Running on Zero 3 3 D2F LLaDA Instruct 8B 👁 Diffusion LLMs Can Do Faster-Than-AR Inference via Discret