Tiny add CPU resource monitoring for overload diagnosis (#16852)

python/sglang/srt/metrics/cpu_monitor.py

@@ -0,0 +1,31 @@
+import threading
+import time
+
+import psutil
+
+
+def start_cpu_monitor_thread(component: str, interval: float = 5.0) -> threading.Thread:
+    from prometheus_client import Counter
+
+    cpu_seconds_total = Counter(
+        name="sglang:process_cpu_seconds_total",
+        documentation="Total CPU time consumed by this process (user + system)",
+        labelnames=["component"],
+    )
+
+    def monitor():
+        process = psutil.Process()
+        last_times = process.cpu_times()
+
+        while True:
+            time.sleep(interval)
+            curr_times = process.cpu_times()
+            delta = (curr_times.user - last_times.user) + (
+                curr_times.system - last_times.system
+            )
+            cpu_seconds_total.labels(component=component).inc(delta)
+            last_times = curr_times
+
+    t = threading.Thread(target=monitor, daemon=True)
+    t.start()
+    return t