[Bugfix] Multiple fixes to tool streaming with hermes and mistral

[cedonley]

REDO of PR #10782

FIX #10781
FIX #10589
FIX #10821
FIX #10900

Key fixed issues:

• UTF-8 escaping mismatches result in argument corruption
• Final delta often not returned when chunks are small
• Short initial arguments can corrupt entire arguments return
• Mistral tool call id generation is inconsistent between stream/non-stream and causes failures

To be reviewed in future PRs:

• Simplifying overall tool streaming -- reduce reliance on json rewriting and diff's
• Identify similar issues in other tool parsers
• When using speculative decoding, final delta can sometimes still be dropped

@K-Mistele @gcalmettes

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[cedonley]

Regarding failure in entrypoints-test:

• Pythonic tool test failure is not a regression
• Appears to be an intermittent issue with that parser similar to the ones I've fixed with hermes and mistral
  • Final delta of arguments seems to be missing
• The issue appears in fastcheck runs from other recent non-tool PRs, so it is not caused by this PR
• My local run of pytest on this branch passes for the same tests

I looked at that parser to see if there was an easy fix, but the logic is somewhat different from the others and I'd rather not try to overload this PR.

[gcalmettes]

Thanks for your work on this.

[gcalmettes]

I agree that the overall tool streaming extraction logic could/should be simplified, and should be addressed in another PR.

[K-Mistele]

I agree that the overall tool streaming extraction logic could/should be simplified, and should be addressed in another PR.

Yes, I have been of half a mind to rewrite from scratch using finite state machines which seems like the most appropriate & clean approach; just haven't had time to dig into it yet

[cedonley]

@K-Mistele Yeah. I was thinking the same thing.

It doesn't help that the OpenAI tool interfaces have changed multiple times and the underlying models don't even stay consistent within families. Going to be hard to completely future-proof this, so may want to let things settle a bit before the rewrite.

[cedonley]

@DarkLight1337 @mgoin This is good to merge per reviews from @K-Mistele and @gcalmettes

Thanks!

[DarkLight1337]

Thanks for fixing!

[DarkLight1337]

Can you merge from main to fix the CI failures?