I think it should, as you dont need to use the encoder layer on the new tokens, you just read the embedding from the cache. that's why cache reads are cheaper