What do you use instead of llama.cpp? With vllm for example most models don't seem to be supported out of the box.