fix: honor skip_special_tokens from extra_body on the vLLM completion path

[DongjiGao]

What

The vLLM text-completion request builder hardcoded skip_special_tokens: False at the top level, while _build_request_body could simultaneously carry a user-supplied skip_special_tokens inside extra_body. The field was sourced twice, so any extra_body.skip_special_tokens was silently dropped on the completion path.

This pops skip_special_tokens out of extra_body into the single top-level field (default False), so inference.extra_body.skip_special_tokens is honored on the completion path — matching how the chat path already behaves.

Why

skip_special_tokens=false keeps special tokens (e.g. speaker tags) in the decoded output, which multispeaker ASR needs. The chat endpoint already respects this via extra_body; this makes the completion endpoint consistent, without adding any new config surface.

Usage (just extra_body, no new options):

++inference.extra_body.skip_special_tokens=false

Test

tests/test_vllm_completion.py:

• default → top-level skip_special_tokens=False, absent from extra_body
• extra_body.skip_special_tokens=True → lifted to the single top-level field, not duplicated

[coderabbitai]

📝 Walkthrough

Walkthrough

Refactor vLLM completion request parameter building to extract skip_special_tokens from merged extra_body payload to the top-level request field, replacing hardcoded defaults. Add unit tests verifying default behavior and override lifting.

Changes

vLLM Detokenization Parameter Support

Layer / File(s)	Summary
vLLM completion request builder refactoring nemo_skills/inference/model/vllm.py	_build_completion_request_params now builds extra_body into a local variable, extracts skip_special_tokens via pop() with False default, and uses the modified payload as request extra_body instead of inlining _build_request_body(...) and hardcoding skip_special_tokens to False.
Completion request builder unit tests tests/test_vllm_completion.py	New test module with _completion_request helper that mocks VLLMModel construction and calls _build_completion_request_params directly. Tests verify skip_special_tokens defaults to False when absent and is properly lifted to top-level without duplication when provided in extra_body.

🎯 2 (Simple) | ⏱️ ~10 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 33.33% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title accurately and specifically describes the main change: honoring skip_special_tokens from extra_body on the vLLM completion path, which aligns with the core modification in VLLMModel._build_completion_request_params.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[Kipok]

@DongjiGao please recreate from a branch

[DongjiGao]

Recreated as in-repo branch PR #1488 so CI can run (fork PRs don't receive CI secrets). Closing this fork PR in favor of #1488.