fix(sglang): inherit reasoning_parser and tool_call_parser in LoRA registration

[julianStreibel]

Summary

Backport of the LoRA runtime_config fix (already on main via commit 8957ec3 / PR #12092) to the release/1.4.2 branch.

Problem

The SGLang backend's load_lora handler publishes adapter ModelDeploymentCards without runtime_config. The adapter card has no reasoning_parser or tool_call_parser, so the Rust frontend postprocessor does not:

• Split <think>...</think> blocks into reasoning_content
• Parse <tool_call> blocks into structured tool_calls

for adapter-name requests. Base model requests work correctly because the base worker's card carries the parsers.

Fix

Call _get_runtime_config() before register_llm() in the load_lora method to construct a ModelRuntimeConfig with the worker's --dyn-reasoning-parser and --dyn-tool-call-parser settings. This mirrors:

• The vLLM backend fix in PR fix: inherit tool_call_parser and reasoning_parser in LoRA model registration #7559
• The same fix already merged to main in PR fix(frontend): reject over-context requests before streaming #12092

Changes

• components/src/dynamo/sglang/request_handlers/handler_base.py:
  • Lazy-import _get_runtime_config from dynamo.sglang.register
  • Build runtime_config from the engine and dynamo args
  • Pass runtime_config=runtime_config to register_llm()

Fixes #14862

[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]

👋 Hi julianStreibel! Thank you for contributing to ai-dynamo/dynamo.

Just a reminder: The NVIDIA Test Github Validation CI runs an essential subset of the testing framework to quickly catch errors.Your PR reviewers may elect to test the changes comprehensively before approving your changes.

🚀

[rmccorm4]

Hi @julianStreibel, thanks for raising this. The 1.4.2 release is already shipped, so we won't backport new changes to it. However, the v1.5.0 release should be coming out very soon, which contains this fix - the recommended path forward is to upgrade to v1.5.0 to pickup this fix, or install a nightly wheel or build from source for custom builds as needed.