Sobek: Streaming Equivariant Tensor Product Convolutions
Sobek avoids edge-sized tensor-product intermediates by streaming products into bounded receiver-side state.
The paper says conventional equivariant graph neural network implementations waste memory by materializing edge-specific weights, messages, and adjoints. Sobek changes the execution schedule while preserving fully connected multiplicity mixing across forward, backward, and double backward. In the authors’ tests, the generated-CUDA backend was faster in all 75 capacity-matched comparisons, with speedups from 1.2x to 49.7x and peak memory reductions up to 99%. ArXiv · AI/CL/LG's note
The paper says conventional equivariant graph neural network implementations waste memory by materializing edge-specific weights, messages, and adjoints. Sobek changes the execution schedule while preserving fully connected multiplicity mixing across forward, backward, and double backward. In the authors’ tests, the generated-CUDA backend was faster in all 75 capacity-matched comparisons, with speedups from 1.2x to 49.7x and peak memory reductions up to 99%. ArXiv · AI/CL/LG's note
score 5