// the take
A small C++/CUDA extension for PyTorch providing optimized graph clustering primitives: Graclus, voxel grid pooling, farthest point sampling, k-NN/radius graph construction, nearest-point clustering, and random walk sampling. It's the low-level op library underneath PyTorch Geometric, aimed at people building graph neural networks or point cloud models who need fast neighbor/cluster computations rather than pure-Python implementations.
Each op has both CPU and CUDA implementations with a consistent, minimal functional API (plain tensors in, tensors out), so it drops into any PyTorch pipeline without a framework dependency. The k-NN/radius graph and FPS routines are the actual bottleneck operations in point cloud and GNN pipelines, and having them in compiled C++/CUDA instead of Python loops is the entire point and genuinely pays off at scale. Prebuilt wheels cover a wide matrix of PyTorch/CUDA/OS combinations, which matters a lot for something that otherwise requires a from-source build with matching toolchains.
The README leads with a deprecation notice: this repo is being folded into pyg-lib, so new projects should probably start there instead of taking on a dependency that's officially being phased out. Being a compiled extension, it's exposed to the usual PyTorch ABI fragility — every new PyTorch/CUDA release requires a matching prebuilt wheel or a from-source rebuild, and the wheel matrix lags behind the latest PyTorch versions. Installing from source needs a correctly configured CUDA toolchain and `Torch_DIR`/cmake setup, which is a real source of friction for anyone outside the PyG ecosystem's usual environment. As a thin ops library it has no higher-level clustering abstractions (no batching helpers beyond the basic `batch` vector, no incremental/streaming support), so it's only useful if you're already building on top of it rather than using it standalone.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →