[GPT-OSS] Structure_Tag support for gpt-oss tool-call in cot

[Hanchenli]

Purpose

The model sometimes generates function calls in CoT that have not been provided in tool_server. This PR uses structural tag to force model to 1. do not generate function calls when tool_server is not provided. 2. Only generate function that are present in the tool_server.

example custom structural tags: https://gist.github.com/frank-wei/4929118fe0881ee561ce48f15c1f80ed

Test Plan

Testing with @frank-wei

Test Result

vllm serve run: https://gist.github.com/frank-wei/86d37061685fd45e5b831b6583d865d6

see number of turns reduced:

• without structural tags: https://gist.github.com/frank-wei/3c89e45e028cb0f61dc56c51f464252f
• with structural tags: https://gist.github.com/frank-wei/45daec7615f0c75ab3d97594746324c2

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @Hanchenli.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[gemini-code-assist]

Code Review

This pull request introduces support for structural tags in gpt-oss tool calls, aiming to constrain the model's output to valid function calls. The changes span model configuration, engine arguments, the OpenAI protocol, and a new reasoning parser for gpt-oss. While the overall approach is sound, I've identified a critical bug in gptoss_reasoning_parser.py involving the mutation of a global dictionary, which could lead to incorrect behavior in subsequent requests. Additionally, there's a high-severity bug in a log statement within the same file that could cause runtime errors. I have provided detailed comments and code suggestions to address these issues. The rest of the changes, including new tests and dependency updates, appear to be in good order.

[bbrowning]

If built-in functions are enabled, my reading of this change is that those will be the only tools we allow to be called. No user-provided tools will be allowed by our guidance here. Am I reading the code and tags wrong? We add a trigger for "<|channel|>commentary to=" and then the only tags we allow are for the built-in functions. Does the generated text not need to match any of our tags? If that were true, then this wouldn't accurately be constraining the CoT from generating function calls when no tool_server is provided, right?

[yeqcharlotte]

@frank-wei ci failure is relevant

[frank-wei]

Looks like it is failed in trunk https://buildkite.com/vllm/ci/builds/35163/steps/canvas?sid=0199eb2d-ccad-4e6b-963d-ef868491a020?
Got this link from: https://app.hex.tech/533fe68e-dcd8-4a52-a101-aefba762f581/app/vLLM-CI-030kdEgDv6lSlh1UPYOkWP/latest?selectedStaticCellId=0199ac4b-8cf6-7ee8-8db1-2dff4166cbcf

[hmellor]

Yes this Phi3 test has been unreliable on main for a while. cc @DarkLight1337

[DarkLight1337]

I should have fixed it on main already. Let me rebase this PR