chore(frontend): Add Gemma 4 parser support + Test Cases

[the-david-oy]

Overview:

Add tool-call and reasoning parsers for Google Gemma 4 thinking models to Dynamo's frontend, plus a vendored chat template. With this change, Dynamo's native preprocessor (mode A in docs/agents/chat-processor-options.md) can serve any Gemma 4 deployment that runs on vLLM ≥ v0.19.0 (where Gemma 4 support landed via vllm-project/vllm#38826 and #39027) end-to-end with structured tool_calls and reasoning_content.

Gemma 4's serialization grammar does not match any existing Dynamo parser family, so this PR introduces a new Gemma4 variant in ParserConfig rather than force-fitting it into JSON/XML/DSML. The grammar uses bare unquoted keys, a custom <|"|> string delimiter, and fixed <|tool_call>...<tool_call|> markers:

<|tool_call>call:get_weather{location:<|"|>Tokyo<|"|>,unit:<|"|>celsius<|"|>,count:5}<tool_call|>

Reasoning is wrapped between channel markers with a thought\n role-label prefix that this parser strips:

<|channel>thought
...chain of thought reasoning...<channel|>
Final answer text.

Recommended pairing for a Gemma 4 deployment:

--dyn-tool-call-parser gemma4 \
--dyn-reasoning-parser gemma4 \
--custom-jinja-template examples/chat_templates/gemma4_tool.jinja

Details:

The PR is two commits to make the upstream-diff alignments easy to review separately from the initial port.

Commit 1 — chore(frontend): Add Gemma 4 parser support + Test Cases

New files:

• lib/parsers/src/tool_calling/gemma4/{mod.rs,parser.rs} — recursive-descent parser for the Gemma 4 argument grammar (booleans, numbers, <|"|>-delimited strings, nested objects, arrays) emitting serde_json::Value. Inline #[cfg(test)] coverage of CASE.1–CASE.24 from lib/parsers/TEST_CASES.md.
• lib/parsers/src/reasoning/gemma4_parser.rs — Gemma4ReasoningParser implementing the ReasoningParser trait directly (mirrors granite_parser.rs shape) with <|channel>...<channel|> delimiters, thought\n prefix stripping (the 3-case streaming logic from upstream's reasoning parser), and partial-marker overlap detection at chunk boundaries.
• examples/chat_templates/gemma4_tool.jinja — verbatim copy of upstream vLLM examples/tool_chat_template_gemma4.jinja with SPDX attribution. Required because the stock HF Gemma 4 chat template does not emit the <|"|>-delimited tool-definition encoding the parser expects.
• examples/chat_templates/README.md — per-template usage notes.

Modified files:

• lib/parsers/src/tool_calling/{config.rs,parsers.rs,mod.rs} — added ParserConfig::Gemma4 variant, ToolCallConfig::gemma4() factory, registered names gemma4 (alias gemma-4), and dispatched try_tool_call_parse_gemma4 / detect_tool_call_start_gemma4 / find_tool_call_end_position_gemma4 from the central registry.
• lib/parsers/src/reasoning/mod.rs — added ReasoningParserType::Gemma4 variant and registered the same names.
• lib/parsers/README.md — bumped parser counts (17 → 18 tool-call, 14 → 15 reasoning) and added rows in the parser-family cheat sheets.
• docs/agents/{tool-calling,reasoning}.md — added parser-table rows and the recommended-pairing line.

Commit 2 — fix(frontend): align Gemma 4 parsers with upstream vLLM main

Behavior-diff against upstream vLLM main (recently-merged Gemma 4 PRs and bugfix follow-ups) surfaced four real bugs in the initial port plus one reasoning-parser gap and one preprocessor optimization gap. All fixed:

