Maybe consider putting "cutlass" in your CUDA/Triton kernels | Some blog
maknee.github.io
Personal website for some random tidbits I work on