RaMP framework optimizes Mixture-of-Experts inference by dynamically selecting kernel configurations based on runtime expert routing, achieving 1.22x kernel speedup and 1.30x end-to-end speedup in vLLM serving

RaMP is a routing-aware dispatch framework that selects the fastest kernel configuration from a runtime expert histogram, achieving 0.93% mean regret versus exhaustive search using a four-parameter wave cost model fitted from 10-24 minutes of one-time profiling per model.

A performance-region analysis derived from hardware constants alone predicts when each optimization helps, correctly predicting all 8 tested architectures including 3 unseen ones. Paired with a CuTe DSL kernel exposing 134-268 polymorphic configurations, RaMP delivers 1.22x kernel speedup over static dispatch.

In vLLM serving, RaMP achieves 1.30x end-to-end speedup over Triton, 1.41x over DeepGEMM, and 1.13x over FlashInfer CUTLASS. When applied to Alpha-MoE, it delivers 1.14x speedup with no source modification, since the model depends only on CTA grid geometry and is kernel-agnostic.