Generating multiple tokens simultaneously rather than one-at-a-time, reducing total decoding steps needed.