As I understand these models, that makes no sense whatsoever. Is it the tokenisation that takes this crazy amount of time? Because each additional token should take exactly the same amount of time as the first.