[ptarasiewiczNV]

Summary

Cherry-pick #16295 onto release/1.6.0 using git cherry-pick -x --signoff (main commit e09ec90990cd1ebb763fd27858c87e1fe7237172). The cherry-pick applied without conflicts or adaptations.

Change the default realtime ASR example to Qwen/Qwen3-ASR-0.6B, using vLLM's realtime support. This gives the release a default example that boots and transcribes audio, avoiding the reported Voxtral startup failure in the default launch path.

• Automatically select Qwen3ASRRealtimeGeneration for the default model when no --hf-overrides is supplied.
• Preserve explicit overrides and require an explicit realtime architecture for other models.
• Update the recipe and voice overview to describe the default and custom-model commands.

The change is limited to one example launcher and two documentation pages. It does not fix the underlying Voxtral startup failure. No additional backport dependencies were needed.

Validation

• On this release branch: pre-commit checks for all three changed files, bash -n, git diff --check, and invalid-argument checks passed. Scoped docs lint passed with zero errors and five existing navigation warnings for untouched pages.
• fern check: zero errors and three warnings. fern docs broken-links reports an existing broken link in the untouched deepseek-v4-pro-0813.mdx page (line 254); the page and navigation are unchanged by this backport.
• All three cherry-picked files exactly match the merged main commit.
• Prior GPU validation of the same launcher in nvcr.io/nvstaging/ai-dynamo/vllm-runtime:1.6.0-rc.2 on an RTX A6000: image packages unchanged; default model and architecture selection; worker booted, registered, and remained healthy after three audio requests.
• That RC2 run used --max-model-len 4096 --max-num-seqs 2 --kv-cache-memory-bytes 1073741824. A paced 6.72-second clip emitted its first delta at 5.48 seconds, before commit at 6.83 seconds, and completed at 7.11 seconds. The documented client passed --require-incremental; residual-only audio also completed without a crash.
• The GPU run was performed for the main PR; it has not been rerun for this conflict-free backport.

Known Upstream Limitations

Transcript repetition and raw ASR markers also reproduced on native vLLM in the original validation. This backport does not address those upstream behaviors:

• Repetition: vllm-project/vllm#51374; proposed context-reset fix vllm-project/vllm#56218.
• Raw ASR markers: vllm-project/vllm#35767; proposed post-processing fix vllm-project/vllm#55786.

Refs DYN-4860.

[github-actions]

🌿 Fern Docs Preview: https://nvidia-preview-01a13c37-8c2a-727e-8315-61ce9a755f19.docs.buildwithfern.com/dynamo/dev

[devin-ai-integration]

✅ Dynamo PR CI passed — run 38636207227 (attempt 1) on 193fa4e193

Gate checks: ✅ backend-status-check · ✅ deploy-status-check · ✅ dynamo-status-check

_{Posted automatically by Devin for run 38636207227. Updated on every full-CI run of this PR.}