PCoMoE: Shifting MoE Inference from Monolithic Expert Selection to Fine-Grained Path Composition
PCoMoE breaks MoE inference below the expert level, composing reusable paths instead of selecting whole experts.
The paper says current MoE systems optimize too coarsely by treating each expert as an atomic unit. PCoMoE reframes expert computation at the path level, then prunes low-value path combinations layer by layer. Its execution engine is built to reuse sub-expert structures while keeping overhead bounded. The authors report up to 1.31x end-to-end inference speedup and a 10% accuracy gain. ArXiv · AI/CL/LG's note
The paper says current MoE systems optimize too coarsely by treating each expert as an atomic unit. PCoMoE reframes expert computation at the path level, then prunes low-value path combinations layer by layer. Its execution engine is built to reuse sub-expert structures while keeping overhead bounded. The authors report up to 1.31x end-to-end inference speedup and a 10% accuracy gain. ArXiv · AI/CL/LG's note
score 5