From 263eab9f5d6659633dfede1634be542fa2e09004 Mon Sep 17 00:00:00 2001
From: ishandhanani <82981111+ishandhanani@users.noreply.github.com>
Date: Fri, 31 Oct 2025 11:34:57 -0700
Subject: [PATCH] fix: dummy health check server not accessible on non-zero
 rank nodes (#12297)

---
 python/sglang/compile_deep_gemm.py | 20 +++++++++++++-------
 python/sglang/srt/utils/common.py  | 26 +++++++++++++++++---------
 2 files changed, 30 insertions(+), 16 deletions(-)

diff --git a/python/sglang/compile_deep_gemm.py b/python/sglang/compile_deep_gemm.py
index 739b907bf..7e1e68301 100644
--- a/python/sglang/compile_deep_gemm.py
+++ b/python/sglang/compile_deep_gemm.py
@@ -104,15 +104,21 @@ def launch_server_process_and_send_one_request(
             if response.status_code == 200:
                 # Rank-0 node send a request to sync with other node and then return.
                 if server_args.node_rank == 0:
+                    payload = {
+                        "input_ids": [0, 1, 2, 3],
+                        "sampling_params": {
+                            "max_new_tokens": 8,
+                            "temperature": 0,
+                        },
+                    }
+                    # In PD mode, include fake bootstrap fields so workers don't assert
+                    if server_args.disaggregation_mode != "null":
+                        payload["bootstrap_host"] = FAKE_BOOTSTRAP_HOST
+                        payload["bootstrap_room"] = 0
+
                     response = requests.post(
                         f"{base_url}/generate",
-                        json={
-                            "input_ids": [0, 1, 2, 3],
-                            "sampling_params": {
-                                "max_new_tokens": 8,
-                                "temperature": 0,
-                            },
-                        },
+                        json=payload,
                         timeout=600,
                     )
                     if response.status_code != 200:
diff --git a/python/sglang/srt/utils/common.py b/python/sglang/srt/utils/common.py
index 0c1a13b5f..244b8fc35 100644
--- a/python/sglang/srt/utils/common.py
+++ b/python/sglang/srt/utils/common.py
@@ -2351,16 +2351,24 @@ def launch_dummy_health_check_server(host, port, enable_metrics):
     )
     server = uvicorn.Server(config=config)
 
-    try:
-        loop = asyncio.get_running_loop()
-        logger.info(
-            f"Dummy health check server scheduled on existing loop at {host}:{port}"
-        )
-        loop.create_task(server.serve())
+    # Run server in a background daemon thread with its own event loop
+    # This prevents blocking the main thread while still serving health checks
+    def run_server():
+        try:
+            asyncio.run(server.serve())
+        except Exception as e:
+            logger.error(f"Dummy health check server failed to start: {e}")
+            raise
+        finally:
+            logger.info(f"Dummy health check server stopped at {host}:{port}")
 
-    except RuntimeError:
-        logger.info(f"Starting dummy health check server at {host}:{port}")
-        server.run()
+    thread = threading.Thread(
+        target=run_server, daemon=True, name="health-check-server"
+    )
+    thread.start()
+    logger.info(
+        f"Dummy health check server started in background thread at {host}:{port}"
+    )
 
 
 def create_checksum(directory: str):