1. Null-keyword case-insensitivity (parser.rs) — upstream's _parse_gemma4_value does value_str.lower() in ("null","none","nil"). The first-pass port hard-coded a small set of mixed-case spellings and missed NONE, Nul, etc. Replaced with try_consume_keyword that does ASCII-case-insensitive matching with a trailing-word-char guard so nullable doesn't get mis-matched as null plus leftovers.
2. Unterminated <|"|> string in args (parser.rs) — upstream's _parse_gemma4_args:146-149 takes "everything after the opening delimiter" as the value when the closing delimiter is missing. The first-pass port bailed instead, which then triggered the entire-args-object empty-fallback in parse_single_call — i.e. a single truncated trailing arg silently lost ALL prior args. Now mirrors upstream.
3. Empty value key: followed by ,/}/EOF (parser.rs) — upstream's _parse_gemma4_args:128-131 emits {"key": ""}. The first-pass port errored out of parse_value and dropped the whole args object. Detect this case in parse_object_body and insert an empty-string.
4. Truncated tool call (parser.rs) — upstream extract_tool_calls:421-424 returns the raw model_output as content when no complete tool call could be parsed. The first-pass port consumed those bytes silently, hiding model truncation from callers. Now echoes the bytes back as normal_text.
5. Reasoning: dangling end marker (gemma4_parser.rs) — when <channel|> is present but <|channel> is absent (tokenizer with skip_special_tokens=True strips the start token), upstream's offline parser still extracts text-before as reasoning. Added a (None, Some(end)) match arm.
6. is_reasoning_disabled_by_request polish (lib/llm/src/preprocessor.rs) — added gemma4/gemma-4 recognition with enable_thinking: false short-circuit, matching the existing pattern for kimi_k25 / deepseek_v4 / nemotron_nano. Without this, the parser still ran on every chunk just to fall through (since no <|channel> markers arrive when the chat template doesn't inject <|think|>).

One intentional divergence from upstream: between-call text is preserved (Sure thing. <call> All set. → "Sure thing. All set.") rather than dropped, because that matches the established kimi_k2_parser.rs pattern in the rest of Dynamo's parser registry. Documented inline.

Where should the reviewer start?

1. lib/parsers/src/tool_calling/gemma4/parser.rs — the centerpiece. Pay particular attention to:
   • The recursive-descent Cursor / parse_object_body / parse_value / parse_array / parse_number / try_consume_keyword chain (lines ~198–402). The grammar parses straight into serde_json::Value.
   • try_tool_call_parse_gemma4 (lines ~91–131) for the outer call-extraction loop and the truncated-call echo behavior.
   • The #[cfg(test)] mod tests block — 31 tests covering CASE.1–CASE.24 plus targeted cases from upstream's tests/tool_parsers/test_gemma4_tool_parser.py (test_unterminated_string, test_empty_value, test_incomplete_tool_call, parallel calls, function names with hyphens/dots, HTML/newlines in string values, case-insensitive null aliases, keyword-prefix non-consumption).
2. lib/parsers/src/reasoning/gemma4_parser.rs — 12 tests including the new detect_dangling_end_marker_* cases that mirror upstream's INVALID_SIMPLE / INVALID_COMPLETE test fixtures, plus the streaming thought\n prefix-split-across-deltas cases.
3. lib/parsers/src/tool_calling/config.rs — the new ParserConfig::Gemma4 unit variant and the ToolCallConfig::gemma4() factory. The variant is unit (no per-instance config) because Gemma 4's markers are fixed and not user-tunable, unlike KimiK2ParserConfig / Glm47ParserConfig.
4. lib/parsers/src/tool_calling/parsers.rs — the registry/dispatch edits: map.insert("gemma4", ...) + the new dispatch arm in try_tool_call_parse and the partial-prefix detection arm in detect_tool_call_start.
5. lib/llm/src/preprocessor.rs — only the is_reasoning_disabled_by_request table entry plus four parametric test cases. Worth glancing at to confirm it follows the existing kimi_k25 / deepseek_v4 pattern.
6. examples/chat_templates/gemma4_tool.jinja — verbatim copy of upstream vLLM's template, only a 3-line SPDX/comment header added at the top. Reviewers can diff against https://github.com/vllm-project/vllm/blob/main/examples/tool_chat_template_gemma4.jinja to verify.

