Comment by yekanchi

i mean 4bit quantized. i can roughly calculate vram for dense models by model size. but i don't know how to do it for MOE models?