Installation¶
Package¶
This installs the library together with
torch and
torch-geometric.
PyTorch and CUDA extensions¶
Select a package manager, torch version, compute platform, and the extras you need; the command below
updates accordingly. Options without a matching wheel are grayed out. flash-attn, mamba, and the
PyG extensions install as prebuilt wheels, from the
Astral GPU indexes and the PyG wheel index.
| Extra | Required by |
|---|---|
pyg-lib |
FPS, kNN, and scatter pooling (torch-scatter, torch-cluster, ...) |
flash-attn |
Point Transformer V3, Sonata, Concerto, Utonia |
mamba |
Point-Mamba, Voxel-Mamba, LION |
spconv |
SpUNet, SPFormer-UNet, voxel-based detectors |
ocnn |
OctFormer (installed together with dwconv) |
torchsparse |
SPVCNN |
sptr |
SphereFormer |
lightning |
The Lightning training modules |
The combination tested in CI and used for the benchmark results is torch==2.10.0 with CUDA 12.8.
Other torch or CUDA versions and exact wheel pins are listed on the Astral GPU indexes
and the
PyG wheel index.
Development¶
To work on the library, clone it and set up the environment with
uv:
For the test and docs tooling, install the dev group and all extras:
Similarly to the above section, install the extras dependencies for your machine (CUDA 12.8, torch 2.10.0, etc.).
We provide several commands to help you get started:
make test # Run the test suite
make format # Format the code
make lint # Lint the code
make type # Type check the code
make clean # Clean the build artifacts
make docs # Build the documentation
make serve # Serve the documentation locally
Compatibility¶
- Python: 3.10+
- PyTorch: the library requires
torch>=2.5. The tested combination istorch==2.10.0with CUDA 12.8 wheels; the selector above covers2.9to2.13across CPU and CUDA 12.6 to 13.2. - CUDA: optional for the point-based families (PointNet, PointNet++, DGCNN, PointNeXt, PointMLP, PointConv, PointCNN, RandLA-Net, and similar), which run inference and training on CPU. The sparse-voxel and flash-attention families (Point Transformer V3, Sonata, Concerto, Utonia, SpUNet, SPVCNN, OctFormer, and the voxel-based detectors) require a CUDA device and their optional dependencies (
spconv,torchsparse,ocnn,flash-attn).