inference/research
verifiedhigh · 0.94updated 2026-10-04

vLLM KV block lifetime histogram

vllm · histogram

Fields

id
metric-exposures/w3m-vllm-vllm-kv-block-lifetime-seconds
Metric name
vllm:kv_block_lifetime_seconds
Metric type
histogram
Unit
seconds
Exposure
prometheus
Endpoint
:8000/metrics
Histogram buckets
0.0010.0020.0050.010.020.050.10.20.51251020306012030060012001800
Labels
model_nameengine
Enabled by default
no
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: kv_cache_residency. NOT ON BY DEFAULT: requires --kv-cache-metrics (ObservabilityConfig.kv_cache_metrics, default False), which also requires log stats. Sampled at --kv-cache-metrics-sample, default 0.01 (1% of blocks) - so a rate computed from these is a sampled estimate, not a census.

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: kv_cache_residency. NOT ON BY DEFAULT: requires --kv-cache-metrics (ObservabilityConfig.kv_cache_metrics, default False), which also requires log stats. Sampled at --kv-cache-metrics-sample, default 0.01 (1% of blocks) - so a rate computed from these is a sampled estimate, not a census.

Cross-references