vLLM cumulative generation-token counter (token throughput numerator)

Fields

id

metric-exposures/w3m-vllm-vllm-generation-tokens

Metric name

vllm:generation_tokens

Metric type

counter

Unit

tokens

Exposure

prometheus

Endpoint

:8000/metrics

Labels

Enabled by default

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: no bucket family. Token THROUGHPUT is not a metric here: vLLM exports cumulative token counters and you compute tok/s as rate(vllm:generation_tokens[$interval]). There is no ready-made throughput gauge, unlike SGLang's sglang:gen_throughput.

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: no bucket family. Token THROUGHPUT is not a metric here: vLLM exports cumulative token counters and you compute tok/s as rate(vllm:generation_tokens[$interval]). There is no ready-made throughput gauge, unlike SGLang's sglang:gen_throughput.

Cross-references

Engine (1)

Inference enginesvLLM