accelerator-utilization

ERASE introduces a novel training schedule that detaches subgraphs to overlap backward passes with forward work, improving throughput by up to 9.51% in large-scale recommendation systems while preserving model performance.