fix: fix the missing metrics on non-rank0 nodes (#7720)

Author: acelyc111

python/sglang/srt/entrypoints/engine.py

@@ -765,7 +765,9 @@ def _launch_subprocesses(
             # When using `Engine` as a Python API, we don't want to block here.
             return None, None, None
 
-        launch_dummy_health_check_server(server_args.host, server_args.port)
+        launch_dummy_health_check_server(
+            server_args.host, server_args.port, server_args.enable_metrics
+        )
 
         for proc in scheduler_procs:
             proc.join()

python/sglang/srt/utils.py

@@ -85,6 +85,8 @@
 from torch.utils._contextlib import _DecoratorContextManager
 from triton.runtime.cache import FileCacheManager
 
+from sglang.srt.metrics.func_timer import enable_func_timer
+
 logger = logging.getLogger(__name__)
 
 show_time_cost = False
@@ -2049,7 +2051,7 @@ def rank0_log(msg: str):
         logger.info(msg)
 
 
-def launch_dummy_health_check_server(host, port):
+def launch_dummy_health_check_server(host, port, enable_metrics):
     import asyncio
 
     import uvicorn
@@ -2067,6 +2069,11 @@ async def health_generate():
         """Check the health of the http server."""
         return Response(status_code=200)
 
+    # Add prometheus middleware
+    if enable_metrics:
+        add_prometheus_middleware(app)
+        enable_func_timer()
+
     config = uvicorn.Config(
         app,
         host=host,