Boosting MoE Training Throughput with Advanced Fusion Kernels

Advanced fusion kernels improve training throughput for mixture-of-experts (MoE) models, which enable larger model capacity while activating only a subset of parameters for each token for improved performance scaling.