[neilkg]

Summary

Beam does not support non-thinking generation, but a request with chat_template_kwargs: {"enable_thinking": false} currently makes Dynamo move the model's analysis into answer content and report zero reasoning tokens. thinking: false has the same effect.

Reject an effective client request to disable Beam reasoning with an actionable HTTP 400 before prompt rendering or backend dispatch. Reuse the existing alias normalization, including top-level thinking and reasoning_effort: "none", without introducing another switch or changing precedence. Validate before deployment defaults so the error describes the client's request. Omitted effort and reasoning_effort: null remain accepted; the string "none" requests reasoning off.

Use a shared reasoning_is_mandatory model policy for both request validation and response reconstruction; Beam is the only family registered by this change. The generic validator reuses the existing family selection and thinking-alias normalization. Reconstructed Beam responses retain their analysis/content split and reasoning-token accounting even if an older request bypassed admission validation.

Other model families and Beam's existing history behavior remain unchanged: supplied reasoning is preserved, preserve_thinking=false retains its compatibility behavior, and the default for allow_missing_thinking is unchanged. Requiring reasoning in every historical assistant message is a separate serving policy from rejecting an explicit request to disable future reasoning.

Targets luckyq/beam-1.0-release, the Beam integration branch used by #16702. This is a frontend fix and does not require changes to the inference engines or parser crate.

Closes #16783.

Validation

• cargo test -p dynamo-llm --no-default-features --lib beam: 24 passed.
• cargo test -p dynamo-llm --no-default-features --lib reasoning_effort: 5 passed.
• cargo test -p dynamo-llm --no-default-features --lib reasoning_usage: 4 passed.
• Two focused regressions cover streaming/non-streaming admission, reasoning-off aliases, the HTTP 400 mapping and public error message, response-policy selection, history preservation, omitted/null effort, and another model's supported non-thinking mode. Existing Beam parser and shared reasoning-token accounting tests provide the parser/usage coverage.
• cargo fmt --all -- --check, changed-file pre-commit hooks, and git diff --check: passed.
• GPU/container end-to-end validation was not rerun; these checks exercise the real frontend preprocessor, parser, response policies, and local mock model metadata without loading model weights.

[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 neilkg! 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.

🚀