{"id":"run_1HS2Qf9AConHL79C","owner":"anonymous-EVjB2S","publishedAt":"2026-08-30T21:09:21.713Z","manifest":{"schema_version":"1.0","kind":"run","client_run_id":"rps-a1419771af6daaea9e6c15ed6256bcd227689d71","name":"google/gemma-4-12B-it | baseline | 256->1024 | C8","notes":"Aggregate-only synthetic online serving measurement. It 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":"5926caa4ec0cac5cbfadaf4077420520de1d5205","identifier":"google/gemma-4-12B-it","modalities":["text"],"architecture":"Gemma4UnifiedForConditionalGeneration with Gemma4UnifiedAssistantForCausalLM","quantization":{"format":"BF16 target plus BF16 MTP assistant"},"modifications":["Multimodal inputs disabled at serving time.","Speculative mtp-assistant checkpoint: google/gemma-4-12B-it-assistant@46d4c6f13f0ac0ad827b915669b8df9b81c64c51."],"revision_kind":"commit","parameter_count_total":12000000000,"parameter_count_active":12000000000},"environment":{"os":"Linux-6.8.0-136-generic-x86_64-with-glibc2.39","cpu":{"model":"x86_64"},"kernel":"6.8.0-136-generic","topology":"One server engine on one GPU.","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"},"host_architecture":"x86_64","system_memory_bytes":269585743872},"runtime":{"engine":"vllm","version":"0.28.0","extensions":{"runpile.server_variant":"baseline","runpile.auxiliary_models":[{"role":"mtp-assistant","revision":"46d4c6f13f0ac0ad827b915669b8df9b81c64c51","repository":"google/gemma-4-12B-it-assistant","quantization":"BF16","weight_bytes":845719296}]},"parameters":{"server":{"dtype":"bfloat16","max_num_seqs":32,"enforce_eager":true,"max_model_len":9216,"kv_cache_dtype":"fp8","prefix_caching":false,"chunked_prefill":true,"expert_parallel":false,"async_scheduling":true,"gpu_memory_utilization":0.88,"max_num_batched_tokens":16384}},"parallelism":{"data":1,"expert":1,"tensor":1,"pipeline":1,"replicas":1},"launch_command":"CUDA_VISIBLE_DEVICES=1 vllm serve google/gemma-4-12B-it --revision 5926caa4ec0cac5cbfadaf4077420520de1d5205 --port 9001 --host 127.0.0.1 --dtype bfloat16 --gpu-memory-utilization 0.88 --max-model-len 9216 --max-num-batched-tokens 16384 --max-num-seqs 32 --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 --enforce-eager --served-model-name google/gemma-4-12B-it --disable-uvicorn-access-log --language-model-only --speculative-config '{\"method\":\"mtp\",\"model\":\"google/gemma-4-12B-it-assistant\",\"revision\":\"46d4c6f13f0ac0ad827b915669b8df9b81c64c51\",\"num_speculative_tokens\":1}'","build_identifier":"vllm-0.28.0-cu130-gguf-0.0.5","safe_environment":{"HF_HUB_OFFLINE":"1","CUDA_VISIBLE_DEVICES":"1","TOKENIZERS_PARALLELISM":"false"}},"workload":{"mode":"online","task":"text_generation","client":{"name":"vllm bench serve","backend":"vllm","version":"0.28.0","placement":"same-host"},"dataset":{"sha256":"43dfac229922b03b51294ab8ee6cb5744d397d2ce10e25a67a45582b31cf678c","generator":"runpile-sweeps-tokenizer-verified-jsonl","synthetic":true,"generator_revision":"cab078b273cceada260983886369dfc896718627d072e44f74a1cca08b67675c","shared_prefix_tokens":0},"sampling":{"mode":"greedy","temperature":0},"stopping":{"ignore_eos":true,"max_output_tokens":1024},"streaming":true,"tokenizer":{"revision":"5926caa4ec0cac5cbfadaf4077420520de1d5205","identifier":"google/gemma-4-12B-it"},"extensions":{"runpile.profile_id":"serve-fixed-256-1k-v1"},"concurrency":8,"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":"warm-host","cache_enabled":false,"shared_prefix_tokens":0},"warmup_request_count":8,"request_rate_per_second":null},"measurement":{"notes":"Aggregate-only