The VRAM budget should equal model weights plus KV cache; it is not sufficient to calculate only the weights.