Accessing Confidence Scores or Perplexity in VILA's Text Generation

[LebinLiang]

I am currently applying VILA in my project, and I would like to obtain the confidence scores or perplexity corresponding to each generated text sequence. I would like to ask whether there is an existing API in the VILA codebase that provides such information, or if it is possible to modify the code to output the confidence score for each response?

[Lyken17]

@hills-code have the experience to add logprob to output. See https://huggingface.co/Efficient-Large-Model/NVILA-Lite-2B-Verifier for details