[Rust Frontend] Fix GLM-5.2 chat template rendering parity

[WoosukKwon]

Purpose

Fix GLM-5.2 chat-template rendering and align Rust tool serialization with the Python frontend.

The MiniJinja 2.22 update from #51235 is now in main. Relative to current main, this PR:

• upgrades MiniJinja and MiniJinja Contrib from 2.22 to 2.24;
• adds a regression test for mid-chain dotted integer lookup such as values.0.name;
• preserves the OpenAI/Python function-field order: name, description, parameters, then strict when present;
• preserves Python field-presence semantics: an absent description serializes as null, a null parameters stays null, an absent strict is omitted, and strict: false is retained;
• adds a real zai-org/GLM-5.2-FP8 render/parse roundtrip covering both reasoning-plus-content and mixed tool-call fixtures.

Version boundary

• MiniJinja 2.18 rejects GLM-5.2's bundled template at a mid-chain dotted integer lookup with unexpected float, expected identifier or integer.
• MiniJinja 2.22 accepts the real GLM-5.2 template and is sufficient for model correctness.
• MiniJinja 2.24, with the existing preserve_order feature, also preserves Serde struct definition order during template iteration. This closes the strict Python prompt-parity gap while allowing the tool projection to remain typed and express field absence with Serde attributes.

Test Plan

cd rust

cargo nextest run -p vllm-chat roundtrip_glm52
cargo nextest run -p vllm-chat chat_template_preserves_openai_tool_field_order
cargo nextest run -p vllm-chat chat_template_preserves_python_optional_tool_fields
cargo nextest run -p vllm-chat 'renderer::hf::'

cargo fmt --all -- --check
cargo clippy -p vllm-chat --all-targets -- -D warnings
git diff --check

The GLM-5.2 roundtrip loads the real tokenizer and bundled chat template, renders fixed conversations, parses the assistant output, appends it to history, and renders the next turn.

Test Result

Red/green evidence:

• MiniJinja 2.18: the real GLM-5.2 roundtrip fails during template compilation at chat:93.
• MiniJinja 2.22: the real GLM-5.2 roundtrip passes.
• MiniJinja 2.22 with the typed tool projection: the field-order regression test fails with description|name|parameters|.
• MiniJinja 2.24 with the typed tool projection: the field-order test passes with name|description|parameters|.
• The previous manual-map projection omits description when absent and fails the optional-field regression test.
• The final typed projection passes the optional-field test with description=null, parameters=null, omitted absent strict, and retained strict=false.

Final validation:

• roundtrip_glm52: passed on both reasoning-plus-content and mixed tool-call fixtures.
• Focused field-order and field-presence tests: passed.
• Hugging Face renderer suite: 76/76 passed.
• cargo fmt, Clippy with warnings denied, and git diff --check: passed.
• A real GLM-5.2 tool prompt rendered by Python and Rust matched exactly: 980 characters, 243 tokens, SHA-256 b3d3bfd8964d9bb3998d4e96a7e7c5d8adc36478e0dbdd4d26ecd3ef31acfd28.

AI assistance was used for source analysis, implementation, testing, and drafting this description. The submitter must review every changed line and validate the results before marking this PR ready for review.

[mergify]

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

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

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}

[WoosukKwon]

/ci run

[github-actions]

✅ Triggered Buildkite CI #84301 for commit fba49b03b23b.

[BugenZhao]

@codex review

[WoosukKwon]

@codex review

This was written by Codex btw

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: fba49b03b2

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve null descriptions in template tools

When a tool omits its optional description, this branch drops description from the template context entirely. The Python frontend uses tool.model_dump() (vllm/renderers/online_renderer.py:180), while FunctionDefinition retains description=None and only removes absent strict/defer_loading (vllm/entrypoints/openai/engine/protocol.py:299-314), so Python supplies description: null. Templates that serialize tools or distinguish undefined from null—such as tool_chat_template_functiongemma.jinja:18—will therefore render different prompts and token sequences, whereas the previous typed serializer preserved null. Insert JsonValue::Null at this canonical position when description is None, while continuing to omit absent strict.

Useful? React with 👍 / 👎.

[BugenZhao]

serializing the Rust tool structs changed the function-field order to description, name, parameters, while the OpenAI request and Python frontend preserve name, description, parameters

Good catch! It turns out the most recent release of minijinja v0.24 has resolved this issue (mitsuhiko/minijinja#920) so we can stick to structs. I'll push some more commits.

[BugenZhao]

@codex review

This was written by Codex btw

My last @codex review was also sent by Codex 🤣

[chatgpt-codex-connector]

😄 Fair point—the latest @codex review request also came from Codex, so there’s no new human-requested action here.

No code changes, commit, or follow-up pull request were created.

View task →

[BugenZhao]

/ci run

[github-actions]

✅ Triggered Buildkite CI #84316 for commit fd7b1785eba3.