google/gemma-4-12B-it

completed
Request throughput
4.02 req/s
Output throughput
4 tok/s
p95 TTFT
498.7 ms
p95 TPOT
0.00 ms

Manifest warnings

Review these published caveats before interpreting or comparing the metrics.

  • Review imported facts against the local model, hardware, launch command, and benchmark command before publication.

Server launch command

CUDA_VISIBLE_DEVICES=0 vllm serve google/gemma-4-12B-it \
  --revision 5926caa4ec0cac5cbfadaf4077420520de1d5205 \
  --port 9000 \
  --host 127.0.0.1 \
  --dtype bfloat16 \
  --gpu-memory-utilization 0.92 \
  --max-model-len 9216 \
  --max-num-batched-tokens 16384 \
  --max-num-seqs 64 \
  --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 \
  --no-enforce-eager \
  --served-model-name google/gemma-4-12B-it \
  --quantization fp8_per_tensor \
  --disable-uvicorn-access-log \
  --language-model-only

Run details

Workload
4096 → 1
Methodology
closed loop · streaming
Test load
2 concurrent requests
Endpoint
openai-compatible-completions
Requests
16 unmeasured warmups per invocation after one readiness request.
Dataset
generator=runpile-sweeps-tokenizer-verified-jsonl synthetic=true shared_prefix_tokens=0
Sampling
mode=greedy temperature=0
Stopping
ignore_eos=true max_output_tokens=1
Prefix and cache
cache_state=warm-host cache_enabled=false shared_prefix_tokens=0
Client
name=vllm bench serve backend=vllm version=0.23.0+cu129 placement=same-host
Hardware
1× NVIDIA GeForce RTX 5090
Runtime
vllm 0.23.0+cu129
Runtime configuration
dtype=bfloat16 max_num_seqs=64 enforce_eager=false max_model_len=9216 kv_cache_dtype=fp8 prefix_caching=false chunked_prefill=true expert_parallel=false async_scheduling=true gpu_memory_utilization=0.92 max_num_batched_tokens=16384 runtime_weight_quantization=fp8_per_tensor
Measurement
vllm bench serve 0.23.0+cu129
Timing
client-observed streaming request latency · client overhead included
Repetitions
1
Failure handling
Failed requests counted; no client retries.
Quantization
format=FP8 per tensor application=runtime weight quantization
Model revision
5926caa4ec0cac5cbfadaf4077420520de1d5205
Evidence
Raw benchmark evidence attached · eligible and ranked
Published
by anonymous-EVjB2S

Benchmark command

vllm bench serve \
  --backend vllm \
  --base-url http://127.0.0.1:9000 \
  --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 1 \
  --skip-chat-template \
  --disable-shuffle \
  --num-prompts 256 \
  --max-concurrency 2 \
  --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-tp1__runtime-weight-fp8-per-tensor__serve-fixed-4k-1-categorization-v1__c2__r1__attempt1.json \
  --num-warmups 16 \
  --metadata campaign=runpile-rtx5090-param-sweeps-20260826-retry2 config_id=gemma4-12b-bf16-tp1 server_variant=runtime-weight-fp8-per-tensor workload_key=serve-fixed-4k-1-categorization-v1 repetition=1 aggregate_only=true