Add sglang:get_loads_duration_seconds metric (#25163)

Co-authored-by: cctry <cctry@meta.com>

python/sglang/srt/entrypoints/v1_loads.py

@@ -19,6 +19,7 @@ metrics for load balancing, monitoring, and capacity planning.
 """
 
 import dataclasses
+import time
 from datetime import datetime, timezone
 from typing import Optional
 
@@ -153,6 +154,7 @@ async def get_loads(
     """
     include_list = [s.strip() for s in include.split(",")] if include else None
 
+    start = time.perf_counter()
     try:
         load_results = await tokenizer_manager.get_loads(
             include=include_list,
@@ -160,6 +162,12 @@ async def get_loads(
         )
     except ValueError as e:
         raise HTTPException(status_code=400, detail=str(e))
+    finally:
+        mc = getattr(tokenizer_manager, "metrics_collector", None)
+        if mc is not None:
+            mc.get_loads_duration_seconds.labels(**mc.labels).observe(
+                time.perf_counter() - start
+            )
 
     if format == "prometheus":
         return _format_loads_prometheus(load_results)

python/sglang/srt/observability/metrics_collector.py

@@ -1352,6 +1352,13 @@ class TokenizerMetricsCollector:
             labelnames=labels.keys(),
         )
 
+        self.get_loads_duration_seconds = Histogram(
+            name="sglang:get_loads_duration_seconds",
+            documentation="Time spent serving /v1/loads requests (seconds).",
+            labelnames=labels.keys(),
+            buckets=(0.0001, 0.0005, 0.001, 0.005, 0.01, 0.05, 0.1, 0.5, 1.0),
+        )
+
         self.num_so_requests_total = Counter(
             name="sglang:num_so_requests_total",
             documentation="Number of structured output requests processed.",

test/registered/language/test_srt_backend.py

@@ -35,6 +35,7 @@ class TestSRTBackend(CustomTestCase):
             mem_fraction_static=0.7,
             incremental_streaming_output=True,
             log_level="info",
+            enable_metrics=True,
         )
         sgl.set_default_backend(cls.backend)