[Bugfix][Structured Output] Fix structural_tag bitmask not applied on reasoning models

[CatherineSue]

Purpose

Fix structural_tag constraints being silently ignored on reasoning models (e.g., gpt-oss with openai_gptoss reasoning parser).

When a reasoning model uses a structural_tag constraint (e.g., triggered_tags format), should_fill_bitmask() and should_advance() in StructuredOutputManager return False during the reasoning phase. This happens because reasoning_ended is computed once from prompt tokens (which do not contain the reasoning end sequence) and never updated during generation. As a result, the grammar bitmask is never applied, the grammar state is never advanced, triggers never fire, and the constraint is completely ignored.

structural_tag grammars handle reasoning/content boundaries internally via triggers — the grammar itself knows when to allow free text (reasoning) and when to constrain output. The external reasoning_ended gate prevents this from working.

Fix: Add an early return in both should_fill_bitmask() and should_advance() when the request uses a structural_tag constraint, bypassing the reasoning_ended gate. Other constraint types (json_schema, regex, etc.) are unaffected and keep the existing behavior.

Test Plan

Tested with gpt-oss-120b (tp=4) via vllm serve HTTP Chat Completions with a structural_tag response_format:

curl -s -X POST http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "/raid/models/openai/gpt-oss-120b",
    "messages": [{"role": "user", "content": "List exactly 2 fruits."}],
    "response_format": {
      "type": "structural_tag",
      "format": {
        "type": "triggered_tags",
        "triggers": ["<|channel|>analysis", "<|channel|>final"],
        "tags": [
          {"begin": "<|channel|>analysis<|message|>", "content": {"type": "any_text"}, "end": "<|end|>"},
          {"begin": "<|channel|>final<|constrain|>json<|message|>", "content": {"type": "json_schema", "json_schema": {"type": "object", "properties": {"items": {"type": "array", "items": {"type": "string"}, "minItems": 2, "maxItems": 2}}, "required": ["items"], "additionalProperties": false}}, "end": ""}
        ],
        "at_least_one": true,
        "stop_after_first": false
      }
    },
    "temperature": 0
  }'

Test Result

Before — structural_tag trigger never fires, Harmony markers leak into content:

{
    "message": {
        "content": "<|channel|>final<|constrain|>json<|message|>{\"items\":[\"Apple\",\"Banana\"]}",
        "reasoning": "The user asks: \"List exactly 2 fruits.\" So we need to output exactly two fruit names..."
    },
    "finish_reason": "stop",
    "usage": {"completion_tokens": 98}
}

After — grammar bitmask applied, trigger fires, clean constrained JSON:

{
    "message": {
        "content": "{\"items\":[\"Apple\",\"Banana\"]}",
        "reasoning": "The user asks: \"List exactly 2 fruits.\" So we need to output exactly two fruit names..."
    },
    "finish_reason": "stop",
    "usage": {"completion_tokens": 78}
}

Note the 20 fewer completion tokens — the text-encoded <|channel|>final<|constrain|>json<|message|> markers are no longer generated.

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR: Fix structural_tag constraint silently ignored on reasoning models
• The test plan: Tested with gpt-oss-120b via HTTP Chat Completions with structural_tag response_format
• The test results: Before/after comparison with token count difference
• (Optional) Documentation update: N/A
• (Optional) Release notes: N/A

[gemini-code-assist]

Code Review

The pull request effectively addresses the bug where structural_tag constraints were being ignored in reasoning models. The changes correctly bypass the reasoning_ended gate for structural_tag grammars in both should_fill_bitmask and should_advance methods, ensuring that the grammar bitmask is always applied and the grammar state is advanced as intended. The inline comments clearly explain the rationale behind these changes. The provided test plan and results confirm the fix and demonstrate the intended behavior, including the reduction in completion tokens due to correct constraint application.

[CatherineSue]

I realized that I can use enable_in_reasoning: True. So closing this PR now.