{"id":"run_mka-eHLzhgXMSPWe","owner":"anonymous-52J7ji","publishedAt":"2026-08-26T12:19:13.737Z","manifest":{"schema_version":"1.0","kind":"run","client_run_id":"vllm-baa386f75bd1-1","name":"nvidia/Qwen3.6-35B-A3B-NVFP4 | tp1 | baseline | 256->1024 | C2","notes":"Text-only synthetic online serving measurement. Exact fixed token lengths, greedy decoding, streaming, EOS ignored, zero shared prefix, and aggregate-only evidence; this reports serving performance and does not evaluate model quality.","visibility":"public","license":"CC-BY-4.0","outcome":"completed","subject":{"type":"model","source":"huggingface","adapters":[],"revision":"491c2f1ea524c639598bf8fa787a93fed5a6fbce","identifier":"nvidia/Qwen3.6-35B-A3B-NVFP4","modalities":["text"],"architecture":"Qwen3_5MoeForConditionalGeneration","quantization":{"format":"W4A16 NVFP4 mixed precision","checkpoint":"modelopt"},"modifications":["Multimodal inputs disabled at serving time."],"revision_kind":"commit","parameter_count_total":35000000000,"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":"One server process on one RTX 5090.","accelerators":[{"count":1,"model":"NVIDIA GeForce RTX 5090","driver":"580.173.02","vendor":"NVIDIA","architecture":"Blackwell SM120","interconnect":"single device","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","extensions":{"runpile.topology":"tp1"},"parameters":{"server":{"dtype":"bfloat16","max_num_seqs":64,"enforce_eager":false,"max_model_len":9216,"kv_cache_dtype":"fp8","server_variant":"baseline","enable_prefix_caching":false,"enable_chunked_prefill":true,"gpu_memory_utilization":0.92,"max_num_batched_tokens":16384},"dependencies":{"pytorch":"2.13.0+cu130","flashinfer":"0.6.16.post3","transformers":"5.15.1","compressed_tensors":"0.17.0"}},"parallelism":{"data":1,"expert":1,"tensor":1,"pipeline":1,"replicas":1},"launch_command":"CUDA_VISIBLE_DEVICES=0 vllm serve nvidia/Qwen3.6-35B-A3B-NVFP4 --revision 491c2f1ea524c639598bf8fa787a93fed5a6fbce --served-model-name nvidia/Qwen3.6-35B-A3B-NVFP4 --host 127.0.0.1 --port 9000 --dtype bfloat16 --gpu-memory-utilization 0.92 --max-model-len 9216 --max-num-batched-tokens 16384 --max-num-seqs 64 --kv-cache-dtype fp8 --no-enable-prefix-caching --enable-chunked-prefill --no-enforce-eager --async-scheduling --distributed-executor-backend mp --tensor-parallel-size 1 --pipeline-parallel-size 1 --disable-uvicorn-access-log --language-model-only --no-enable-flashinfer-autotune","build_identifier":"vllm-0.27.1","safe_environment":{"HF_HUB_OFFLINE":"1","CUDA_VISIBLE_DEVICES":"0","TOKENIZERS_PARALLELISM":"false","VLLM_MARLIN_USE_ATOMIC_ADD":"1"}},"workload":{"mode":"online","task":"text_generation","client":{"name":"vllm bench serve","backend":"vllm","version":"0.27.1","placement":"same-host"},"dataset":{"generator":"harness-random-token-ids","synthetic":true,"shared_prefix_tokens":0},"sampling":{"mode":"greedy","temperature":0},"stopping":{"ignore_eos":true,"max_output_tokens":1024},"streaming":true,"tokenizer":{"revision":"491c2f1ea524c639598bf8fa787a93fed5a6fbce","identifier":"nvidia/Qwen3.6-35B-A3B-NVFP4"},"extensions":{"runpile.profile_id":"serve-fixed-256-1k-v1"},"concurrency":2,"random_seed":0,"input_tokens":{"kind":"fixed","value":256},"endpoint_type":"openai-compatible-completions","output_tokens":{"kind":"fixed","value":1024},"request_count":128,"arrival_process":"closed_loop","prefix_behavior":{"cache_state":"cold","cache_enabled":false,"shared_prefix_tokens":0},"warmup_request_count":8,"request_rate_per_second":null},"measurement":{"notes":"Aggregate-only output; readiness request is separate from