Add Python build/pytest CI workflow on colossus runner#167
Conversation
Mirrors the cpp-build-test workflow's container + runner setup, then builds the Python package via pip + scikit-build and runs the pytest suite (excluding the `long` substructure integration tests, matching pip-build.yml's smoke job). Single matrix entry pinned to py3.13 / rdkit 2025.9.2, the latest CUDA 12.x pair tracked in conda-build.yml.
|
| Filename | Overview |
|---|---|
| .github/workflows/python-build-test.yml | New CI workflow that source-builds nvmolkit and runs pytest on a colossus V100 runner; structure mirrors the existing C++ workflow with minor inconsistencies in shell flags. |
Reviews (3): Last reviewed commit: "Drop CI matrix back to py3.12 for V100 r..." | Re-trigger Greptile
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
PyPI torch wheels from 2.5+ dropped sm_70 from their default arch list, which produces cudaErrorNoKernelImageForDevice on the colossus V100.
py3.13 torch wheels start at 2.5, which dropped sm_70 from default arches. Use py3.12 + rdkit 2025.3.6 instead.
The default PyPI torch wheel ships an older bundled libcudart that predates cudaGraphAddNode (added in CUDA 12.4), and the cu128/cu129 channels dropped sm_70.
|
Greptile encountered an error while reviewing this PR. Please reach out to support@greptile.com for assistance. |
1 participant
Builds the Python package and runs pytest on a single matrix entry (py3.13 / rdkit 2025.9.2). Source build for now; in principle the conda artifact from conda-build.yml could be reused, but cross-workflow artifact passing is awkward.