FastEEP: Fast Intra-Instance Elastic Scaling for Expert-Parallel MoE Serving

Yunfei Feng, Jingkai He, Shiqi Liu +3

Large-scale expert parallelism (EP) has become the standard deployment configuration for Mixture-of-Experts (MoE) models in production. Meanwhile, the highly dynamic nature of LLM…