Contributing

Reporting issues

Use the Bug report / Feature request templates under Issues → New issue.

Development setup

git clone https://github.com/hcompai/late-interaction-kernels
cd late-interaction-kernels
uv sync --extra dev --extra pylate --extra torch-cuda   # or --extra torch-cpu on CPU-only boxes

uv run ruff check . && uv run ruff format --check .
uv run pytest -q                                        # CUDA tests auto-skip without a GPU

Pick exactly one of torch-cuda (CUDA index, cu124) or torch-cpu (CPU-only wheel, what CI uses) — the two are declared conflicting in pyproject.toml. On macOS, torch-cpu falls back to PyPI's default MPS-capable wheel.

GPU tests

GPU tests run on AWS CodeBuild (A10G). They do not fire on pushes to main (CodeBuild spend); they run automatically on v* tag pushes and on PRs carrying the run-gpu-tests label (applying the label requires triage+, so ping a maintainer if your PR needs it). Maintainers can also trigger an on-demand run via the GPU CI workflow workflow_dispatch.

Style

• Python 3.10+; type hints on public APIs.
• Comments explain why, not what.
• Docstrings: short, concrete, no marketing.

Autotune for a new GPU

Run the relevant benchmarks/kernels/bench_*.py, add a shortlist in late_interaction_kernels/_autotune.py keyed on the device-name prefix, re-run and include before / after in the PR.

New kernel variant

Keep it in its own module under late_interaction_kernels/ and follow the existing split: internal _forward returning (scores, argmax), torch.autograd.Function wrapper, pure-PyTorch reference in reference.py, parity tests in tests/.

Publishing a release

1. main is green and CHANGELOG.md Unreleased is filled in.
2. Releases → Draft a new release on GitHub, tag vX.Y.Z off main.
3. Paste the matching CHANGELOG.md section as the body, then Publish.

publish.yml builds and uploads to PyPI via OIDC — no token needed.

License

By contributing you agree your work is licensed under Apache 2.0 (see LICENSE).