Installation¶

pip install torch-pointcloud

This installs the library with torch and torch-geometric. The kernels below ship as prebuilt wheels that must match your torch and CUDA build, so install torch first. The selector writes the commands.

Dependency	Needed by
`pyg-lib`	FPS, kNN and radius search: nearly every point-based model
`spconv`	Point Transformer V3, Sonata, Concerto, Utonia, SpUNet, SPFormer-UNet, SphereFormer, SECOND, VoxelNeXt, Voxel-Mamba, LION
`flash-attn`	Point Transformer V3, Sonata, Concerto, Utonia (`use_flash_attn=False` runs the dense attention instead)
`mamba`	PointMamba, Voxel-Mamba, LION (`mamba-ssm` and `causal-conv1d`)
`ocnn`	OctFormer and the octree transforms, with `dwconv` for its depthwise convolutions
`torchsparse`	SPVCNN (built from source, needs `libsparsehash-dev`)
`sptr`	SphereFormer (with `torch-scatter`, up to torch 2.12)
`kaolin`	A faster nearest-neighbor kernel for the Chamfer distance
`lightning`	The Lightning training modules (with `torchmetrics`)

Package manager

PyTorch

Compute

Extras

uv pip install torch==2.10.0 \
  --index-url https://download.pytorch.org/whl/cu128
uv pip install torch-pointcloud

uv pip install pyg-lib \
  -f https://data.pyg.org/whl/torch-2.10.0+cu128.html

Compatibility¶

Python 3.10+ and torch>=2.8. CI runs the tests on CPU with torch==2.8.0; the benchmarks use torch==2.10.0 with CUDA 12.8.
The point-based families (PointNet, PointNet++, DGCNN, PointNeXt, RandLA-Net, ...) run on CPU. The sparse-voxel and flash-attention families (Point Transformer V3, Sonata, Concerto, Utonia, SpUNet, SPVCNN, OctFormer, the voxel-based detectors) need a CUDA device.

To work on the library itself, see CONTRIBUTING.md.