Really, it's just the Python frameworks on AMD consumer GPUs that suck. AMD datacenter GPUs seem to work with them well enough, and llama.cpp crushes it for the rest of us.