[krishung5]

Overview

Backport OpenEngine sidecar audio support from #16871 to release/1.7.0. Audio URLs and data URIs are forwarded to audio-capable TensorRT-LLM engines, including text-only input models such as Whisper.

Key Changes

Approximately 122 core added lines, excluding tests, documentation, and CI configuration/harness code.

• Advertise and enforce audio/input capabilities, preserving token-ID behavior for older servers and sending the formatted prompt to text-only models.
• Preserve ordered media sources and caller UUIDs through the existing OpenEngine request mapping.
• Add the Whisper launch configuration and audio coverage to the existing official TRT-LLM upstream-nightly integration workflow.

Validation

• Clean git cherry-pick -x --signoff of main commit bb160083f62edb30c3bcfbb8fd7c4af1ecb09d68; patch ID matches the merged source.
• All 22 CPU nightly-selection, launch, and media-harness tests passed on the release branch.
• The nightly audio test explicitly executes the real OpenEngine test and checks the expected Whisper transcription and authoritative token usage. GPU validation has not been rerun for this backport; this test covers the sidecar-to-engine path, rather than full HTTP frontend E2E.

Audio serving requires an audio-capable upstream TensorRT-LLM runtime. Existing image/video coverage is retained; the Whisper launch configuration uses --migration-limit 0.

Issues and Ongoing Work

Refs DIS-3339

Source: #16871 (merged to main). This PR carries only that audio change; image-aware routing will be backported separately after #17437 merges.

Pull Request Checklist

• I used Dynamo's /pr-checklist skill.