vLLM per-step inter-token latency histogram
vllm · histogram
Fields
id
metric-exposures/w3m-vllm-vllm-inter-token-latency-secondsMetric name
vllm:inter_token_latency_seconds
Metric type
histogram
Unit
seconds
Exposure
prometheus
Endpoint
:8000/metrics
Histogram buckets
0.010.0250.050.0750.10.150.20.30.40.50.7512.557.510204080
Labels
model_nameengine
Enabled by default
yes
Phase scoped
yes
notes
vLLM main, PrometheusStatLogger in vllm/v1/metrics/loggers.py, read 2026-10-04. On by default (EngineArgs.disable_log_stats defaults False); served at :8000/metrics. Bucket family name in --custom-histogram-buckets: inter_token_latency. One sample per streamed output event: the wall-clock gap between successive outputs, NOT divided by token count. Distinct from request_time_per_output_token_seconds and not interchangeable with it.
Notes
vLLM main, PrometheusStatLogger in vllm/v1/metrics/loggers.py, read 2026-10-04. On by default (EngineArgs.disable_log_stats defaults False); served at :8000/metrics. Bucket family name in --custom-histogram-buckets: inter_token_latency. One sample per streamed output event: the wall-clock gap between successive outputs, NOT divided by token count. Distinct from request_time_per_output_token_seconds and not interchangeable with it.