TGI router/src/server.rs - PrometheusBuilder bucket config and prometheus_port

router/src/server.rs on main, read 2026-10-04. The Prometheus recorder is built at line 1974 with PrometheusBuilder::new().with_http_listener(p_addr) where p_addr is the server address with the port replaced by prometheus_port, then per-metric bucket overrides: duration_matcher = Matcher::Suffix("duration") with 35 buckets formed geometrically from 0.0001 x 1.5^n (so EVERY metric whose name ends in 'duration' shares one bucket ladder, in seconds, starting at 0.00015); input_length_matcher = Matcher::Full("tgi_request_input_length") with 100 linear buckets to max_input_tokens; generated_tokens_matcher and max_new_tokens_matcher are Matcher::Full on tgi_request_generated_tokens / tgi_request_max_new_tokens with 100 linear buckets to max_total_tokens; batch_size_matcher = Matcher::Full("tgi_batch_next_size") with buckets 1..1024. The skipped-tokens override is commented out, citing metrics-rs issue 467. The CLI default is prometheus_port 9000 (backends/v3/src/main.rs: #[clap(default_value = "9000", long, short, env)]). Note the practical consequence: a bucket ladder whose first bound is 0.00015 s cannot be quoted as a fixed list without naming max_input_tokens, and tgi_batch_next_size is a Histogram despite reading like a gauge.