Validation:

• cargo build -p dynamo-parsers — clean (0 errors, 0 warnings).
• cargo test -p dynamo-parsers --lib gemma4 — 44/44 pass.
• cargo test -p dynamo-parsers (full crate) — 443 pass, 5 ignored, no regressions in other parsers.
• cargo fmt --check --all — workspace clean.
• cargo clippy -p dynamo-parsers — no warnings on the new code (one pre-existing warning in glm47_parser.rs:177, untouched by this PR).
• pre-commit run --files <13 changed files> — all hooks pass (codespell, case-conflict, merge-conflict, mixed-line-ending, trailing-whitespace, pytest-marker-report).
• cargo test -p dynamo-llm --lib is_reasoning_disabled_by_request — could not run on the development host because the dynamo-llm crate has a pre-existing macOS link error (libstdc++ not available on modern macOS, replaced by libc++). The four new parametric test cases are small and surgical, mirroring the existing pattern for kimi_k25 and deepseek_v4; CI on Linux will exercise them.

Out of scope (explicit non-goals — file as follow-ups if desired):

• Native Rust chat-template formatter for Gemma 4 (analogous to lib/llm/src/preprocessor/prompt/deepseek_v4.rs). The vendored jinja covers Phase 1 via --custom-jinja-template; a native formatter is a several-day port of the 333-line jinja and not blocking.
• SGLang frontend wiring (components/src/dynamo/frontend/sglang_prepost.py). Only relevant for --dyn-chat-processor sglang (mode C). Whether SGLang has its own gemma4 parser is a separate question.
• Streaming-jail integration tests under lib/llm/tests/data/vllm/gemma4/. Those need real captured chunks from a Gemma 4 vLLM deployment; will follow up after a real-world capture.
• Backporting to older Dynamo release branches.

Watch list (upstream vLLM bugfix PRs the maintainers may want to track):

The vLLM Gemma 4 parsers are still settling — there are several open bugfix PRs touching internals (#39484 STRING_DELIM leak, #39311 nested-array bracket handling, #39070 argument truncation, #39588 selective reasoning, #39885 multi-turn streaming leak, #39570 chat-template HF resync). The public marker contract and parser name gemma4 are stable, so this PR is unaffected, but the vendored jinja may want a periodic re-sync against upstream once #39570 lands. Filing a tracking issue is recommended.

AI-Generated Code Disclosure:

Per Dynamo's contribution guide (docs/contribution-guide.md §AI-Generated Code): AI assistance was used to draft this change. Every changed line was manually reviewed against the upstream vLLM reference implementations (vllm/tool_parsers/gemma4_tool_parser.py, vllm/reasoning/gemma4_reasoning_parser.py, examples/tool_chat_template_gemma4.jinja) and against the analog Dynamo parsers (kimi_k2_parser.rs, granite_parser.rs, dsml/parser.rs). The submitter understands and can defend every line of the change end-to-end, including the recursive-descent grammar parser, the streaming thought\n prefix-stripping state machine, and the four upstream-diff fixes in commit 2.

Related Issues: (use one of the action keywords Closes / Fixes / Resolves / Relates to)

• Relates to GitHub issue: [CONTRIBUTION]: Add Gemma 4 tool-call and reasoning parsers #8851

---

Summary by CodeRabbit

• New Features

  • Added support for Gemma 4 models with tool-calling and reasoning parsing capabilities
  • New custom Jinja chat template for Gemma 4 tool formatting

• Documentation

  • Added documentation for Gemma 4 tool-calling and reasoning parsers
  • Added README guide for custom Jinja chat templates with Gemma 4 example

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[github-actions]

🌿 Fern Docs Preview: https://nvidia-preview-c0db9638-dc48-4efd-adfc-a8f8ee236842.docs.buildwithfern.com/dynamo/dev

[coderabbitai]

Walkthrough

This PR adds comprehensive Gemma 4 model support by introducing a tool-calling parser with custom <|tool_call> grammar, a channel-based reasoning parser for thinking outputs, a corresponding Jinja chat template, and infrastructure to register these parsers in the existing framework.

