bility to export/view logprobs before and after constraint masking

[Krusher11]

I am looking to analyze the impact of SGLang’s constraint-guided decoding (e.g., regex or JSON schema) on the model's output distribution. Currently, it is easy to see the final logprobs for the selected tokens, but it is difficult to see the "tension" between what the model originally predicted and what the constraint allowed.

Is there a recommended way or can a new feature be added to:

1. Raw Logprobs: The top-k logprobs generated by the model before any LogitProcessor or bitmask is applied.

2. Masked Logprobs: The logprobs after the constraint mask has been applied (setting illegal tokens to $-\infty$) [I assume this is what is already returned]