MLX-graphs
A Graph Neural Network library built on Apple's MLX framework, designed to train and run GNNs on Apple Silicon's unified CPU and GPU memory, with reported speedups over other frameworks on large graphs.
MLX-graphs implements Graph Neural Network operations on top of Apple's MLX library so that GNN training and inference run on Apple Silicon's GPU and CPU using MLX's low-level primitives, reporting up to a tenfold speed improvement over other frameworks on large datasets in its own benchmarks.
Because MLX uses a unified memory architecture, graph data lives in memory shared between CPU and GPU, letting a Mac with enough RAM train GNNs on graphs spanning tens of gigabytes directly on its GPU without device-to-device transfers, and letting specific operations run explicitly on either processor without added overhead.
It installs from PyPI or from source, and new GNN layers are built by implementing its message-passing base class, which handles the message-passing mechanics so a contributor can focus on the layer's own logic; example notebooks cover a quickstart and graph classification.