GeoT
A tensor-centric library for GNNs, built on efficient GPU segment reduction.
Graph neural network frameworks usually reach for message-passing abstractions that fight the tensor programming model underneath them. GeoT takes the other route: express GNN computation in terms of segment reduction, and then make that one primitive genuinely fast on the GPU.
The library provides tuned segment-reduction kernels along with the index handling around them, so GNN operators compose with ordinary tensor code instead of living in a parallel universe of framework-specific ops.