The htop for LLM inference see exactly where every GB of VRAM goes and get measured quantization savings.
