Installation

From PyPI

pip install NuMPI

NuMPI requires Python 3.10 or later and depends only on numpy, scipy, primefac and array-api-compat. Without mpi4py, NuMPI uses its MPI stub and runs serially.

Optional dependencies

Extra	Installs	Needed for
mpi	mpi4py	running in parallel under mpirun
fft	muGrid (>= 1.6.0)	FFTs of domain-decomposed grids
gpu	CuPy	optimization on the GPU
test	pytest, pytest-cov, array-api-strict	running the test suite

The gpu extra installs the generic cupy package, which compiles from source. A prebuilt CuPy wheel for your platform (e.g. cupy-cuda12x or cupy-rocm-6-0) installed instead works as well.

pip install NuMPI[mpi,fft]

Compile mpi4py and muGrid against your MPI library

The wheels of mpi4py and muGrid on PyPI may be built against a different MPI library than the one your mpirun belongs to, or without MPI at all. Build both from source against the MPI library of your system:

CC=mpicc CXX=mpicxx pip install --no-binary mpi4py --no-binary muGrid NuMPI[mpi,fft]

muGrid refuses communicators from an mpi4py linked against a different MPI library, and muGrid.has_mpi tells whether muGrid was compiled with MPI.

Development installation

Clone the repository and install it in editable mode:

git clone https://github.com/muSpectre/NuMPI.git
cd NuMPI
pip install -e .[test]

Running the tests

python -m pytest

This runs serially and exercises the MPI stub, so it works without mpi4py. With mpi4py installed, run the suite in parallel with

mpirun --oversubscribe -n 4 python -m pytest

CI runs the suite serially without mpi4py, and with mpi4py on 1, 2, 4, 8 and 10 processes. Tests that need a communicator of a given size are skipped when there are fewer processes (see Testing utilities).

Building this documentation

pip install -r docs/requirements.txt
mkdocs serve