output.","warmup":"8 unmeasured warmups per invocation after one readiness request.","harness":"vllm bench serve","version":"0.28.0","extensions":{"runpile.declared_warmup_count":8,"runpile.max_gpu_temperature_c":54,"runpile.preemption_log_mentions":0,"runpile.readiness_request_count":1,"runpile.controller_source_sha256":"cab078b273cceada260983886369dfc896718627d072e44f74a1cca08b67675c","runpile.speculative_accept_length":1.5011918501181065,"runpile.independent_repetition_index":3},"repetitions":1,"timing_scope":"client-observed streaming request latency","benchmark_command":"vllm bench serve --backend vllm --base-url http://127.0.0.1:9001 --endpoint /v1/completions --model google/gemma-4-12B-it --tokenizer google/gemma-4-12B-it --dataset-name custom --dataset-path <EXACT_SYNTHETIC_DATASET> --custom-output-len 1024 --skip-chat-template --disable-shuffle --num-prompts 128 --max-concurrency 8 --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 gemma4-12b-bf16-mtp1-tp1__baseline__serve-fixed-256-1k-v1__c8__r3__attempt1.json --num-warmups 8 --metadata campaign=runpile-rtx5090-nemotron-gemma4-specdecode-pilot-20260830-v2 config_id=gemma4-12b-bf16-mtp1-tp1 server_variant=baseline workload_key=serve-fixed-256-1k-v1 repetition=3 aggregate_only=true","retry_failure_treatment":"Failed requests counted; no client retries.","client_overhead_included":true},"results":{"latency_ms":{"end_to_end":{"p50":12401.367955375463,"p90":23439.399258466437,"p95":23515.037050982937,"p99":23624.339227820747,"mean":15238.79954086442,"stddev":4526.453674328642,"sample_count":128,"measurement_status":"measured"},"inter_token":{"p50":22.641795920208097,"p90":22.954963939264417,"p95":23.109056241810322,"p99":34.59547746926546,"mean":22.794226673850083,"stddev":2.058597611790882,"sample_count":128,"measurement_status":"measured"},"time_to_first_token":{"p50":86.56908199191093,"p90":93.38790988549584,"p95":206.10358391422676,"p99":248.57166721485555,"mean":96.31000324952765,"stddev":37.322149921328496,"sample_count":128,"measurement_status":"measured"},"time_per_output_token":{"p50":12.037968863200538,"p90":22.827687697532426,"p95":22.901544994028253,"p99":23.008996239473692,"mean":14.802042558763336,"stddev":4.428830011601178,"sample_count":128,"measurement_status":"measured"}},"throughput":{"requests_per_second":{"value":0.5101942752278915,"sample_count":128,"measurement_status":"measured"},"total_tokens_per_second":{"value":653.0486722917011,"sample_count":128,"measurement_status":"measured"},"output_tokens_per_second":{"value":522.4389378333609,"sample_count":128,"measurement_status":"measured"}},"failed_requests":0,"duration_seconds":250.8848221451044,"total_input_tokens":32768,"successful_requests":128,"total_output_tokens":131072},"artifacts":[{"kind":"vllm_bench_serve_json","sha256":"8086efdee42221c062b4d26a1fc4fd4859a73935fa6fdde0e9a25a57455423ae","license":"CC-BY-4.0","filename":"gemma4-12b-bf16-mtp1-tp1__baseline__serve-fixed-256-1k-v1__c8__r3.json","media_type":"application/json","size_bytes":2615,"visibility":"private","client_artifact_id":"raw-vllm-serve","raw_requests_included":false}],"provenance":{"tool":"runpile HTTP importer","source":"import","tool_version":"1.0"},"warnings":["Review imported facts against the local model, hardware, launch command, and benchmark command before publication."]},"indexingState":"indexed","resultKey":"res_BgM1GuRRqOJKMPNAwlx_J2kV","eligibilityReasons":[],"artifactBacked":true,"downloads":[]}