pytorch3d-slicer-wheels

Prebuilt PyTorch3D wheels for 3D Slicer's bundled Python.

What this is

PyTorch3D doesn't publish Windows wheels on PyPI, so installing it into Slicer's bundled Python normally requires compiling from source with a matching CUDA toolkit and MSVC. This repo builds those wheels once in CI and serves them through a PEP 503 "simple" index hosted on GitHub Pages.

Current build matrix

Slicer	Python	torch	CUDA	OS
5.10	3.12	2.6.0	cu124	win_amd64
5.10	3.12	2.8.0	cu129	win_amd64
5.10	3.12	2.8.0	cpu	win_amd64
5.10	3.12	2.12.0	cu130	win_amd64
5.10	3.12	2.2.2	cpu	macosx_10_13_x86_64

PyTorch3D version: 0.7.9

Using the wheels directly

From a Slicer Python console (pin to the local-version tag of the wheel that matches your installed torch — +pt260cu124 for torch 2.6.0+cu124, +pt280cu129 for torch 2.8.0+cu129, +pt280cpu for torch 2.8.0+cpu):

import slicer.util
slicer.util.pip_install(
    "pytorch3d==0.7.9+pt260cu124 "
    "--extra-index-url https://ImageMindAnalytics.github.io/pytorch3d-slicer-wheels/simple/"
)

Building locally

You shouldn't need to. CI builds on every push to main and on manual dispatch. If you do need to reproduce locally, see scripts/build_one.py and the matching workflow in .github/workflows/.

License

The build scripts here are MIT. The wheels they produce are PyTorch3D, which is BSD-3-Clause.