>>106381653
I've tried both vulkan and rocm neither remedy the Mi50s biggest problem - the fuckslow PP. Lost my notes but there's several threads on both llama.cpp github and /r/locallama about it.
I only tried llama.cpp though, might be other engines that work better.