Anatomy of a high-performance EP kernel

This article discusses the design and implementation of a high-performance expert parallelism (EP) kernel for large language models, focusing on the anatomy of the kernel and its optimization for throughput and latency.