dmlc/dgl

Provides GPU-accelerated message passing primitives and distributed training across multiple GPUs/machines for billion-scale graphs, with framework-agnostic design supporting PyTorch, TensorFlow, and MXNet backends. Includes pre-built GNN layers, comprehensive model zoo, and sampling-based stochastic training for efficient large-graph learning. Optimized communication and memory management reduce distributed training overhead while maintaining compatibility with standard deep learning ecosystems.

14,245 stars. Used by 4 other packages. No commits in the last 6 months. Available on PyPI.