Add customized renormalized moe routing kernel for moe cutlass backend#4955

Merged

byshiue merged 3 commits intoNVIDIA:mainfrom

ChristinaZ:opt_Topk_Qwen3_onHopper

Jun 9, 2025

Commits on Jun 8, 2025

Add customized renormalized moe routing kernel for moe cutlass backend. Replace name Qwen3 with RenormalizeNaive.
ChristinaZ
committed
Add support for arbitrary expert number less than or equal to 128 and data type half
ChristinaZ
committed

Commits on Jun 9, 2025

Prioritize elements with smaller indices
ChristinaZ
committed