MAYBE CONSIDER PUTTING "CUTLASS" IN YOUR CUDA/TRITON KERNELS

Summary from the TLDR AI newsletter

Adding 'cutlass' to your kernel name can give you 100+ TFLOPs or -20% FLOPS.