[emme1t]

Motivation

The GPT-OSS parser rejects tool_choice="required" when XGrammar cannot build the tool grammar. The equivalent mandatory forms, a named function and allowed_tools.mode="required", currently fall back to unconstrained generation. A request that requires a tool call can therefore proceed with a plain-text response.

This completes the grammar-failure behavior introduced in #4997.

Modification

• Recognize ToolChoice and required AllowedToolChoice alongside the string required when grammar construction fails.
• Raise the existing request-validation error for all three mandatory forms before generation starts.
• Preserve prompt-only fallback for auto and allowed_tools.mode="auto", including response-format prompt injection.
• Extend the existing regression to five tool-choice modes with and without a structured response format; update the parser docstring.

Validation

• The expanded regression on upstream aba987ad produced 4 failed / 6 passed. Both named-function and allowed-required modes failed with and without a response format.
• python -m pytest tests/test_lmdeploy/serve/parsers/test_gpt_oss_parser.py tests/test_lmdeploy/serve/parsers/test_required_tool_choice.py -q: 110 passed.
• All applicable pre-commit hooks and git diff --check passed.

These are parser-level CPU tests on Python 3.13.13 with real LMDeploy imports and XGrammar 0.2.8. The existing Harmony test fixture supplies a scripted token parser, and the regression simulates grammar construction failure. GPU model inference was not exercised.

Checklist

• Pre-commit checks passed.
• Added regression coverage for mandatory and automatic tool choices.
• Updated the relevant docstring.
• No dependency requirement changes.

[Copilot]

🟢 Approval recommended

The focused validation change covers all mandatory forms, preserves automatic fallback, and includes targeted regression coverage.

No open issues 🎉

What changed in this PR

Completes the GPT-OSS grammar-failure handling introduced in #4997 so mandatory tool choices cannot fall back to unconstrained generation.

Changes:

• Reject failed grammar construction for named functions and required allowed-tool choices.
• Preserve prompt-only fallback for automatic modes and clarify the parser docstring.
• Expand regression coverage to five tool-choice modes, with and without a structured response format.

File	Description
tests/​test_lmdeploy/​serve/​parsers/​test_gpt_oss_parser.py	Covers mandatory rejection, automatic fallback, and response-format prompt injection.
lmdeploy/​serve/​parsers/​_openai_harmony.py	Extends mandatory-choice validation and documents grammar-failure behavior.

_{Review effort: Balanced}

[windreamer]

LGTM