RTX 5090 Omni stack qualification

Point-in-time record, 2026-07-27. The Primary Node RTX 5090 topology now uses one exclusive Nemotron 3 Nano Omni tier for auxiliary text, general image understanding, and OCR. This replaces three independently resident model containers; it does not replace embeddings, reranking, voice audio, or ComfyUI. Those remain separate, mutually exclusive stacks on the same GPU.

Qualified identity

Field	Value
Checkpoint	nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4
Revision	dc5f0b0bfddf8b6e0f5891475be9af05b80126fe
Served name	nemotron3-omni-30b-a3b-nvfp4
Engine	vLLM 0.23.1rc1.dev531+ga65f93fb2
Image	vllm/vllm-openai:nightly@sha256:907377dddef392f6b679d9c071e1c33c3935b4dc993b61d0352e391a5319ff3e
Managed serve	omni, container vllm-nemotron3-omni, port 30003
Hardware	NVIDIA GeForce RTX 5090, 32,607 MiB
Context / concurrency	65,536 tokens / 2 sequences
Router aliases	llm.voice, vision.general, vision.ocr -> omni-local

The controller-managed production container reached /health with HTTP 200. Observed GPU use was 27,706 MiB after startup. The reservation owns the full 27,999 MiB usable budget (32,607 MiB capacity less the 4,608 MiB system/audio reserve), so the stack is intentionally exclusive.

Gates

The text gate ran with thinking disabled and passed smoke, JSON, 4K retrieval, and three tool calls. The new multimodal preflight checks used the bounded, content-addressed PNG input already published with the vision evidence. General image understanding found all of RTX 5090, Error 503, and Retry; the OCR check reproduced all of Anvil Serving Dashboard, VRAM used: 30057 MiB, No available tier, and planning.

The bounded capacity run completed 6/6 requests at concurrency two with a 2,048-token prompt plan and 128-token output cap. TTFT p50/p95 was 122/164 ms, end-to-end p50/p95 was 423/515 ms, and aggregate output throughput was 224.08 tok/s. This is a small operational capacity probe, not a model-quality comparison.

Raw evidence:

• Multimodal preflight
• Capacity probe
• Routed auxiliary text
• Routed general vision
• Routed OCR

Operational result and caveats

omni-stack and llm-stack can start Omni; auxiliary-stack starts only embeddings and reranking. ComfyUI retains its independent project and can use the guarded --evict flow to quiesce and drain omni-local before taking its on-demand reservation. Voice remains a separate lifecycle and is covered by the 4,608 MiB non-ledger reserve.

The engine is pinned to a nightly digest because this architecture is not supported by the older stable/NGC images tested previously. Do not replace the image tag, digest, or checkpoint revision without rerunning text, tools, image, OCR, context, and capacity gates. Audio input was not tested here; “Omni” qualification in this record means auxiliary text plus image/OCR only.

The deployed router began with the legacy four-tier ID set. The guarded router install-config verb quiesced and drained that current set, used the existing atomic volume-write/restart/rollback implementation, then verified that only ready primary-local and omni-local tiers remained.