pyg-team/pytorch_geometric

Provides a message-passing API for implementing custom GNN layers and pre-built convolution operators (GCNConv, EdgeConv, etc.) that handle node aggregation and feature propagation. Supports heterogeneous graphs, dynamic temporal graphs, and large-scale models with millions of nodes, alongside specialized data loaders for mini-batch training on both small and giant graphs. Includes built-in benchmark datasets and graph transforms for point clouds and 3D meshes, with `torch.compile` and multi-GPU support for production deployments.

23,561 stars and 1,165,902 monthly downloads. Used by 42 other packages. Actively maintained with 19 commits in the last 30 days. Available on PyPI.