[Bugfix][Higgs-Audio-V3] Propagate Sampling Parameters to Sampler

[sphinxkkkbc]

PLEASE FILL IN THE PR DESCRIPTION HERE.

Purpose

Fix #5387

Enable per-request sampling parameters for Higgs Audio v3 speech generation.

• Promotes temperature, top_p, and top_k from speech request extra_params into the stage-0 SamplingParams.
• Allows vLLM to propagate these values through InputBatch into SamplingMetadata.
• Passes SamplingMetadata to the Higgs Audio v3 codebook sampler.
• Expands request-level sampling parameters from [batch] to [batch * num_vq], so every VQ logits row uses the parameters of its owning request.

Test Plan

• Validate speech request sampling parameters are propagated through stage-0 SamplingParams and SamplingMetadata to the model-owned sample() method.
• Validate Higgs Audio v3 correctly applies per-request temperature, top_p, and top_k across batched requests and their VQ logits, top_k=-1 and invalid_temperature test

pytest -v tests/unit/higgs_audio_v3/test_higgs_audio_v3.py
pytest -v tests/worker/test_gpu_ar_model_runner.py 

vLLM Version:
0.25.0
vLLM-Omni Commit:
b57b76a

Test Result

• test_gpu_ar_model_runner: 21 passed
• test_higgs_audio_v3: 77 passed

BEFORE SUBMITTING: read CONTRIBUTING.md and run the precheck-pr skill with the code agent for a self-check against project conventions.
(anything written below this line will be removed by GitHub Actions)

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[Gaohan123]

top_k=-1 is commonly used by vLLM callers to mean “disable top-k filtering,” but clamp(min=1, ...) converts it to 1. That makes audio-code sampling greedy for those requests, rather than leaving the distribution unrestricted.

Could we preserve the no-top-k sentinel before calling apply_top_k_top_p—for example, map top_k <= 0 to x.shape[-1] (or the helper’s documented disable value)? Please also add a regression test covering top_k=-1, ensuring more than one candidate remains sampleable.

[sphinxkkkbc]

fixed, test added and descriptions updated

[Gaohan123]

LGTM. Thanks