[26.04 POR] Benchmarking for Audio: stage-wise profiling across modalities

Goal: Stage-wise profiling across modalities (Text / Audio / Image / Video / Reasoning) to identify bottlenecks and prioritize optimizations.

Acceptance (draft)

Define profiling methodology + metrics per stage.
Produce baseline results per modality.
Summarize top bottlenecks + proposed fixes.

Checklist:
Release Readiness Checklist
Benchmarking and Performance

• Add benchmarking scripts to the repository and integrate them into the regular CI matrix (https://github.com/NVIDIA-NeMo/Curator/issues/1563).
• Run benchmarks for all modality pipelines and capture performance results.

Security and Dependencies

• Review vulnerability scans for CVEs.
• If dependency updates are required for a modality (due to CVEs or breakage), the modality PIC must help resolve and unblock the release.
• Validate that all open-source packages comply with licensing policies.
• For any packages or binaries with unknown or unclear licensing, the modality PIC must coordinate with Legal to verify dependencies and ensure release readiness.
• Validate that all NVIDIA internal packages included in the container are authorized and correctly versioned.

Quality and Documentation

• Identify and address all release-blocking bugs.
• Update all relevant documentation and release notes.

Visibility and Communication

• Maintain clear visibility of work status to stakeholders and the project team.

Ref: https://nvidia.atlassian.net/wiki/spaces/NLLMS/pages/3142129441/NeMo+Curator+Release+Readiness#Modality-PIC-Responsibilities