Contributing

Bug reports, benchmarks on new hardware and algorithmic improvements are all welcome. The short version lives in CONTRIBUTING.md; the pages in this section go into detail.

Contributor License AgreementCopy Anchor Link

Before we can merge your contribution, you must sign the Contributor License Agreement. The CLA bot posts a one-time comment on your first pull request with instructions.

Reporting bugsCopy Anchor Link

Open an issue with:

• the versions of src_method, numpy and, if relevant, cupy;
• a minimal reproducer, ideally with a fixed seed=;
• the observed and expected behaviour.

For numerical-accuracy reports, include the bond dimensions, the number of sites, the dtype and the error metric you used.

Development setupCopy Anchor Link

The repository ships a Dev Container configuration and a Nix flake that install every dependency for you. By hand, you need uv:

uv sync --all-groups --all-extras
uv run prek install --prepare-hooks

Quality gatesCopy Anchor Link

All of these must pass before a pull request can be merged; prek runs the lint, format and type checks on every commit, and the Lint workflow runs the same hooks in CI.

uv run prek run --all-files     # lint, format, type check, lockfile
uv run pytest -m "not slow"     # fast test suite
uv run pytest                   # full suite, including slow tests

New behaviour needs a test. Numerical changes need a test that pins the accuracy, not just the shapes.

Pull requestsCopy Anchor Link

• Keep pull requests focused on a single concern.
• Use Conventional Commits, optionally with a gitmoji after the colon, e.g. fix: 🐛 trim terminal bond on the left-to-right sweep.
• Update the documentation and the docstrings alongside the code.
• If the change affects performance, include before and after numbers and the hardware they were measured on.

Code styleCopy Anchor Link

The project targets Python 3.11+, is checked with ruff under a strict rule set and type-checked with ty. Public functions carry type hints and Google-style docstrings without types in the argument list. The algorithms go through src_method.utils._backend rather than importing NumPy or CuPy directly, so the CPU and GPU paths stay in sync.