There's also ZLUDA, which can run llama.cpp and some other CUDA workloads already without any modification, but it's still maturing.