I created a repo that implements GEMM kernels for most datatypes supported on B200. It provides a python library that can also emit the `.cu` and `.cuh` files, so if you or an agent needs the raw files, the OpenGEMM library can provide the kernel without spending any tokens.
1 comment
[ 0.25 ms ] story [ 6.8 ms ] threadBlog: https://aramesh10.github.io/opengemm/opengemm.html