Contributing

Contributions are welcome from any “array-consuming” library contributors who have found themselves writing private array-agnostic functions in the process of converting code to consume the standard.

Thanks to all contributors so far!

How to contribute a new function

  • Open an issue to propose the new function. You may want to wait for initial feedback on the issue before diving into an implementation. Feel free to skip this step if there is already an open issue for the function.

  • Add the implementation of your function to src/array_api_extra/_lib/_funcs.py.

    • Ensure that your function includes type annotations and a numpydoc-style docstring.

    • Add your function to __all__ at the top of the file.

  • Import your function to src/array_api_extra/__init__.py and add it to __all__ there.

  • Add a test class for your function in tests/test_funcs.py.

  • Add your function to docs/api-reference.md.

  • Make a PR!

How to add delegation to a function

See the tracker for adding delegation.

  • If you would like to discuss the task before diving into the implementation, click on the three dots next to the function on the tracker issue, and choose “Convert to sub-issue”.

  • Create a function in src/array_api_extra/_delegation.py with a signature matching the function in src/array_api_extra/_lib/_funcs.py, and move the docstring to the new function. Leave a one-line docstring in _funcs.py, pointing to _delegation.py to see the full docstring.

  • Also move the initial array_namespace call and any input validation over to the new function.

  • Add delegation to backends using the if _delegate pattern. See src/array_api_extra/_lib/_backends.py for the full list of backends we have worked with so far.

  • After all delegation layers, return the result from the implementation in _funcs.

  • Simplify the signature in _funcs.py to remove impossible arguments now that it is only called internally via _delegation. For example, the xp parameter can be changed from type ModuleType | None to ModuleType.

  • Don’t worry if you are not sure how to do some of the above steps or think you might have done something wrong - make a PR!

Development workflow

If you are an experienced contributor to Python packages, feel free to develop however you feel comfortable! However, if you would like some guidance, development of array-api-extra is made easy with Pixi:

pixi shell -e dev

  • To run the tests:

pixi run tests

  • To build the docs locally:

pixi run docs

  • To build and preview the docs locally:

pixi run open-docs

  • To install pre-commit hooks:

pixi run hooks

  • To run pre-commit checks on staged files:

pixi run pre-commit

  • To run the full lint suite:

pixi run --environment=lint lint

  • To enter an interactive Python prompt:

pixi run ipython

  • To run individual parts of the lint suite separately (for example):

pixi run --environment=lint pyright
pixi run --environment=lint dprint

  • To generate the coverage report:

pixi run coverage

  • To generate and display the coverage report:

pixi run open-coverage

Alternative environments are available with a subset of the dependencies and tasks available in the dev environment:

pixi shell --environment=docs
pixi shell --environment=tests
pixi shell --environment=tests-backends
pixi shell --environment=lint

If you run on a host with CUDA hardware, you can enable extra tests:

pixi shell --environment=dev-cuda
pixi shell --environment=tests-cuda
pixi run --environment=tests-cuda tests