I have only used nvidia cards, but in my experience kobold.cpp and ollama forks of llama.cpp can be easier to compile.