{"id":"run_DapqPLH83TG71M1Y","owner":"anonymous-TWTejS","publishedAt":"2026-08-23T23:38:46.150Z","manifest":{"schema_version":"1.0","kind":"run","client_run_id":"vllm-b7ed86ed158a-1","name":"NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 | 4x RTX 5090 | 1024->256 | C128","notes":"Text-only synthetic serving measurement. Exact token lengths, greedy decoding, streaming responses, and EOS ignored. No prompts, generated text, request bodies, or response bodies are retained in the aggregate artifact.","visibility":"public","license":"CC-BY-4.0","outcome":"completed","subject":{"type":"model","source":"huggingface","adapters":[],"revision":"e8f3c7c4de75ad84fe1bcef95d38eca76214480b","identifier":"nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4","modalities":["text"],"architecture":"NemotronHForCausalLM","quantization":{"format":"NVFP4","checkpoint":"modelopt"},"modifications":[],"revision_kind":"commit","parameter_count_total":30000000000,"parameter_count_active":3000000000},"environment":{"os":"Ubuntu 24.04.4 LTS","cpu":{"model":"AMD Ryzen Threadripper PRO 9955WX 16-Cores","sockets":1,"numa_nodes":1,"logical_cpus":32,"physical_cores":16},"kernel":"6.8.0-136-generic","topology":"Four independent one-GPU replicas, one per GPU, behind the recorded round-robin HTTP proxy.","accelerators":[{"count":4,"model":"NVIDIA GeForce RTX 5090","driver":"580.173.02","vendor":"NVIDIA","architecture":"Blackwell SM120","interconnect":"PCIe within one NUMA node; no NVLink","memory_bytes_per_device":34190917632}],"driver_versions":{"nvidia":"580.173.02","cuda_compiler":"Build cuda_13.1.r13.1/compiler.37061995_0"},"host_architecture":"x86_64","system_memory_bytes":269585743872},"runtime":{"engine":"vllm","version":"0.27.1","parameters":{"server":{"dtype":"bfloat16","max_num_seqs":"87","max_model_len":"16384","kv_cache_dtype":"fp8","prefix_caching":false,"gpu_memory_utilization":"0.90","max_num_batched_tokens":"16384"},"dependencies":{"pytorch":"2.13.0+cu130","flashinfer":"0.6.16.post3","transformers":"5.15.1"},"load_balancer":{"type":"round_robin_streaming_http","retry_other_replica_on_connect_failure":true}},"parallelism":{"tensor":1,"pipeline":1,"replicas":4},"launch_command":"CUDA_VISIBLE_DEVICES=0 vllm serve nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 --revision e8f3c7c4de75ad84fe1bcef95d38eca76214480b --served-model-name nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 --host 127.0.0.1 --dtype bfloat16 --kv-cache-dtype fp8 --gpu-memory-utilization 0.90 --max-model-len 16384 --max-num-batched-tokens 16384 --max-num-seqs 87 --no-enable-prefix-caching --disable-uvicorn-access-log --mamba-backend flashinfer --mamba-cache-mode align --port 9001 --tensor-parallel-size 1\nCUDA_VISIBLE_DEVICES=1 vllm serve nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 --revision e8f3c7c4de75ad84fe1bcef95d38eca76214480b --served-model-name nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 --host 127.0.0.1 --dtype bfloat16 --kv-cache-dtype fp8 --gpu-memory-utilization 0.90 --max-model-len 16384 --max-num-batched-tokens 16384 --max-num-seqs 87 --no-enable-prefix-caching --disable-uvicorn-access-log --mamba-backend flashinfer --mamba-cache-mode align --port 9002 --tensor-parallel-size 1\nCUDA_VISIBLE_DEVICES=2 vllm serve nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 --revision e8f3c7c4de75ad84fe1bcef95d38eca76214480b --served-model-name nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 --host 127.0.0.1 --dtype bfloat16 --kv-cache-dtype fp8 --gpu-memory-utilization 0.90 --max-model-len 16384 --max-num-batched-tokens 16384 --max-num-seqs 87 --no-enable-prefix-caching --disable-uvicorn-access-log --mamba-backend flashinfer --mamba-cache-mode align --port 9003 --tensor-parallel-size 1\nCUDA_VISIBLE_DEVICES=3 vllm serve nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 --revision e8f3c7c4de75ad84fe1bcef95d38eca76214480b --served-model-name nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 --host 127.0.0.1 --dtype bfloat16 --kv-cache-dtype fp8 --gpu-memory-utilization 0.90 --max-model-len 16384 --max-num-batched-tokens 16384 --max-num-seqs 87 --no-enable-prefix-caching --disable-uvicorn-access-log --mamba-backend flashinfer --mamba-cache-mode align --port 9004 --tensor-parallel-size 1\nRUNPILE_BACKENDS=<replica-urls> RUNPILE_PROXY_PORT=9000 