Running RAP with LLMs Missing get_next_token_logits and get_loglikelihood

[ERoszczyk]

I've noticed that many LLMs available through APIs do not support functions like get_next_token_logits or get_loglikelihood. Since RAP relies on these functions, I'm wondering if it's possible to run RAP with an LLM that doesn't provide them.

I'm particularly interested in the prontoqa implementation using RAP. However, I found that both of these functions are used in the fast_reward function within ProntoQAConfig (see this link). Although there’s a comment referring to this as "OLD fast reward code," it appears that this implementation is still in use across several parts of the codebase.

Could someone clarify whether there’s an alternative approach for LLMs that lack these functions, or if this part of the implementation is planned for replacement? Any insights or suggestions would be appreciated.