[None][doc] Gemma 4: usage examples

[Hudayday]

Summary

Append a new "Run Gemma 4" section to examples/models/core/gemma/README.md
covering the PyTorch backend flow that we actually develop and test against:

• variant table (E2B / E4B / 26B-A4B / 31B with modalities)
• trtllm-serve (OpenAI-compatible REST) launch + curl /v1/chat/completions
• quickstart_multimodal.py for image and audio inputs (E2B / E4B only)
• trtllm-eval mmmu and trtllm-eval covost2 entry points

Gemma 4 runs on the PyTorch backend with HF checkpoints loaded directly;
the legacy convert_checkpoint.py / trtllm-build flow is intentionally
omitted since it does not apply to Gemma 4.

Also registers the new section and its three subsections in the
top-of-file Table Of Contents.

Summary by CodeRabbit

• Documentation
  • Added detailed Gemma 4 documentation including deployment setup, OpenAI-compatible API integration with usage examples, multimodal inference capabilities for image and audio processing, accuracy benchmarking and evaluation tools with sample commands, practical quick-start examples, and guidance on supported model variants.

[coderabbitai]

📝 Walkthrough

Walkthrough

This PR adds documentation for Gemma 4 model inference and evaluation. A new "Run Gemma 4" section describes launching trtllm-serve with Gemma 4 checkpoints, querying the OpenAI-compatible API, running multimodal inference, and evaluating accuracy using trtllm-eval on MMMU and CoVoST 2 benchmarks. The Table of Contents is updated to reflect these subsections.

Changes

Gemma 4 Documentation

Layer / File(s)

Summary

Gemma 4 server and API documentation examples/models/core/gemma/README.md

Table of Contents updated with new "Run Gemma 4" entry and sub-links. New section documents supported Gemma 4 HuggingFace variants/modalities, provides trtllm-serve launch examples using PyTorch backend, curl chat-completions query examples with internal chat-template handling, multimodal inference commands for image/audio via quickstart script, and trtllm-eval benchmark commands for MMMU and CoVoST 2 evaluation.

🎯 1 (Trivial) | ⏱️ ~5 minutes

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description check	✅ Passed	The description explains the change clearly (new README section for Gemma 4) and provides sufficient detail about what is being added.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
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.
Title check	✅ Passed	The title '[None][doc] Gemma 4: usage examples' is directly related to the main change, which is adding usage examples for Gemma 4 to the README documentation.

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

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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

[coderabbitai]

Actionable comments posted: 2

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@examples/models/core/gemma/README.md`:
- Around line 853-854: The README's MMMU support line incorrectly lists
text-only checkpoints (26B-A4B and 31B) as supported for this vision
multiple-choice task; update the MMMU support matrix so it only lists
multimodal-capable checkpoints (e.g., E2B / E4B) and remove or mark 26B-A4B and
31B as "text-only / not supported" in the MMMU entry so the modality claim is
consistent with the earlier note about text-only checkpoints.
- Around line 833-845: The commands in the README invoke
quickstart_multimodal.py using a repo-root-relative path which will fail when
run from examples/models/core/gemma/; update the examples to either prepend the
correct relative path to quickstart_multimodal.py or add an explicit instruction
to cd to the repository root before running. Specifically edit the README
entries that reference quickstart_multimodal.py (the multimodal example lines)
to use a correct relative path from examples/models/core/gemma/ to the script or
add a preceding "cd" instruction so the quickstart_multimodal.py invocation
resolves correctly.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: bbb42d75-0cae-47d7-8bfb-b07cfcff52dd

📥 Commits

Reviewing files that changed from the base of the PR and between 989671b and a704288.

📒 Files selected for processing (1)

• examples/models/core/gemma/README.md

[Hudayday]

/bot skip --comment "document updates, all commands tested"

[tensorrt-cicd]

PR_Github #49345 [ skip ] triggered by Bot. Commit: f5383b8 Link to invocation

[tensorrt-cicd]

PR_Github #49345 [ skip ] completed with state SUCCESS. Commit: f5383b8
Skipping testing for commit f5383b8

Link to invocation