Will llama.cpp be the go-to local inference framework for every device?