Fix heuristic scheduling for UAG

[jmamou]

What does this PR do?

Fix #34803

The number of matching tokens according to the draft vocabulary is missing in candidate_generator.py. To fix the issue, we propose to modify update_candidate_strategy by comparing num_matches (the number of matched tokens according to the target vocabulary) with len(scores[0]) - 1 that represents the number of tokens generated by the draft model according to the target vocabulary.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

@gante @ArthurZucker @zucchini-nlp

[zucchini-nlp]

@jmamou thanks for fixing! Quick clarification, why is there a -1, shouldn't scores reflect the length of generated tokens?

[jmamou]

@jmamou thanks for fixing! Quick clarification, why is there a -1, shouldn't scores reflect the length of generated tokens?

@zucchini-nlp

scores contains also the additional token produced by the target model independently of the draft generation process as explained here

[zucchini-nlp]

Ah, I see, thanks for explaining. Can be merged after core maintainer's review :)

[ArthurZucker]

Cool! Thanks for fixing!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.