Read the original whitepaper or go look at how any framework implements it. You will see that tokens not predicted by greedy sampling of the target model are r

Read the original whitepaper or go look at how any framework implements it. You will see that tokens not predicted by greedy sampling of the target model are rejected. Ergo, they are mathematically identical.