Changes

Cohort / File(s)	Summary
Documentation & Examples docs/agents/reasoning.md, docs/agents/tool-calling.md, examples/chat_templates/README.md, lib/parsers/README.md	Added documentation for new gemma4 reasoning and tool-calling parsers, including model family support, vLLM aliases, grammar details, configuration guidance, and updated parser registry counts (tool-calling: 17→18, reasoning: 14→15).
Chat Template examples/chat_templates/gemma4_tool.jinja	New Jinja template for Gemma 4 tool formatting, with macros for tool/function declarations, argument serialization, reasoning/thought channels, and tool response blocks. Handles multimodal content (text/image/audio/video) and streaming generation prompts.
Reasoning Parser Implementation lib/parsers/src/reasoning/gemma4_parser.rs	New Gemma4ReasoningParser struct extracting thinking content from `<
Reasoning Parser Infrastructure lib/parsers/src/reasoning/mod.rs	Added Gemma4ReasoningParser re-export, ReasoningParserType::Gemma4 enum variant, and registry entries for "gemma4" and "gemma-4" aliases.
Tool-Calling Parser Implementation lib/parsers/src/tool_calling/gemma4/parser.rs, lib/parsers/src/tool_calling/gemma4/mod.rs	New Gemma 4 tool-calling parser detecting `<
Tool-Calling Infrastructure lib/parsers/src/tool_calling/mod.rs, lib/parsers/src/tool_calling/config.rs, lib/parsers/src/tool_calling/parsers.rs	Added gemma4 module, ParserConfig::Gemma4 variant, ToolCallConfig::gemma4() factory, and registry dispatch for both "gemma4" and "gemma-4" aliases routing to Gemma 4 parser helpers.
Preprocessor Logic lib/llm/src/preprocessor.rs	Extended is_reasoning_disabled_by_request to recognize gemma4/gemma-4 aliases and consult chat_template_args["enable_thinking"] to determine reasoning disable state, with fallback to false when missing.

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~65 minutes

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately describes the main changes: addition of Gemma 4 parser support and test cases for the frontend.
Description check	✅ Passed	The PR description comprehensively covers the required template sections: Overview, Details, Where should the reviewer start, and Related Issues, with extensive technical depth and validation results.
Docstring Coverage	✅ Passed	Docstring coverage is 85.71% which is sufficient. The required threshold is 80.00%.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Review rate limit: 7/8 reviews remaining, refill in 7 minutes and 30 seconds.}

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

🧹 Nitpick comments (2)

lib/parsers/src/tool_calling/gemma4/parser.rs (1)

4-5: Incomplete reference link in comment.

The PR reference link on line 5 contains a placeholder (...) instead of the actual PR number. Consider either completing the link or removing it if the upstream PR hasn't been merged yet.

 // Reference implementations:
-// https://github.com/vllm-project/vllm/pull/.../vllm/tool_parsers/gemma4_tool_parser.py
+// https://github.com/vllm-project/vllm/blob/main/vllm/tool_parsers/gemma4_tool_parser.py

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@lib/parsers/src/tool_calling/gemma4/parser.rs` around lines 4 - 5, The inline
comment in parser.rs contains an incomplete PR URL
("https://github.com/vllm-project/vllm/pull/.../vllm/tool_parsers/gemma4_tool_parser.py");
update that comment by replacing the "..." with the actual PR number/path if
available, or remove the placeholder link entirely (keep or replace with a note
like "reference implementation in vllm/tool_parsers/gemma4_tool_parser.py" if
the PR is not merged) so the reference in the header comment of
lib/parsers/src/tool_calling/gemma4/parser.rs is accurate and not misleading.

lib/parsers/src/reasoning/gemma4_parser.rs (1)

102-122: Verify underflow safety in resolve_prefix.

Line 104 computes accum.len() - raw_reasoning.len(). While the current call sites ensure raw_reasoning is always a suffix of accum (pushed before calling), this invariant isn't documented or enforced at the function boundary.

