Isn't that exactly how draft models speed up inference, though? Validating a batch of tokens is significantly faster than generating them.