[Gemma4] Fix chat template and stop tokens for OpenAI tool calling compatibility

[lucianommartins]

[Gemma4] Fix chat template and stop tokens for OpenAI tool calling compatibility

What does this PR do?

Rewrites the _patch_template_for_openai_tool_role() function in convert_gemma4_weights.py to fully support OpenAI Chat Completions tool-calling semantics for Gemma4 (E4B and 31B).

Chat template patcher

• Forward-scan tool rendering: role: "tool" messages are skipped in the outer loop and rendered proactively as <|tool_response> blocks from the preceding assistant turn that issued the tool_calls
• Turn suppression: Suppresses duplicate <|turn>model when consecutive assistant messages are separated only by tool messages (multi-round tool-call loops)
• tool_call_id resolution: Matches tool results back to the originating tool_calls array by ID to resolve function names correctly (prevents response:unknown)
• Content-parts robustness: Handles tool response content as both plain strings and OpenAI content-parts arrays ([{type: "text", text: "..."}])
• format_tool_response_block macro: Injects a reusable macro to centralize tool response rendering (used by both legacy Gemma native tool_responses and OpenAI-style role: "tool" paths)
• reasoning/reasoning_content support: Renders thinking fields as <|channel>thought blocks (compatible with vLLM, DeepSeek, and o1-style inference servers)
• Legacy compat: Preserves native tool_responses on assistant messages (Google/Gemma format)

Stop tokens (eos_token_id)

• Removed <tool_call|> (etc_token) from the stop token list
• Keeps only <eos> + <turn|> (eot_token)
• Enables parallel tool calls without premature truncation after the first <tool_call|>; <turn|> still terminates the model turn correctly

Testing

Validated with 17 functional test scenarios across both E4B and 31B templates:

• Simple chat, tool declarations, single/multi/parallel tool calls
• Multi-round tool loops (exactly 1 <|turn>model emitted)
• Legacy tool_responses, tool_call_id resolution, content-parts arrays
• reasoning/reasoning_content field rendering
• add_generation_prompt correctness, Jinja2 syntax validation

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline, Pull Request section?
• Was this discussed/approved via a Github issue or the forum?
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

Who can review?

Models:

• multimodal models: @zucchini-nlp

Library:

• generate: @zucchini-nlp (visual-language models)

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: gemma4

[zucchini-nlp]

cc @Rocketknight1 for chat templates and tool calling, but this seems to be only the conversion. Prob the latest changes you made before release, didn't make it to conversion script 😅

[Rocketknight1]

The template updates seem good, but I think it'd make more sense and be easier to review if we just embedded the entire file in this script, rather than all the replacements, since I think the size of the replacements approaches the size of the original file! That would mean we're not dependent on the files in the model repos too.

[Rocketknight1]

Also linking an additional bug report in the template: #45331

[RyanMullins]

@lucianommartins and I pushed chat template updates directly to the Hub yesterday, so I think these changes are no longer needed and we can probably close this.