TGI current batch max-tokens gauge
text-generation-inference · gauge
Fields
id
metric-exposures/w3m-text-generation-inference-tgi-batch-current-max-tokensMetric name
tgi_batch_current_max_tokens
Metric type
gauge
Unit
tokens
Exposure
prometheus
Endpoint
:9000/metrics
Enabled by default
yes
notes
TGI main, router/src/server.rs plus backends/v3/src/{backend,queue,lib}.rs, read 2026-10-04. On by default: the Prometheus recorder is built unconditionally with PrometheusBuilder and an HTTP listener on prometheus_port, default 9000, so nothing has to be opted into. Label keys are method (prefill|decode) on the batch families, err on tgi_request_failure, and reason on tgi_batch_concat - there are no model or engine labels, so every series is fleet-wide. Not in the published docs table - it appears in the source only, which is one of six metrics where the doc table lags the code.
Notes
TGI main, router/src/server.rs plus backends/v3/src/{backend,queue,lib}.rs, read 2026-10-04. On by default: the Prometheus recorder is built unconditionally with PrometheusBuilder and an HTTP listener on prometheus_port, default 9000, so nothing has to be opted into. Label keys are method (prefill|decode) on the batch families, err on tgi_request_failure, and reason on tgi_batch_concat - there are no model or engine labels, so every series is fleet-wide. Not in the published docs table - it appears in the source only, which is one of six metrics where the doc table lags the code.