So I did test this, and it doesn't work because the quantized layers need torch._int_mm, which PyTorch's MPS backend doesn't implement. It just throws NotImplementedError.

This is good advice if you have nvidia, but for Mac does not apply currently.

GGUF is unsupported by ComfyUI’s memory management system that enables running models much larger than fit in VRAM with tolerable efficiency via weight streaming, but for unified memory systems that system is less relevant (unless using models too big to run in unified memory AND having fast enough mass storage to benefit from direct-from-disk weight streaming.)