from the container apt install). Without this the workflow referenced the deleted ./.github/actions/setup-pixi. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
pixi run testdid not complete from a clean checkout. Two independent root causes, both downstream of CUDA toolkit drift:CUDA version pin lag. The bindings were regenerated against CUDA 13.3.0 (cc50515, "Support CUDA 13.3"), adding NVRTC symbols (
NVRTC_ERROR_BUSY,nvrtcBundledHeadersInfo,nvrtcGetBundledHeadersInfo), but the pixicuda-versionpins stayed at 13.2 incuda_bindings/pixi.tomlandcuda_core/pixi.toml. The local build then compiled 13.3-referencing Cython code against a 13.2nvrtc.h:CI was unaffected because it builds wheels from
ci/versions.yml(13.3.0), not viapixi run test.Cython-test
.soplacement bug (latent since fix(test): unblock root pixi run test workflow #1978).tests/cython/build_tests.pyrunsbuild_ext --inplace, which writes the.sorelative to cwd. pixi runs the task from the project root, so the.solanded in the package root instead oftests/cython/, where pytest imports it by bare module name. fix(test): unblock root pixi run test workflow #1978's verification silently reused a correctly-placed.sofrom an earlier build; a clean checkout fails withModuleNotFoundError.Changes
13.2.* → 13.3.*(build-variants +feature.cu13) in bothpixi.toml, matching the regenerated sources andci/versions.yml; re-solvedpixi.lock×2.os.chdir(script_dir)beforebuild_ext --inplacein bothbuild_tests.py(kept aligned per fix(test): unblock root pixi run test workflow #1978) so the.solands next to its.pyxregardless of cwd.Verification
pixi run test(envcu13) now completes, exit 0:Confirmed the rebuilt
.solands intests/cython/(package root clean), proving the placement fix rather than a leftover artifact.🤖 Generated with Claude Code
Closes #2182. Follow-up: #2183.