Not one token at a time.

Intel shows real revisable diffusion drafts. NVIDIA streams committed output from vLLM, not intermediate denoising drafts.