Comment by kbumsik

So the draft model's performance is directly linked to the overall speed. Thank you for the explanation!

By the way, can it be slower than without speculative decoding in worst case then?