PCoMoE: Shifting MoE Inference from Monolithic Expert Selection to Fine-Grained Path Composition
The episode discusses the paper "PCoMoE: Shifting MoE Inference from Monolithic Expert Selection to Fine-Grained Path Composition." The hosts explain that PCoMoE fundamentally shifts how large models operate by moving away from selecting a single expert, instead enabling custom, fine-grained sub-transformation compositions for better efficiency and performance.