Adaptive P sampler: update review logic, delete old code comments, put prep stage after logit bias

[dungquixote42]

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

This PR deletes old code comments left in #1359 .

The review logic now tracks slot.token_buffer.size(), instead of n_rewind. Looking at the end result of bans/rewinds should be more reliable than recreating the math. As a side effect, this corerctly handles a case where send_token_results(2) does not actually send all tokens.

This removes all llama_review_adaptive_p_impl() related variables from buffer_and_check_string_ban(). Hopefully, this helps with the supposed bug regarding #1243 .

[dungquixote42]

Converting to draft until #1243 is merged.

[dungquixote42]

Tokens belonging to banned strings often have high logits, and llama_prep_adaptive_p() drops tokens with logit<max-30 while collecting the probabilities. It looks like collecting before logit bias leaves too few tokens as a result. Restoring the original llama_prep_adaptive_p() with dd3bfac.

[ikawrakow]

@dungquixote42

I see you keep changing things around. When you get to a version that does not need multiple follow up PRs, let me know.

[dungquixote42]

@dungquixote42

I see you keep changing things around. When you get to a version that does not need multiple follow up PRs, let me know.

Yeah. I just keep finding these cases I did not account for. It is over now, at least from my end.

[dungquixote42]

@ikawrakow This PR is ready for review.