Add chat template kwargs to tokenizer API

[jennyf19]

Summary

• add backward-compatible C and C++ tokenizer APIs for typed chat-template kwargs
• preserve the existing ApplyChatTemplate behavior by delegating with null kwargs
• pass model-specific context such as enable_thinking and reasoning_effort to ONNX Runtime Extensions
• add focused C API coverage for typed values and legacy equivalence

Dependency

This PR depends on microsoft/onnxruntime-extensions#1102. The production dependency pin is intentionally unchanged; this PR should remain draft until #1102 merges and ORT GenAI updates its normal Extensions pin.

Motivation

Foundry Local issue microsoft/foundry-local#808 needs a request path for:

{"chat_template_kwargs":{"enable_thinking":false}}

Qwen 3 defaults to reasoning and can exhaust bounded output budgets before producing an answer or tool call. Qwen 2.5 has no equivalent reasoning-template branch.

Validation

The code was stacked against the tested #1102 head only inside a disposable fork workflow. The native build compiled and CAPITests.ChatTemplate passed:
https://github.com/jennyf19/onnxruntime-genai/actions/runs/32071927312

No personal-fork dependency appears in this PR's production diff.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

Pull request overview

This PR extends ONNX Runtime GenAI’s chat templating pipeline to accept additional template context (“kwargs”) as a JSON object, exposing the capability through the C API and the zero-overhead C++ wrapper, and validating behavior via C API tests.

Changes:

• Add OgaTokenizerApplyChatTemplateWithOptions to the C API to pass template_kwargs (JSON) into chat template rendering.
• Update the C++ wrapper (OgaTokenizer) and internal Tokenizer implementation to route templating through the new “with options” path (with legacy behavior preserved by passing nullptr kwargs).
• Add/extend C API tests to validate kwargs rendering and parity with the legacy API when kwargs are nullptr.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
test/c_api_tests.cpp	Adds coverage for kwargs-based template rendering and legacy/null-kwargs equivalence.
src/ort_genai.h	Adds C++ wrapper ApplyChatTemplateWithOptions and forwards legacy ApplyChatTemplate to it.
src/ort_genai_c.h	Introduces the new exported C API function and documentation for template_kwargs.
src/ort_genai_c.cpp	Implements the new C API entrypoint and forwards the legacy entrypoint to it.
src/models/model.h	Adds internal Tokenizer::ApplyChatTemplateWithOptions declaration.
src/models/model.cpp	Implements internal kwargs-aware chat templating via OrtxApplyChatTemplateWithOptions.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[jennyf19]

Ready for review — Extensions leaf landed

microsoft/onnxruntime-extensions#1102 merged as bd0e21c11187e0b8a2385d1c61122a4d259a53a0 (merged by Sayan Shaw (@sayanshaw24)).

This PR now:

• Keeps the backward-compatible OgaTokenizerApplyChatTemplateWithOptions / C++ parallel API
• Pins cmake/deps.txt onnxruntime_extensions GIT_TAG to that merge commit (no personal-fork pin)
• Includes latest main merge

Stack

1. ✅ Extensions pip install onnxruntime-genai does not work #1102 merged
2. 👉 This PR (GenAI) — please review / undraft path
3. Foundry Local update version #1009 stays draft until a GenAI NuGet (Microsoft.ML.OnnxRuntimeGenAI.Foundry) includes this API (Baiju release train)

Evidence

• Motivating Foundry issue + Qwen 3 before/after: Add ability to modify chat template kwargs foundry-local#808
• Prior stacked fork validation (Extensions symbols + CAPITests.ChatTemplate): https://github.com/jennyf19/onnxruntime-genai/actions/runs/32071927312
• Official Extensions CI green on the merged tip: https://dev.azure.com/aiinfra/PublicPackages/_build/results?buildId=1373337

Follow-up (non-blocking, Extensions)

Sayan noted a future direction: fold chat-template kwargs into OrtxUpdateTokenizerOptions and eventually remove the per-call WithOptions surface. Dependents should stay thin adapters.

cc anyone owning GenAI releases / packaging for Foundry Local.

[jennyf19]

Looks like the red Windows CUDA x64 check is an infra flake, not a code failure:

The self-hosted runner lost communication with the server…

It died mid Build with build.py with no compile error. The twin windows-cuda-x64-build job on this same SHA is green, as are linux-cuda-x64, Windows CPU, lint, etc.

Please Re-run failed jobs when you have a moment — no source change needed from our side.

[Baiju Meswani (baijumeswani)]

/azp run macOS arm64 Build

[azure-pipelines]

Azure Pipelines:
Successfully started running 1 pipeline(s).