I've had success using my AMD GPU with the OpenCL backend for llamacpp. The ROCm backend had pretty bad performance though.