python lb_proxy.py","safe_environment":{"HF_HUB_OFFLINE":"1","CUDA_VISIBLE_DEVICES":"one distinct GPU index per replica","VLLM_MARLIN_USE_ATOMIC_ADD":"1"}},"workload":{"mode":"online","task":"text_generation","client":{"name":"vllm bench serve","backend":"vllm","version":"0.27.1"},"dataset":{"name":"vllm-random","range_ratio":0,"random_prefix_tokens":0},"sampling":{"mode":"greedy","temperature":0},"stopping":{"ignore_eos":true,"max_tokens":256},"streaming":true,"tokenizer":{"revision":"e8f3c7c4de75ad84fe1bcef95d38eca76214480b","identifier":"nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4"},"concurrency":128,"random_seed":20260823,"input_tokens":{"kind":"fixed","value":1024},"endpoint_type":"openai_completions","output_tokens":{"kind":"fixed","value":256},"request_count":512,"arrival_process":"closed_loop","prefix_behavior":{"shared_prefix_tokens":0,"server_prefix_caching":false},"warmup_request_count":16,"request_rate_per_second":null},"measurement":{"notes":"Aggregate-only output; --save-detailed was not enabled.","warmup":"16 explicit requests before the curve; the harness also issues one unmeasured endpoint test immediately before each timed point.","harness":"vllm bench serve","version":"0.27.1","repetitions":1,"timing_scope":"Client-observed, streaming online serving interval after the harness initial test request.","benchmark_command":"vllm bench serve --backend vllm --host localhost --port 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 random --random-input-len 1024 --random-output-len 256 --random-range-ratio 0 --num-prompts 512 --request-rate inf --max-concurrency 128 --ignore-eos --temperature 0 --seed 20260823 --percentile-metrics ttft,tpot,itl,e2el --metric-percentiles 50,90,95,99 --disable-tqdm --save-result --result-dir results --result-filename nemotron35-lightning-30b-a3b-nvfp4__r4__i1024__o256__c128.json","retry_failure_treatment":"No client retries; failed requests are counted as failures.","client_overhead_included":true},"results":{"latency_ms":{"end_to_end":{"p50":2966.4372509578243,"p90":5490.631008474156,"p95":6802.620931470301,"p99":8359.395038352814,"mean":4024.2256299343353,"stddev":1523.108843459091,"sample_count":512,"measurement_status":"measured"},"inter_token":{"p50":7.5165185844525695,"p90":8.11028191819787,"p95":8.316507586278021,"p99":43.17774103255943,"mean":8.429048794335127,"stddev":13.442187849839469,"sample_count":512,"measurement_status":"measured"},"time_to_first_token":{"p50":783.2536435453221,"p90":3304.021734627895,"p95":5166.965388064273,"p99":6109.741709588561,"mean":1874.8182944386826,"stddev":1521.5157122457958,"sample_count":512,"measurement_status":"measured"},"time_per_output_token":{"p50":8.44755327812962,"p90":9.359509325312342,"p95":9.734562315302444,"p99":10.100359786374895,"mean":8.429048374492757,"stddev":0.9582117921098696,"sample_count":512,"measurement_status":"measured"}},"throughput":{"requests_per_second":{"value":28.988052257595307,"sample_count":512,"measurement_status":"measured"},"total_tokens_per_second":{"value":37104.706889721994,"sample_count":512,"measurement_status":"measured"},"output_tokens_per_second":{"value":7420.9413779443985,"sample_count":512,"measurement_status":"measured"}},"failed_requests":0,"duration_seconds":17.662449185969308,"total_input_tokens":524288,"successful_requests":512,"total_output_tokens":131072},"artifacts":[{"kind":"vllm_bench_serve_json","sha256":"b7ed86ed158a97960bfd3aa3a9b905e0ac366e94ba70466c86e4a9b7f60681af","license":"CC-BY-4.0","filename":"nemotron35-lightning-30b-a3b-nvfp4__r4__i1024__o256__c128.json","media_type":"application/json","size_bytes":1723,"visibility":"private","client_artifact_id":"raw-vllm-serve","raw_requests_included":false}],"provenance":{"tool":"Runpile vllm importer completed from local evidence","source":"import","generated_at":"2026-08-23T23:38:45.731852+00:00","tool_version":"1.0"},"warnings":["vLLM reported uncalibrated FP8 attention q/prob scales of 1.0; serving performance is measured, but model quality was not evaluated.","vLLM auto-selected weight-only Marlin FP4 kernels and reported no native FP4 support for this ModelOpt W4A16 checkpoint on RTX 5090."]},"indexingState":"indexed","resultKey":"res_tk8NmIkGph31lunKSMsRVffB","eligibilityReasons":[],"artifactBacked":true,"downloads":[]}