[Gemma4] Updated chat template, reasoning property

[bbrowning]

This documents the updated chat template to use with Gemma 4 models for reasoning and/or tool calling that was merged in
vllm-project/vllm#39027 .

It also adds instructions for how to enable thinking by default, if a user prefers to always think.

And, it replaces the deprecated reasoning_content field with the updated reasoning field.

[gemini-code-assist]

Code Review

This pull request updates the Gemma 4 documentation to reflect the change from reasoning_content to reasoning in the API response and adds instructions for using the optimized chat template for reasoning and tool calling. A formatting issue was identified in the key metrics table where a row was missing its closing pipe and third column, which would cause rendering errors.

[faradawn]

Hey @bbrowning, thanks for the PR! Do you know which version of vLLM incorporates this change? So that we can add a
▎ Note: The reasoning field requires vLLM >= X.Y.Z. On older versions, use reasoning_content.

[faradawn]

We have just merged a pull request, can you update your branch? Then It looks great to me.

[bbrowning]

Hey @bbrowning, thanks for the PR! Do you know which version of vLLM incorporates this change? So that we can add a ▎ Note: The reasoning field requires vLLM >= X.Y.Z. On older versions, use reasoning_content.

@faradawn vLLM v0.16.0 moved to reasoning instead of reasoning_content. But, Gemma 4 models weren't supported until vLLM v0.19.0. So, I think in the context of the Gemma 4 docs, there's no reason to ever mention reasoning_content as an input/output field. Anyone using Gemma 4 models is on a new enough vLLM to use reasoning instead.