[mickqian]

Summary

• Trigger the NVIDIA multimodal generation CI from exact origin/main revision e9fe58139f48ec2a22fe4000ef604fc6862b1bfe.
• Collect H100, B200, and RTX 5090 performance logs.
• Tighten existing performance baselines from those measurements without increasing any existing value.

Method

The first CI run changes metadata only and therefore preserves all performance thresholds. After it completes, this PR will apply the same guardrail as #28123:

• update only fields observed in CI
• use min(existing, measured)
• do not increase any existing numeric baseline
• keep scenario and field sets stable
• leave unsupported or zero placeholders unchanged

Validation

• pre-commit passed for all three NVIDIA baseline files
• no local tests were run; GPU validation will come from the PR CI

---

CI States

Latest PR Test (Base): ❌ Run #31924624572
Latest PR Test (Extra): ❌ Run #31924635960

[mickqian]

Closing the bootstrap PR; performance sampling will use the dedicated multimodal_gen workflow on main instead.