Skip to content

README.md

Latest commit

 

History

History
102 lines (75 loc) · 4.09 KB

README.md

File metadata and controls

102 lines (75 loc) · 4.09 KB
Raw

pytorch3d-wheels

Prebuilt PyTorch3D wheels for Windows, Linux, and macOS (Apple Silicon), served from a PEP 503 simple index hosted on GitHub Pages.

Why this exists

PyTorch3D doesn't publish wheels for Windows on PyPI, and Linux/macOS support upstream is limited. Installing it normally requires compiling from source with a matching CUDA toolkit and host compiler — a painful experience for users who just want to pip install pytorch3d.

This repo builds those wheels in CI for a bounded set of (Python, torch, CUDA, OS) combinations and serves them through a simple index.

Install

pip install pytorch3d \
  --extra-index-url https://ImageMindAnalytics.github.io/pytorch3d-wheels/simple/

If you have a specific torch + CUDA already installed and want pip to pick the matching wheel, pin to the local-version tag:

# torch 2.11.0 + CUDA 12.6 (Windows / Linux)
pip install "pytorch3d==0.7.9+pt2110cu126" \
  --extra-index-url https://ImageMindAnalytics.github.io/pytorch3d-wheels/simple/

# torch 2.11.0 + CUDA 12.8 (Windows / Linux)
pip install "pytorch3d==0.7.9+pt2110cu128" \
  --extra-index-url https://ImageMindAnalytics.github.io/pytorch3d-wheels/simple/

# torch 2.12.0 + CUDA 12.6 (Windows / Linux)
pip install "pytorch3d==0.7.9+pt2120cu126" \
  --extra-index-url https://ImageMindAnalytics.github.io/pytorch3d-wheels/simple/

# torch 2.8.0 + CPU only (macOS arm64)
pip install "pytorch3d==0.7.9+pt280cpu" \
  --extra-index-url https://ImageMindAnalytics.github.io/pytorch3d-wheels/simple/

The local-version tag format is +pt<torch>cu<CUDA> (or +pt<torch>cpu), with dots stripped — pt280cu129 for torch 2.8.0 + CUDA 12.9.

Supported matrix

This is the support contract. Combinations outside this list are not built; requests to add new combinations are evaluated based on CI cost and demand.

OS Python torch CUDA
Windows x86_64 3.10, 3.11, 3.12, 3.13 2.11.0 12.6, 12.8
Windows x86_64 3.10, 3.11, 3.12, 3.13 2.12.0 12.6
Linux x86_64 (*) 3.10, 3.11, 3.12, 3.13 2.11.0 12.6, 12.8
Linux x86_64 (*) 3.10, 3.11, 3.12, 3.13 2.12.0 12.6
macOS arm64 3.10, 3.11, 3.12, 3.13 2.8.0 CPU

torch 2.12 has no cu128 builds on PyPI's torch index — that's why torch 2.12 here is cu126-only.

(*) Linux wheels are manylinux_2_28_x86_64.

PyTorch3D version: 0.7.9 (the latest upstream release).

CUDA 13.x — experimental (fix being verified)

CUDA 13.x wheels (torch 2.12 + cu132) are kept in a separate matrix and workflow while a fix is being verified. CUDA 13 changed the default for -static-global-template-stub to true, which gives pulsar's explicit template instantiations hidden visibility without emitting their definitions — causing _C.so to fail at link with "hidden symbol ... isn't defined". build_one.py now passes -static-global-template-stub=false via NVCC_PREPEND_FLAGS when CUDA major ≥ 13 (and /Zc:preprocessor on Windows for CCCL). See NVIDIA's writeup.

The matrix and workflows live in matrix-{windows,linux}-cu132.yml and .github/workflows/build-{windows,linux}-cu132.yml. These workflows default to publish: false so a failed run doesn't touch gh-pages. Once a green end-to-end run is confirmed, rows will move into the stable matrices.

Scope

  • This repo only ships wheels — runtime bugs belong on the upstream pytorch3d tracker.
  • Issues here should be about wheel availability, build failures, or installation problems caused by the wheels themselves.

Building locally

You shouldn't need to. CI builds on push to main and on manual workflow dispatch. If you do, 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.