Skip to content

Installation

Install the published package for simulation work. A development checkout is needed only when changing the source or building the docs.

Install from PyPI

pip install polarism

Pin a release with pip install "polarism==0.3.8".

The development version can be installed straight from GitHub:

pip install "git+https://github.com/jakrog01/Polarism.git@dev"

Check the installed CLI:

polarism --help

On a CUDA machine, polarism-bench --help shows the benchmark options; a small benchmark can also confirm that the GPU backend is available. See Running on GPUs for a command and guidance. For a source checkout and test workflow, see Contributing.

Optional GPU support

The runtime backend uses NumPy by default and switches to CuPy when cfg.compute_engine.use_gpu = True and CuPy is available. If CuPy is missing, the code falls back to CPU execution.

Recommended workflow:

  • Start with a CPU install and validate correctness first.
  • Add a CuPy build that matches the CUDA toolkit on the target machine: pip install "polarism[cuda12]" for CUDA 12.x or pip install "polarism[cuda13]" for CUDA 13.x.
  • V100 (Volta) systems need polarism[cuda12]; CUDA 13 no longer supports Volta.
  • Re-run the compliance or benchmark suites on GPU hardware before trusting production-scale timings.

For solver choice, precision, and CUDA graph chunks, see Running on GPUs.