{"id":"run_hWNIAajUCzgQN7C3","owner":"anonymous-52J7ji","publishedAt":"2026-08-26T12:19:17.794Z","manifest":{"schema_version":"1.0","kind":"run","client_run_id":"vllm-4793d55a926e-1","name":"nvidia/Qwen3.6-35B-A3B-NVFP4 | tp1 | baseline | 4096->1 | 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":1},"streaming":true,"tokenizer":{"revision":"491c2f1ea524c639598bf8fa787a93fed5a6fbce","identifier":"nvidia/Qwen3.6-35B-A3B-NVFP4"},"concurrency":2,"random_seed":0,"input_tokens":{"kind":"fixed","value":4096},"endpoint_type":"openai-compatible-completions","output_tokens":{"kind":"fixed","value":1},"request_count":256,"arrival_process":"closed_loop","prefix_behavior":{"cache_state":"cold","cache_enabled":false,"shared_prefix_tokens":0},"warmup_request_count":16,"request_rate_per_second":null},"measurement":{"notes":"Aggregate-only output; readiness request is separate from workload warmups; TPOT and ITL are not applicable for the one-output-token categorization workload.","warmup":"16 declared unmeasured warmup requests before 256 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":16,"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 4096 --random-output-len 1 --random-range-ratio 0 --random-prefix-len 0 --num-prompts 256 --num-warmups 16 --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-4k-1-categorization-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-4k-1-categorization-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":252.326785819605,"p90":252.9980931431055,"p95":253.17781674675643,"p99":253.84864907246086,"mean":251.84831012848008,"stddev":7.185604035186342,"sample_count":256,"measurement_status":"measured"},"inter_token":{"p50":0,"p90":0,"p95":0,"p99":0,"mean":0,"stddev":0,"sample_count":256,"measurement_status":"measured"},"time_to_first_token":{"p50":252.32696579769254,"p90":252.9982936102897,"p95":253.1781451543793,"p99":253.84876721072942,"mean":251.84852106758626,"stddev":7.185601395960459,"sample_count":256,"measurement_status":"measured"},"time_per_output_token":{"p50":0,"p90":0,"p95":0,"p99":0,"mean":0,"stddev":0,"sample_count":256,"measurement_status":"measured"}},"throughput":{"requests_per_second":{"value":7.924075252419903,"sample_count":256,"measurement_status":"measured"},"total_tokens_per_second":{"value":32464.936309164343,"sample_count":256,"measurement_status":"measured"},"output_tokens_per_second":{"value":7.924075252419903,"sample_count":256,"measurement_status":"measured"}},"failed_requests":0,"duration_seconds":32.30660889064893,"total_input_tokens":1048576,"successful_requests":256,"total_output_tokens":256},"artifacts":[{"kind":"benchmark_result","sha256":"4793d55a926e7733c7f3af8fc289f34411a0e32fb05abdb66181ea6ac407e8c6","license":"CC-BY-4.0","filename":"qwen36-35b-a3b-nvfp4-tp1__tp1__baseline__serve-fixed-4k-1-categorization-v1__c2__rep1.json","media_type":"application/json","size_bytes":1494,"visibility":"private","client_artifact_id":"aggregate-result-4793d55a926e7733","raw_requests_included":false}],"provenance":{"tool":"Runpile vllm importer completed from local evidence","source":"import","generated_at":"2026-08-26T12:19:17.447417+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_jkp8W6l8RjOQzDbHQ3uKkf8g","eligibilityReasons":[],"artifactBacked":true,"downloads":[]}