nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4

completed
Request throughput
9.65 req/s
Output throughput
2,471 tok/s
p95 TTFT
1051.3 ms
p95 TPOT
12.55 ms

Manifest warnings

Review these published caveats before interpreting or comparing the metrics.

  • Review imported facts against the local model, hardware, launch command, and benchmark command before publication.

Server launch command

CUDA_VISIBLE_DEVICES=0 vllm serve nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 \
  --revision e8f3c7c4de75ad84fe1bcef95d38eca76214480b \
  --port 9000 \
  --host 127.0.0.1 \
  --dtype bfloat16 \
  --gpu-memory-utilization 0.92 \
  --max-model-len 9216 \
  --max-num-batched-tokens 16384 \
  --max-num-seqs 64 \
  --tensor-parallel-size 1 \
  --pipeline-parallel-size 1 \
  --kv-cache-dtype fp8 \
  --distributed-executor-backend mp \
  --no-enable-prefix-caching \
  --enable-chunked-prefill \
  --async-scheduling \
  --no-enforce-eager \
  --served-model-name nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 \
  --disable-uvicorn-access-log \
  --mamba-backend flashinfer \
  --mamba-cache-mode align

Run details

Workload
1024 → 256
Methodology
closed loop · streaming
Test load
32 concurrent requests
Endpoint
openai-compatible-completions
Requests
16 unmeasured warmups per invocation after one readiness request.
Dataset
sha256=0e0945825728f70eee1188087607793934a8cbde9fe12a4cc0c6192429ce7fd3 generator=runpile-sweeps-tokenizer-verified-jsonl synthetic=true generator_revision=f08cb0bccfbd93f7ea4cee52d5b921a8250619c0ed7abd572cd2159a829f7d28 shared_prefix_tokens=0
Sampling
mode=greedy temperature=0
Stopping
ignore_eos=true max_output_tokens=256
Prefix and cache
cache_state=warm-host cache_enabled=false shared_prefix_tokens=0
Client
name=vllm bench serve backend=vllm version=0.28.0 placement=same-host
Hardware
1× NVIDIA GeForce RTX 5090
Runtime
vllm 0.28.0
Runtime configuration
dtype=bfloat16 max_num_seqs=64 enforce_eager=false max_model_len=9216 kv_cache_dtype=fp8 prefix_caching=false chunked_prefill=true expert_parallel=false async_scheduling=true gpu_memory_utilization=0.92 max_num_batched_tokens=16384
Measurement
vllm bench serve 0.28.0
Timing
client-observed streaming request latency · client overhead included
Repetitions
1
Failure handling
Failed requests counted; no client retries.
Quantization
format=W4A16 NVFP4 mixed precision checkpoint=modelopt
Model revision
e8f3c7c4de75ad84fe1bcef95d38eca76214480b
Evidence
Raw benchmark evidence attached · eligible and ranked
Published
by anonymous-EVjB2S

Benchmark command

vllm bench serve \
  --backend vllm \
  --base-url http://127.0.0.1:9000 \
  --endpoint /v1/completions \
  --model nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 \
  --tokenizer nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 \
  --dataset-name custom \
  --dataset-path <EXACT_SYNTHETIC_DATASET> \
  --custom-output-len 256 \
  --skip-chat-template \
  --disable-shuffle \
  --num-prompts 256 \
  --max-concurrency 32 \
  --request-rate inf \
  --seed 0 \
  --ignore-eos \
  --temperature 0 \
  --percentile-metrics ttft,tpot,itl,e2el \
  --metric-percentiles 50,90,95,99 \
  --disable-tqdm \
  --save-result \
  --result-dir <local-path> \
  --result-filename nemotron35-lightning-30b-a3b-nvfp4-tp1__baseline__serve-fixed-1k-256-v1__c32__r2__attempt1.json \
  --num-warmups 16 \
  --metadata campaign=runpile-rtx5090-remaining-models-20260827-2 config_id=nemotron35-lightning-30b-a3b-nvfp4-tp1 server_variant=baseline workload_key=serve-fixed-1k-256-v1 repetition=2 aggregate_only=true