Mixture-of-Experts Serving

Fine-grained MoE architectures where a learned router activates a small fraction of total parameters per token, and the expert-routing traffic patterns that shape how they must be served.