{"id":"run_ZijkPzwF5OpCpk96","owner":"anonymous-TWTejS","publishedAt":"2026-08-24T03:59:59.071Z","manifest":{"schema_version":"1.0","kind":"run","client_run_id":"vllm-d17d81946db9-1","name":"Qwen3.8-27B-NVFP4 | 4x RTX 5090 | 1024->256 | C16","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":"7d6f8d4d72f56b92b3cdbf22f156b90e1bab0108","identifier":"unsloth/Qwen3.8-27B-NVFP4","modalities":["text"],"architecture":"Qwen3_5ForConditionalGeneration","quantization":{"format":"NVFP4","recipe":"Unsloth Dynamic V3.0 preview","checkpoint":"compressed-tensors"},"modifications":["Multimodal inputs disabled at serving time."],"revision_kind":"commit","parameter_count_total":27000000000,"parameter_count_active":27000000000},"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":"26","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 unsloth/Qwen3.8-27B-NVFP4 --revision 7d6f8d4d72f56b92b3cdbf22f156b90e1bab0108 --served-model-name unsloth/Qwen3.8-27B-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 26 --no-enable-prefix-caching --disable-uvicorn-access-log --language-model-only --port 9001 --tensor-parallel-size 1\nCUDA_VISIBLE_DEVICES=1 vllm serve unsloth/Qwen3.8-27B-NVFP4 --revision 7d6f8d4d72f56b92b3cdbf22f156b90e1bab0108 --served-model-name unsloth/Qwen3.8-27B-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 26 --no-enable-prefix-caching --disable-uvicorn-access-log --language-model-only --port 9002 --tensor-parallel-size 1\nCUDA_VISIBLE_DEVICES=2 vllm serve unsloth/Qwen3.8-27B-NVFP4 --revision 7d6f8d4d72f56b92b3cdbf22f156b90e1bab0108 --served-model-name unsloth/Qwen3.8-27B-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 26 --no-enable-prefix-caching --disable-uvicorn-access-log --language-model-only --port 9003 --tensor-parallel-size 1\nCUDA_VISIBLE_DEVICES=3 vllm serve unsloth/Qwen3.8-27B-NVFP4 --revision 7d6f8d4d72f56b92b3cdbf22f156b90e1bab0108 --served-model-name unsloth/Qwen3.8-27B-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 26 --no-enable-prefix-caching --disable-uvicorn-access-log --language-model-only --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":"7d6f8d4d72f56b92b3cdbf22f156b90e1bab0108","identifier":"unsloth/Qwen3.8-27B-NVFP4"},"concurrency":16,"random_seed":20260823,"input_tokens":{"kind":"fixed","value":1024},"endpoint_type":"openai_completions","output_tokens":{"kind":"fixed","value":256},"request_count":200,"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 unsloth/Qwen3.8-27B-NVFP4 --tokenizer unsloth/Qwen3.8-27B-NVFP4 --dataset-name random --random-input-len 1024 --random-output-len 256 --random-range-ratio 0 --num-prompts 200 --request-rate inf --max-concurrency 16 --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 qwen38-27b-nvfp4__r4__i1024__o256__c16.json","retry_failure_treatment":"No client retries; failed requests are counted as failures.","client_overhead_included":true},"results":{"latency_ms":{"end_to_end":{"p50":4526.578180957586,"p90":4581.51210884098,"p95":4611.894588416908,"p99":4666.759836238343,"mean":4524.178917914396,"stddev":66.81774562197464,"sample_count":200,"measurement_status":"measured"},"inter_token":{"p50":16.3278270047158,"p90":16.58285497687757,"p95":16.668414568994194,"p99":17.104378209915012,"mean":16.78658074835845,"stddev":5.493058095071478,"sample_count":200,"measurement_status":"measured"},"time_to_first_token":{"p50":228.86457643471658,"p90":351.2868609279394,"p95":356.7832377157174,"p99":363.9145768759772,"mean":243.6009937012568,"stddev":70.58685348399806,"sample_count":200,"measurement_status":"measured"},"time_per_output_token":{"p50":16.82482012730165,"p90":17.162205153327506,"p95":17.186378660137017,"p99":17.32730751624331,"mean":16.78658009495349,"stddev":0.33925033213107914,"sample_count":200,"measurement_status":"measured"}},"throughput":{"requests_per_second":{"value":3.4126035248624818,"sample_count":200,"measurement_status":"measured"},"total_tokens_per_second":{"value":4368.1325118239765,"sample_count":200,"measurement_status":"measured"},"output_tokens_per_second":{"value":873.6265023647953,"sample_count":200,"measurement_status":"measured"}},"failed_requests":0,"duration_seconds":58.606280671898276,"total_input_tokens":204800,"successful_requests":200,"total_output_tokens":51200},"artifacts":[{"kind":"vllm_bench_serve_json","sha256":"d17d81946db939ccce8864e890efc32a2f2134bd0e228e3c4b15afe8bc72b7d9","license":"CC-BY-4.0","filename":"qwen38-27b-nvfp4__r4__i1024__o256__c16.json","media_type":"application/json","size_bytes":1647,"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":"2026-08-24T03:59:58.621058+00:00","tool_version":"1.0"}},"indexingState":"indexed","resultKey":"res_6BKLEV6XtRkGNoRK4zLfkLhK","eligibilityReasons":[],"artifactBacked":true,"downloads":[]}