pynumerics

RMC.Numerics for Python. The wheel ships the Numerics.dll assembly from the published RMC.Numerics NuGet package, byte for byte, with a loader that starts one .NET runtime and one Numerics assembly per process, and typed wrappers over every Numerics area: univariate and multivariate distributions, copulas, statistics, hypothesis tests, sampling, MCMC and the bootstrap, integration, differentiation, root finding and ODEs, optimization, special functions, linear algebra and Fourier transforms, data and time series, and machine learning.

The wrappers do no numerical work in Python. Arrays cross between NumPy and .NET in bulk, bit for bit, every wrapper raises Python exceptions with the Numerics messages, and functions can be Python callables or C# source compiled with the bundled Roslyn compiler. The distribution version equals the RMC.Numerics version of the bundled assembly.

Install

pip install pynumerics==2.3.0

The 2.3.0 wheel bundles Numerics.dll from the RMC.Numerics 2.3.0 NuGet package. Requirements: CPython 3.11 to 3.14 on Linux or Windows, and the .NET 10 runtime (Microsoft.NETCore.App 10.0). Set DOTNET_ROOT when .NET is not installed in the default location.

from pynumerics import distributions

gev = distributions.GeneralizedExtremeValue(location=10.0, scale=2.0, shape=-0.1)
print(gev.inverse_cdf([0.5, 0.99, 0.999]))   # one call into InverseCDF(IList<double>)

Documentation

The user guide and the API reference are in docs/: start with getting started, which covers installation, PYNUMERICS_DLL_DIR and runtime selection, arrays and errors, value semantics, callbacks and migration from rmc-numerics. The changelog lists the changes of each release. Build the HTML pages, including the API reference generated from the docstrings, from python/pynumerics:

python -m pip install -e ".[docs]"
python -m sphinx -W --keep-going -b html docs docs/_build/html

Releases are published to PyPI by .github/workflows/PythonPublish.yml with trusted publishing. It starts by itself when NuGetPublish.yml has pushed RMC.Numerics X.Y.Z to nuget.org for the GitHub Release vX.Y.Z: it waits until nuget.org serves the package, builds the wheel from it, tests it on Linux and Windows, and uploads it to TestPyPI and then, after the pypi environment's reviewer approves, to PyPI. _version.py must already declare X.Y.Z; no commit is needed after the NuGet release, because the wheel build writes the package's SHA-256 into the wheel.

Build and test

From python/pynumerics:

python scripts/build_wheel.py --stage-only       # stage _bin/Numerics.dll and the Roslyn assemblies
python -m pip install -e ".[test]"
python -m pytest                                  # includes tests/test_docs.py: every docs code block
python scripts/build_wheel.py                    # build dist/pynumerics-<version>-py3-none-any.whl

The script takes Numerics.dll from lib/net10.0 of the RMC.Numerics package in the NuGet global packages folder (or --nupkg PATH, or --download from nuget.org), or a locally built assembly with --dll PATH. It reads the version, the assembly version and the SHA-256 of that file, refuses a version or assembly version that differs from src/pynumerics/_version.py, and writes all three into the wheel's _version.py (the tracked file keeps NUMERICS_SHA256 empty, so assert_numerics_identity() needs an explicit sha256 in a source checkout). It stages Microsoft.CodeAnalysis.dll and Microsoft.CodeAnalysis.CSharp.dll from lib/net9.0 of the Microsoft.CodeAnalysis.Common and Microsoft.CodeAnalysis.CSharp 4.14.0 packages (NuGet cache, or --download), each checked against its ROSLYN_*_SHA256 pin; --roslyn-only stages only those. To test against a locally built assembly, set PYNUMERICS_DLL_DIR to Numerics/bin/Release/net10.0, or build a wheel from it as the pull request check does:

dotnet build ../../Numerics/Numerics.csproj -c Release -f net10.0
python scripts/build_wheel.py --dll ../../Numerics/bin/Release/net10.0/Numerics.dll --download

The same checks run on Linux in the .github/docker/linux-test image, from the repository root:

docker build -t numerics-linux-test .github/docker/linux-test
docker run --rm -v "$PWD:/src:ro" -v "$PWD/python/pynumerics/dist:/out" numerics-linux-test
docker run --rm -v "$PWD/python/pynumerics/dist:/dist:ro" numerics-linux-test identity-check.sh /dist

The first run builds and tests Test_Numerics (net10.0), builds the wheel from that build and runs ruff and pytest against it; CHECKS="wheel pytest" skips the .NET tests. The second installs the wheel in a fresh container and checks the identity of the loaded assembly.