Add argument for configuring Prometheus port

README.md

@@ -307,6 +307,12 @@ Options:
           [env: OTLP_SERVICE_NAME=]
           [default: text-embeddings-inference.server]
 
+      --prometheus-port <PORT>
+          The Prometheus metrics port to listen on
+
+          [env: PROMETHEUS_PORT=]
+          [default: 9000]
+
       --cors-allow-origin <CORS_ALLOW_ORIGIN>
           Unused for gRPC servers
 

router/src/lib.rs

@@ -63,6 +63,7 @@ pub async fn run(
     api_key: Option<String>,
     otlp_endpoint: Option<String>,
     otlp_service_name: String,
+    prometheus_port: u16,
     cors_allow_origin: Option<Vec<String>>,
 ) -> Result<()> {
     let model_id_path = Path::new(&model_id);
@@ -314,7 +315,7 @@ pub async fn run(
         }
     };
 
-    let prom_builder = prometheus::prometheus_builer(info.max_input_length)?;
+    let prom_builder = prometheus::prometheus_builer(addr, prometheus_port, info.max_input_length)?;
 
     #[cfg(all(feature = "grpc", feature = "http"))]
     compile_error!("Features `http` and `grpc` cannot be enabled at the same time.");

router/src/main.rs

@@ -164,6 +164,10 @@ struct Args {
     #[clap(default_value = "text-embeddings-inference.server", long, env)]
     otlp_service_name: String,
 
+    /// The Prometheus port to listen on.
+    #[clap(default_value = "9000", long, short, env)]
+    prometheus_port: u16,
+
     /// Unused for gRPC servers
     #[clap(long, env)]
     cors_allow_origin: Option<Vec<String>>,
@@ -227,6 +231,7 @@ async fn main() -> Result<()> {
         args.api_key,
         args.otlp_endpoint,
         args.otlp_service_name,
+        args.prometheus_port,
         args.cors_allow_origin,
     )
     .await?;

router/src/prometheus.rs

@@ -1,6 +1,15 @@
+use std::net::SocketAddr;
+
 use metrics_exporter_prometheus::{BuildError, Matcher, PrometheusBuilder};
 
-pub(crate) fn prometheus_builer(max_input_length: usize) -> Result<PrometheusBuilder, BuildError> {
+pub(crate) fn prometheus_builer(
+    addr: SocketAddr,
+    port: u16,
+    max_input_length: usize,
+) -> Result<PrometheusBuilder, BuildError> {
+    let mut addr = addr;
+    addr.set_port(port);
+
     // Duration buckets
     let duration_matcher = Matcher::Suffix(String::from("duration"));
     let n_duration_buckets = 35;
@@ -30,6 +39,7 @@ pub(crate) fn prometheus_builer(max_input_length: usize) -> Result<PrometheusBui
 
     // Prometheus handler
     PrometheusBuilder::new()
+        .with_http_listener(addr)
         .set_buckets_for_metric(duration_matcher, &duration_buckets)?
         .set_buckets_for_metric(input_length_matcher, &input_length_buckets)?
         .set_buckets_for_metric(batch_size_matcher, &batch_size_buckets)?