Contributing

Contributions are welcome. The bar is specific and non-negotiable in two places: benchmarks are mandatory, and honesty is policy. The full rules live in GUIDELINES.md; this page is the short version.

Before you add a function

Every addition must solve a recurring problem for Numba users, and must answer one of the three benchmark-honesty questions — faster, similar-but-more-ergonomic, or slower-but-unique. "It seems faster" is not accepted; a reproducible table is.

What a PR must include

• A benchmark with a pinned seed, input size, and iteration count, added to benchmarks/ and reflected in BENCHMARKS.md — losing rows included.
• Tests built on reference validation, not just self-consistency: compare against an independent NumPy or pure-Python implementation. See Testing for why this is the only trustworthy coverage in nopython mode.
• A docstring following the fixed structure: what, why, complexity, memory, example, benchmark, limitations, related functions.

Design decisions

Non-obvious trade-offs are recorded in design records — read the relevant one before changing behavior around caching, parallelism, collections, or RNG, and add a new record when you make a decision worth remembering.

Running the suite

.venv/Scripts/python -m pytest        # POSIX: .venv/bin/python

For the machine-specific cache crash, set NUMBA_UTILS_CACHE=0 (see Installation).

Reporting bugs and asking questions

See SUPPORT.md: questions go to Discussions, bugs to Issues, vulnerabilities to a private advisory.