The Kernel I Wrote Without Knowing CUDA

My DGX Spark was running slower MoE inference than it should have — because of a two-line bug that blocked every SM12x GPU

Ride the Lightning