vllm:request_prefill_kv_computed_tokens

Version history

• In vLLM v0.23.0, vllm:request_prefill_kv_computed_tokens is known to be registered by the PrometheusStatLogger (V1 metrics) as a Prometheus histogram carrying the label names model_name and engine.
• In that same v0.23.0 scope, it records one histogram sample per finished request equal to prompt tokens minus non-negative cached tokens.
• Also in v0.23.0, the value is computed during finished-request logging after the request completes, from num_prompt_tokens and num_cached_tokens.
• Interpretively for v0.23.0, the metric reflects newly computed prefill KV work per completed request rather than total prompt length whenever caching contributes.
• Guidance recorded for v0.23.0 is to compare its distributions under matched workloads alongside cached-token counters and to avoid reading a single bucket count as token throughput.

Identity

Prometheus type

histogram

Labels

model_nameengine

Buckets

Not recorded — bucket boundaries are computed at runtime from max_model_len, so no fixed values are true

Unit

Not recorded — vLLM registration code has no structured unit field, and units are not inferred from name suffixes

Semantic coverage

Metric knowledge (6)

Related Use Cases (1)

• #40696: Full-block eligibility creates prefix-cache cliffs

Availability