[model-gateway] support customizing Prometheus duration buckets (#14716)

sgl-model-gateway/bindings/python/sglang_router/router_args.py

@@ -54,6 +54,7 @@ class RouterArgs:
     # Prometheus configuration
     prometheus_port: Optional[int] = None
     prometheus_host: Optional[str] = None
+    prometheus_duration_buckets: Optional[List[float]] = None
     # Request ID headers configuration
     request_id_headers: Optional[List[str]] = None
     # Request timeout in seconds
@@ -342,6 +343,12 @@ class RouterArgs:
             default="0.0.0.0",
             help="Host address to bind the Prometheus metrics server. Supports IPv4, IPv6 (e.g., ::, ::1), or 0.0.0.0 for all interfaces",
         )
+        parser.add_argument(
+            f"--{prefix}prometheus-duration-buckets",
+            type=float,
+            nargs="+",
+            help="Buckets for Prometheus duration metrics",
+        )
         parser.add_argument(
             f"--{prefix}request-id-headers",
             type=str,

sgl-model-gateway/bindings/python/src/lib.rs

@@ -177,6 +177,7 @@ struct Router {
     bootstrap_port_annotation: String,
     prometheus_port: Option<u16>,
     prometheus_host: Option<String>,
+    prometheus_duration_buckets: Option<Vec<f64>>,
     request_timeout_secs: u64,
     request_id_headers: Option<Vec<String>>,
     pd_disaggregation: bool,
@@ -439,6 +440,7 @@ impl Router {
         bootstrap_port_annotation = String::from("sglang.ai/bootstrap-port"),
         prometheus_port = None,
         prometheus_host = None,
+        prometheus_duration_buckets = None,
         request_timeout_secs = 1800,
         request_id_headers = None,
         pd_disaggregation = false,
@@ -516,6 +518,7 @@ impl Router {
         bootstrap_port_annotation: String,
         prometheus_port: Option<u16>,
         prometheus_host: Option<String>,
+        prometheus_duration_buckets: Option<Vec<f64>>,
         request_timeout_secs: u64,
         request_id_headers: Option<Vec<String>>,
         pd_disaggregation: bool,
@@ -606,6 +609,7 @@ impl Router {
             bootstrap_port_annotation,
             prometheus_port,
             prometheus_host,
+            prometheus_duration_buckets,
             request_timeout_secs,
             request_id_headers,
             pd_disaggregation,
@@ -695,6 +699,7 @@ impl Router {
                 .prometheus_host
                 .clone()
                 .unwrap_or_else(|| "127.0.0.1".to_string()),
+            duration_buckets: self.prometheus_duration_buckets.clone(),
         });
 
         let runtime = tokio::runtime::Runtime::new()