This PR completes the removal of the deprecated vllm:time_per_output_token_seconds
metric that was deprecated in v0.11, hidden in v0.12, scheduled for removal in v0.13,
but delayed until v0.15.
Signed-off-by: carlory <baofa.fan@daocloud.io>
Co-authored-by: Claude Haiku 4.5 <noreply@anthropic.com>
In Prometheus Counters always expose their actual numeric value with a metric name that ends in _total. We should document the base name, as this what appears in the get_metrics() API.
Signed-off-by: CYJiang <86391540+googs1025@users.noreply.github.com>
Introduces three new Prometheus histograms for fine-grained observability of KV cache residency behavior:
vllm:kv_block_lifetime_seconds — total lifetime from allocation to free
vllm:kv_block_idle_before_evict_seconds — idle duration before eviction
vllm:kv_block_reuse_gap_seconds — time between consecutive reuses of the same block
These metrics help operators analyze KV cache efficiency, reuse patterns, and eviction timing beyond simple utilization rates.
Implementation uses monotonic timestamps for accuracy, 1% sampling for minimal overhead (~48 bytes/block), and is fully thread-safe with zero runtime cost when disabled.
Two new runtime flags are introduced:
--kv-cache-metrics – enable KV cache residency metrics
--kv-cache-metrics-sample – control sampling ratio (default: 0.01)
Signed-off-by: Shivam <shivamprasad91@gmail.com>