Consider adding a debug assertion or a brief comment documenting this precondition:

📝 Suggested documentation

 /// Case 3: accumulated reasoning diverged from the prefix — emit the
 ///   buffered reasoning verbatim (data preservation).
+///
+/// # Precondition
+/// `raw_reasoning` must be a suffix of `accum` (i.e., `accum.len() >= raw_reasoning.len()`).
 fn resolve_prefix<'a>(accum: &'a str, raw_reasoning: &'a str) -> (&'a str, bool) {
+    debug_assert!(accum.len() >= raw_reasoning.len(), "raw_reasoning must be a suffix of accum");
     if accum.starts_with(THOUGHT_PREFIX) {

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@lib/parsers/src/reasoning/gemma4_parser.rs` around lines 102 - 122, The
function resolve_prefix relies on the precondition that raw_reasoning is a
suffix of accum (used when computing accum.len() - raw_reasoning.len()), so add
a debug assertion and brief comment documenting this invariant: insert something
like a comment above resolve_prefix mentioning "precondition: raw_reasoning must
be a suffix of accum" and a debug_assert!(accum.ends_with(raw_reasoning)) at the
start of resolve_prefix to make the precondition explicit and fail fast in debug
builds; reference resolve_prefix and the THOUGHT_PREFIX logic when adding the
assertion and comment.

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Nitpick comments:
In `@lib/parsers/src/reasoning/gemma4_parser.rs`:
- Around line 102-122: The function resolve_prefix relies on the precondition
that raw_reasoning is a suffix of accum (used when computing accum.len() -
raw_reasoning.len()), so add a debug assertion and brief comment documenting
this invariant: insert something like a comment above resolve_prefix mentioning
"precondition: raw_reasoning must be a suffix of accum" and a
debug_assert!(accum.ends_with(raw_reasoning)) at the start of resolve_prefix to
make the precondition explicit and fail fast in debug builds; reference
resolve_prefix and the THOUGHT_PREFIX logic when adding the assertion and
comment.

In `@lib/parsers/src/tool_calling/gemma4/parser.rs`:
- Around line 4-5: The inline comment in parser.rs contains an incomplete PR URL
("https://github.com/vllm-project/vllm/pull/.../vllm/tool_parsers/gemma4_tool_parser.py");
update that comment by replacing the "..." with the actual PR number/path if
available, or remove the placeholder link entirely (keep or replace with a note
like "reference implementation in vllm/tool_parsers/gemma4_tool_parser.py" if
the PR is not merged) so the reference in the header comment of
lib/parsers/src/tool_calling/gemma4/parser.rs is accurate and not misleading.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 7704350f-d1fc-436c-8990-6ab2d110a063

📥 Commits

Reviewing files that changed from the base of the PR and between 665b54b and c65e5da.

📒 Files selected for processing (13)

• docs/agents/reasoning.md
• docs/agents/tool-calling.md
• examples/chat_templates/README.md
• examples/chat_templates/gemma4_tool.jinja
• lib/llm/src/preprocessor.rs
• lib/parsers/README.md
• lib/parsers/src/reasoning/gemma4_parser.rs
• lib/parsers/src/reasoning/mod.rs
• lib/parsers/src/tool_calling/config.rs
• lib/parsers/src/tool_calling/gemma4/mod.rs
• lib/parsers/src/tool_calling/gemma4/parser.rs
• lib/parsers/src/tool_calling/mod.rs
• lib/parsers/src/tool_calling/parsers.rs

[the-david-oy]

Pushed f9c7a6d with two updates:

1. Rebased onto current main (was on 2e31245, now on 665b54b). The rebase was clean, no conflicts.
2. Addressed both CodeRabbit nits in a third commit chore(frontend): address CodeRabbit nits on Gemma 4 parsers:
   • parser.rs:4 — replaced the placeholder PR URL with the resolved upstream link and noted vLLM PR #38826 / #39027 are merged.
   • gemma4_parser.rs::resolve_prefix — made the "raw_reasoning is a suffix of accum" precondition explicit via doc-block + debug_assert!. Behavior unchanged in release builds.

