Pure Go hardware accelerated local inference on VLMs using llama.cpp