It would also kill speculative decoding. You would have to run a full inference pass for every token instead of being able to generate multiple tokens with a single pass.