Parakeet RTX first-run startup guidance (NVBugs 6195601)

[kheiss-uwzoo]

Summary

• NVBugs 6195601 — Document that the Parakeet/Riva speech NIM does not ship prebuilt TensorRT profiles on RTX workstation SKUs and may need a long first-run engine build before the audio pod becomes Ready.
• Update the hardware support matrix: B200 remains unsupported for self-hosted audio (footnote 4); RTX Pro 6000 and RTX PRO 4500 show supported with footnote 5 operational guidance (profile build, default startupProbe behavior, kubectl wait up to 30m, do not submit audio jobs until Ready).
• Add a one-line pointer from the audio/video Helm section to matrix footnote 5.

Helm chart startupProbe defaults are wontfix for the Retriever Library team per QA disposition; this PR is documentation only on main.

Test plan

• Review prerequisites-support-matrix.md footnotes 1, 4, and 5 and audio matrix rows for RTX vs B200.
• Review audio-video.md Helm section link to footnote 5.
• Confirm PR diff is two markdown files only.

[greptile-apps]

Greptile Summary

This documentation-only PR adds first-run startup guidance for the Parakeet/Riva speech NIM on RTX workstation GPUs (RTX Pro 6000 Blackwell and RTX PRO 4500 Blackwell), documenting that these GPUs lack prebuilt TensorRT profiles and require a runtime engine build before the audio pod becomes Ready.

• Updates the hardware support matrix to move RTX Pro 6000 and RTX PRO 4500 from "Not supported⁴" to "1¹⁵" (supported with footnotes 1 and 5), scoping footnote ⁴ to B200 only.
• Adds new footnote ⁵ explaining the startupProbe / CrashLoopBackOff behavior during the engine build, with kubectl wait --timeout=30m guidance and a link to the Helm chart README for adjusting startupProbe.failureThreshold.
• Adds a one-line pointer in audio-video.md from the Helm deployment section to the new footnote ⁵.

Confidence Score: 5/5

Documentation-only change with no code or configuration modifications; safe to merge.

Both changed files are Markdown documentation. The hardware matrix updates and new footnote ⁵ are internally consistent: B200 remains unsupported, RTX workstation GPUs are now marked supported with appropriate caveats, and the kubectl wait / startupProbe guidance is technically accurate. The issues previously flagged in review threads (kubectl wait stdout behavior, link anchor precision, footnote ordering) are style/accuracy nits in prose that do not affect deployed software.

No files require special attention beyond the open review thread items already on record.

Important Files Changed

Filename	Overview
docs/docs/extraction/prerequisites-support-matrix.md	Table rows updated to mark RTX Pro 6000 and RTX PRO 4500 as supported (1¹⁵); new footnote ⁵ added with kubectl wait / startupProbe guidance; footnote ⁴ narrowed to B200 only. Footnotes now appear out of numerical order (¹, ⁴, ⁵, ², ³) and the kubectl wait output description is inaccurate — both flagged in prior review threads.
docs/docs/extraction/audio-video.md	Adds a one-line first-run guidance paragraph linking to the matrix footnote ⁵; the link target resolves to the section heading rather than the footnote anchor — flagged in a prior review thread.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Deploy Parakeet NIM via Helm] --> B{GPU type?}
    B -->|B200| C[Not supported — use hosted Parakeet or nimOperator.audio.enabled=false]
    B -->|RTX Pro 6000 / RTX PRO 4500| D[NIM starts runtime Riva/TensorRT engine build]
    B -->|H100 / H200 / A100 / A10G / L40S| E[May need runtime engine build per footnote 1]
    D --> F{startupProbe passes before build finishes?}
    F -->|No — CrashLoopBackOff| G[Treat as build in progress kubectl wait --timeout=30m]
    F -->|Yes| H[Audio NIM pod Ready]
    G --> H
    G -->|Still not Ready after 30m| I[Inspect logs kubectl logs deploy/audio Check Speech NIM support matrix]
    I --> J[Increase startupProbe.failureThreshold e.g. ~360 in Helm values]
    J --> K[Redeploy and wait again]
    K --> H
    E --> H

Loading

_{Reviews (3): Last reviewed commit: "docs(extraction): revert duplicate footn..." | Re-trigger Greptile}

[greptile-apps]

kubectl wait output description is inaccurate

The sentence "The command exits with code 0 when the pod is Ready and prints nothing until then" is incorrect — kubectl wait does print a line like pod/audio-xxxxx condition met to stdout on success. A reader who sees that output after being told to expect silence may think something went wrong.

Prompt To Fix With AI

This is a comment left during a code review.
Path: docs/docs/extraction/prerequisites-support-matrix.md
Line: 144

Comment:
**`kubectl wait` output description is inaccurate**

The sentence "The command exits with code 0 when the pod is Ready and prints nothing until then" is incorrect — `kubectl wait` does print a line like `pod/audio-xxxxx condition met` to stdout on success. A reader who sees that output after being told to expect silence may think something went wrong.

How can I resolve this? If you propose a fix, please make it concise.

[greptile-apps]

The link destination #model-hardware-requirements lands at the top of the "Model Hardware Requirements" section, not at footnote ⁵ itself. Since footnote ⁵ appears roughly 30 lines below that heading, users clicking the link must scroll to find the relevant guidance. Consider adding an explicit anchor (e.g. { #rtx-startup-guidance }) to footnote ⁵ in the matrix doc and updating this link to point to it.

Suggested change

	On first deploy, the Parakeet/Riva NIM may need a long Riva/TensorRT engine build (especially on RTX GPUs without a prebuilt profile). Wait until the audio NIM pod is Ready before you run extraction jobs—see [Pre-Requisites & Support Matrix — footnote ⁵](prerequisites-support-matrix.md#model-hardware-requirements).
	On first deploy, the Parakeet/Riva NIM may need a long Riva/TensorRT engine build (especially on RTX GPUs without a prebuilt profile). Wait until the audio NIM pod is Ready before you run extraction jobs—see [Pre-Requisites & Support Matrix — footnote ⁵](prerequisites-support-matrix.md#rtx-startup-guidance).

Prompt To Fix With AI

This is a comment left during a code review.
Path: docs/docs/extraction/audio-video.md
Line: 66

Comment:
The link destination `#model-hardware-requirements` lands at the top of the "Model Hardware Requirements" section, not at footnote ⁵ itself. Since footnote ⁵ appears roughly 30 lines below that heading, users clicking the link must scroll to find the relevant guidance. Consider adding an explicit anchor (e.g. `{ #rtx-startup-guidance }`) to footnote ⁵ in the matrix doc and updating this link to point to it.

```suggestion
On first deploy, the Parakeet/Riva NIM may need a long Riva/TensorRT engine build (especially on RTX GPUs without a prebuilt profile). Wait until the audio NIM pod is Ready before you run extraction jobs—see [Pre-Requisites & Support Matrix — footnote ⁵](prerequisites-support-matrix.md#rtx-startup-guidance).
```

How can I resolve this? If you propose a fix, please make it concise.