MoE places extremely stringent demands on multi-card communication, memory bandwidth, and operator scheduling, and general-purpose CUDA operator libraries often cannot directly bring out the best performance of Ascend hardware.