Pre-commit job failure

The pre-commit check was failing on pre-existing breakage in main, not on anything in this PR. Specifically the pytest-marker-report hook reports 49 tests missing required markers across files I do not touch — most prominently components/src/dynamo/common/tests/test_video_protocol.py and test_video_utils.py, which were added 2 hours ago by #8491 (feat: formalize video streaming). Confirming this is upstream:

• gh run list --repo ai-dynamo/dynamo --branch main shows the most recent Pre Merge run on main commit 665b54b (the one I rebased onto) is also failure.
• The hook has pass_filenames: false in .pre-commit-config.yaml, so it always validates the full repo regardless of which files a given PR changed.
• All other CI checks on this PR are passing or pending: cargo clippy, DCO, copyright-checks, lychee, dco-comment, changed-files, pr_reminder, Validate PR title, rust-clippy (.) etc. all green.

This is due to:

• test: enforce pytest markers on components/src tests #7796 (test: enforce pytest markers on components/src tests) merged at 2026-04-29 18:06 UTC, making the pytest-marker-report hook strict for tests under components/src/.
• feat: formalize video streaming #8491 (feat: formalize video streaming) merged at 2026-04-29 18:45 UTC — its CI ran against the pre-strict hook, so it passed; once test: enforce pytest markers on components/src tests #7796 was merged 39 minutes earlier, feat: formalize video streaming #8491's tests retroactively became non-compliant.
• The same situation applies to test_audio_protocol.py (added by an earlier PR whose CI ran before test: enforce pytest markers on components/src tests #7796).

So three test files now block every external-contributor PR rebased onto main:

• components/src/dynamo/common/tests/test_video_protocol.py (25 tests)
• components/src/dynamo/common/tests/test_video_utils.py (7 tests)
• components/src/dynamo/common/tests/test_audio_protocol.py (17 tests)

[rmccorm4]

Thanks @the-david-oy ! CC @richardhuo-nv @keivenchang to help review

[rmccorm4]

/ok to test f9c7a6d

[keivenchang]

FYI I added comments on test-case categories (and pushed it). No logic change.

[keivenchang]

two notes — both pointers, no inline comments needed.

1. Pre-commit hook unrelated to this PR. pytest-marker-report currently fails on this branch with 49 missing-marker tests across test_audio_protocol.py, test_video_protocol.py, test_video_utils.py. those 49 are pre-existing on the branch — main has 0 missing. fix already merged on main as cc1af6a90 (#8849). merging main / rebasing should clear the hook.

2. CASE. coverage gaps to close (per lib/parsers/TEST_CASES.md).*

• To add: CASE.9 (paired reasoning + tool in same response), CASE.14 (empty tool_calls / null at API edge), CASE.15 (duplicate tool calls — universal gap, no parser covers this today).
• Framework-level (cross-parser suites, not in this PR's scope): CASE.11 (tool_choice modes), CASE.12 (finish_reason).
• N/A (worth a one-line // CASE.X — N/A per the doc's stated convention): CASE.23 (markers fixed, no variants), CASE.xml1/2 (not XML-family), CASE.harmony1 (not Harmony).

[keivenchang]

hey David, both round-1 comments addressed cleanly — regex-based extraction is the right shape, and the skip_special_tokens hook is a nice unprompted add. tiny future-facing nit inline.

can you also go through all the cases in lib/parsers/TEST_CASES.md and confirm coverage / mark explicit N/A for the ones that don't apply? i think 9, 14, 15 are still uncovered.

Almost there, thanks!

[richardhuo-nv]

/ok to test bd44405

[the-david-oy]

can you also go through all the cases in lib/parsers/TEST_CASES.md and confirm coverage / mark explicit N/A for the ones that don't apply? i think 9, 14, 15 are still uncovered.

Absolutely, done!

[richardhuo-nv]

/ok to test 54a5fca

[the-david-oy]

The failing docs link check is coming from the main branch, not any of the changes in this PR.

[keivenchang]

nice work David — all CASE gaps closed, N/A annotations match the doc convention. thanks, ship it. Oh, and that Doc link error, it is not a Mandatory check.

[the-david-oy]

Thank you! Are you able to merge? I don't think I have permission.