“Drafter models improve inference efficiency by using a small model to generate multiple token candidates, which are then verified in parallel by a larger model.”