workload warmups.","warmup":"8 declared unmeasured warmup requests before 128 measured requests; one separate endpoint-readiness request precedes the declared warmups and is not counted.","harness":"vllm bench serve","version":"0.27.1","extensions":{"runpile.declared_warmup_count":8,"runpile.readiness_request_count":1},"repetitions":1,"timing_scope":"client-observed streaming request latency","benchmark_command":"vllm bench serve --backend vllm --base-url http://127.0.0.1:9000 --endpoint /v1/completions --model nvidia/Qwen3.6-35B-A3B-NVFP4 --tokenizer nvidia/Qwen3.6-35B-A3B-NVFP4 --dataset-name random --random-input-len 256 --random-output-len 1024 --random-range-ratio 0 --random-prefix-len 0 --num-prompts 128 --num-warmups 8 --request-rate inf --max-concurrency 2 --ignore-eos --temperature 0 --seed 0 --percentile-metrics ttft,tpot,itl,e2el --metric-percentiles 50,90,95,99 --disable-tqdm --save-result --result-dir results --result-filename qwen36-35b-a3b-nvfp4-tp1__tp1__baseline__serve-fixed-256-1k-v1__c2__rep1.json --metadata campaign=runpile-rtx5090-param-sweeps-20260826 config_id=qwen36-35b-a3b-nvfp4-tp1 server_variant=baseline workload_key=serve-fixed-256-1k-v1 repetition=1 aggregate_only=true","retry_failure_treatment":"Failed requests counted; no client retries.","client_overhead_included":true},"results":{"latency_ms":{"end_to_end":{"p50":7009.8535707220435,"p90":7020.1829333323985,"p95":7022.899534506723,"p99":7025.443941601552,"mean":7002.2283240250545,"stddev":23.67858397226059,"sample_count":128,"measurement_status":"measured"},"inter_token":{"p50":6.764063378795981,"p90":6.901628011837602,"p95":6.943402672186494,"p99":7.019506623037159,"mean":6.774237687422371,"stddev":0.7725529799148304,"sample_count":128,"measurement_status":"measured"},"time_to_first_token":{"p50":71.35457592085004,"p90":91.05309681035578,"p95":91.58714818768203,"p99":92.02624147757888,"mean":72.18338492020848,"stddev":18.10589021486788,"sample_count":128,"measurement_status":"measured"},"time_per_output_token":{"p50":6.773009014116529,"p90":6.808103419219454,"p95":6.8095252359830365,"p99":6.813548891764145,"mean":6.774237477130837,"stddev":0.029088007428798274,"sample_count":128,"measurement_status":"measured"}},"throughput":{"requests_per_second":{"value":0.2856167501838331,"sample_count":128,"measurement_status":"measured"},"total_tokens_per_second":{"value":365.5894402353063,"sample_count":128,"measurement_status":"measured"},"output_tokens_per_second":{"value":292.4715521882451,"sample_count":128,"measurement_status":"measured"}},"failed_requests":0,"duration_seconds":448.1529879379086,"total_input_tokens":32768,"successful_requests":128,"total_output_tokens":131072},"artifacts":[{"kind":"benchmark_result","sha256":"baa386f75bd1c541c9ce71629642b1d349108f9e673e51746e5563566d878267","license":"CC-BY-4.0","filename":"qwen36-35b-a3b-nvfp4-tp1__tp1__baseline__serve-fixed-256-1k-v1__c2__rep1.json","media_type":"application/json","size_bytes":1683,"visibility":"private","client_artifact_id":"aggregate-result-baa386f75bd1c541","raw_requests_included":false}],"provenance":{"tool":"Runpile vllm importer completed from local evidence","source":"import","generated_at":"2026-08-26T12:19:13.356660+00:00","tool_version":"1.0"},"warnings":["Review imported facts against the local model, hardware, launch command, and benchmark command before publication.","FP8 KV-cache serving performance measured; no quality evaluation was performed."]},"indexingState":"indexed","resultKey":"res_6pvisK8CUGtAZDvvkSwTeAVB","eligibilityReasons":[],"artifactBacked":true,"downloads":[]}