GPU VRAM only for small models with llama.cpp: is it possible?(reddit.com)