Be sure you compare inputs tokens to pre-fill rates and output tokens to generation rates.