Recipe

nemotron-3-ultra-modelopt-mixed-rtxpro6000-vllm-tp4

nemotron-3-ultra-modelopt-mixed-rtxpro6000-vllm-tp4

Controller-backed candidate with an exact model revision and retained server-init proof of TP4, MTP5, FP8 KV, and full/piecewise graphs. Client timings survive only as controller summaries, so the index has not promoted or replayed them.

Record

Status: candidate
Source: 0xsero
Engine: vllm
Engine version: 0.22.0
Graph: full-and-piecewise
Accelerators: 4
Tensor parallel: 4
Context tokens: 261,056
Max concurrency: 1
KV cache tokens: 376,988
chat: yes
reasoning: yes
tools: yes
vision: no

Hugging Face model card

Identity

https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4

Repository: nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4
Status: known
Link type: Exact Hub repository

Public Hugging Face repository confirmed by the Hub API.

Candidate: useful compatibility or speed evidence. The registry does not offer Run until promotion requirements are met.

Controller configuration

vllm

No container · candidate · controller

Candidate evidence — not a Run contract

Environment

Variable	Value
`CUDA_DEVICE_ORDER`	`PCI_BUS_ID`
`CUDA_VISIBLE_DEVICES`	`GPU-fa982c97-64af-db6a-2ffb-08380e1f9375,GPU-c6ac75f2-cadf-6ff3-4cab-76c6033c1006,GPU-3cece4bc-432e-705e-7324-3f441d9cb4cc,GPU-7b5db8b3-3a49-c0a7-b8e4-80dc1bd3e853`
`ENABLE_CHUNKED_PREFILL`	`1`
`ENABLE_MAMBA_CACHE_STOCHASTIC_ROUNDING`	`1`
`ENABLE_MTP`	`1`
`ENABLE_PREFIX_CACHING`	`1`
`GLOO_SOCKET_IFNAME`	`lo`
`GPU_MEMORY_UTILIZATION`	`0.97`
`MAMBA_BACKEND`	`flashinfer`
`MAMBA_CACHE_PHILOX_ROUNDS`	`5`
`MAX_MODEL_LEN`	`262144`
`MAX_NUM_BATCHED_TOKENS`	`1024`
`MAX_NUM_SEQS`	`2`
`MOE_BACKEND`	`auto`
`NCCL_CUMEM_HOST_ENABLE`	`0`
`NCCL_DEBUG`	`WARN`
`NCCL_IB_DISABLE`	`1`
`NCCL_MNNVL_ENABLE`	`0`
`NCCL_NVLS_ENABLE`	`0`
`NCCL_P2P_DISABLE`	`1`
`NCCL_P2P_LEVEL`	`PIX`
`NCCL_SOCKET_IFNAME`	`lo`
`NVIDIA_TF32_OVERRIDE`	`1`
`OMP_NUM_THREADS`	`8`
`PYTORCH_CUDA_ALLOC_CONF`	`expandable_segments:True`
`RUN_FOREGROUND`	`1`
`SAFETENSORS_FAST_GPU`	`1`
`TORCH_NCCL_ASYNC_ERROR_HANDLING`	`1`
`TORCH_NCCL_BLOCKING_WAIT`	`1`
`VLLM_ALLOW_LONG_MAX_MODEL_LEN`	`1`
`VLLM_ALLREDUCE_USE_SYMM_MEM`	`0`
`VLLM_DISABLE_PYNCCL`	`0`
`VLLM_ENGINE_READY_TIMEOUT_S`	`1800`
`VLLM_FLASHINFER_ALLREDUCE_BACKEND`	`trtllm`
`VLLM_FLASHINFER_MOE_BACKEND`	`latency`
`VLLM_LOGGING_LEVEL`	`INFO`
`VLLM_SKIP_P2P_CHECK`	`0`
`VLLM_USE_FLASHINFER_MOE_FP4`	`1`
`VLLM_USE_FLASHINFER_MOE_FP8`	`1`
`VLLM_WORKER_MULTIPROC_METHOD`	`spawn`

Measured speed

Concurrency	Context	Prefill	Decode	TTFT ms	Status	Sweep
1	564	—	84.4	—	historical	nemotron-3-ultra-modelopt-mixed-rtxpro6000-vllm-tp4-sweep
1	32,768	—	102.4	—	historical	nemotron-3-ultra-modelopt-mixed-rtxpro6000-vllm-tp4-sweep
1	131,128	—	101.5	—	historical	nemotron-3-ultra-modelopt-mixed-rtxpro6000-vllm-tp4-sweep
1	131,128	—	76.5	—	historical	nemotron-3-ultra-modelopt-mixed-rtxpro6000-vllm-tp4-sweep
1	261,056	—	120.7	—	historical	nemotron-3-ultra-modelopt-mixed-rtxpro6000-vllm-tp4-sweep
1	261,056	—	81.7	—	historical	nemotron-3-ultra-modelopt-mixed-rtxpro6000-vllm-tp4-sweep

Remaining fields

Identity, launch, related records, and measured speed are shown above. This is the rest of the normalized record.

hardware count: 4
hardware id: rtx-pro-6000-blackwell-96gb
id: nemotron-3-ultra-modelopt-mixed-rtxpro6000-vllm-tp4
model instance id: nvidia-nvidia-nemotron-3-ultra-550b-a55b-nvfp4--nvfp4-mixed
recipe source: 0xsero
schema version: local-ai-registry/v1
speed sweep ids: nemotron-3-ultra-modelopt-mixed-rtxpro6000-vllm-tp4-sweep
status: candidate

capabilities

chat: Yes
reasoning: Yes
tools: Yes
vision: No

engine

graph mode: full-and-piecewise
name: vllm
version: 0.22.0

serving

kv cache tokens: 376,988
max concurrency: 1
max context tokens: 261,056
tensor parallel: 4

Provenance & metadata (3)

facts

metadata

provenance

captured at: 2026-08-27T06:04:13.773Z

sources

captured at	kind	url
2026-08-27T06:04:13.773Z	normalized-recipe	github.com/0xSero/local-ai-registry ↗