I think it's llama.cpp that simply #defines all cuda_ functions to rocm_ (99% name-name). Porting seems to be that trivial.