mps-pointops reference

Metal point-cloud, graph, and geometry operators for PyTorch tensors on Apple Silicon.

This reference describes the repository source and the bounded v1.0.0 release scope. The latest published package and source on main may differ during release preparation. Check PyPI and the release tags before attributing a feature to an installed version. Opt-in research paths and private modules do not imply complete upstream-library compatibility.

Interactive explainer Source and issues Version archive series

Start here

python -m pip install mps-pointops

Native Metal calls require an Apple Silicon Mac, Python 3.10 or later, an MPS-capable PyTorch build (package metadata requires torch>=2.7), and torch.backends.mps.is_available() == True. CPU tensors use the PyTorch reference path where the public function documents one; an MPS request does not silently move to CPU. See the tested combinations and limits before deploying on another OS or GPU generation.

The API index distinguishes public, opt-in, and private modules. The numerical contracts explain ordering, rounding, and gradients. The evidence guide links raw logs and synchronized measurements to the exact hardware and workload they cover.

Reproduce a Safe Math run

Set the mode before starting Python. The MPS compiler caches the setting inside a process.

PYTORCH_ENABLE_MPS_FALLBACK=0 PYTORCH_MPS_FAST_MATH=0 \
  python -m pytest -ra tests

Run Fast Math in a fresh process with PYTORCH_MPS_FAST_MATH=1. Fast Math is a separate observed configuration, not an extension of the Safe Math numerical guarantee.