expert-folding

ExFold is a unified training-free framework that accelerates MoE model inference by folding excluded expert contributions into retained experts, achieving up to 1.41× speedup while maintaining high quality.