The measured speedups belong to uniform FIXED-K, which preserves uniformly shaped expert GEMMs; per-token dynamic k is harder to serve efficiently, and the paper reports no matched dynamic-rule throughput results. · CiteArk
The measured speedups belong to uniform FIXED-K, which preserves uniformly shaped expert GEMMs; per-token dynamic k is harder to serve efficiently, and the paper reports no matched dynamic-rule throughput results.
来源：fixed-paper:PDF pp. 7, 12, and 24, Sections 4.3 and 7.3 and Appendix G
