[router] Expose worker startup secs & Return error instead of panic for router init (#3016)

sgl-router/py_src/sglang_router/launch_router.py

@@ -33,6 +33,7 @@ class RouterArgs:
 
     # Routing policy
     policy: str = "cache_aware"
+    worker_startup_timeout_secs: int = 300
     cache_threshold: float = 0.5
     balance_abs_threshold: int = 32
     balance_rel_threshold: float = 1.0001
@@ -87,6 +88,12 @@ class RouterArgs:
             choices=["random", "round_robin", "cache_aware"],
             help="Load balancing policy to use",
         )
+        parser.add_argument(
+            f"--{prefix}worker-startup-timeout-secs",
+            type=int,
+            default=RouterArgs.worker_startup_timeout_secs,
+            help="Timeout in seconds for worker startup",
+        )
         parser.add_argument(
             f"--{prefix}cache-threshold",
             type=float,
@@ -147,6 +154,9 @@ class RouterArgs:
             host=args.host,
             port=args.port,
             policy=getattr(args, f"{prefix}policy"),
+            worker_startup_timeout_secs=getattr(
+                args, f"{prefix}worker_startup_timeout_secs"
+            ),
             cache_threshold=getattr(args, f"{prefix}cache_threshold"),
             balance_abs_threshold=getattr(args, f"{prefix}balance_abs_threshold"),
             balance_rel_threshold=getattr(args, f"{prefix}balance_rel_threshold"),
@@ -188,9 +198,10 @@ def launch_router(args: argparse.Namespace) -> Optional[Router]:
 
         router = Router(
             worker_urls=router_args.worker_urls,
-            policy=policy_from_str(router_args.policy),
             host=router_args.host,
             port=router_args.port,
+            policy=policy_from_str(router_args.policy),
+            worker_startup_timeout_secs=router_args.worker_startup_timeout_secs,
             cache_threshold=router_args.cache_threshold,
             balance_abs_threshold=router_args.balance_abs_threshold,
             balance_rel_threshold=router_args.balance_rel_threshold,
@@ -205,7 +216,7 @@ def launch_router(args: argparse.Namespace) -> Optional[Router]:
 
     except Exception as e:
         logger.error(f"Error starting router: {e}")
-        return None
+        raise e
 
 
 class CustomHelpFormatter(
@@ -239,10 +250,7 @@ Examples:
 
 def main() -> None:
     router_args = parse_router_args(sys.argv[1:])
-    router = launch_router(router_args)
-
-    if router is None:
-        sys.exit(1)
+    launch_router(router_args)
 
 
 if __name__ == "__main__":

sgl-router/py_src/sglang_router/launch_server.py

@@ -68,7 +68,7 @@ def run_server(server_args, dp_rank):
     # create new process group
     os.setpgrp()
 
-    setproctitle(f"sglang::server")
+    setproctitle("sglang::server")
     # Set SGLANG_DP_RANK environment variable
     os.environ["SGLANG_DP_RANK"] = str(dp_rank)
 
@@ -120,9 +120,26 @@ def find_available_ports(base_port: int, count: int) -> List[int]:
 
 def cleanup_processes(processes: List[mp.Process]):
     for process in processes:
-        logger.info(f"Terminating process {process.pid}")
-        process.terminate()
-    logger.info("All processes terminated")
+        logger.info(f"Terminating process group {process.pid}")
+        try:
+            os.killpg(process.pid, signal.SIGTERM)
+        except ProcessLookupError:
+            # Process group may already be terminated
+            pass
+
+    # Wait for processes to terminate
+    for process in processes:
+        process.join(timeout=5)
+        if process.is_alive():
+            logger.warning(
+                f"Process {process.pid} did not terminate gracefully, forcing kill"
+            )
+            try:
+                os.killpg(process.pid, signal.SIGKILL)
+            except ProcessLookupError:
+                pass
+
+    logger.info("All process groups terminated")
 
 
 def main():
@@ -173,7 +190,12 @@ def main():
     ]
 
     # Start the router
-    router = launch_router(router_args)
+    try:
+        launch_router(router_args)
+    except Exception as e:
+        logger.error(f"Failed to start router: {e}")
+        cleanup_processes(server_processes)
+        sys.exit(1)
 
 
 if __name__ == "__main__":

sgl-router/py_src/sglang_router/router.py

@@ -17,6 +17,7 @@ class Router:
             - PolicyType.CacheAware: Distribute requests based on cache state and load balance
         host: Host address to bind the router server. Default: '127.0.0.1'
         port: Port number to bind the router server. Default: 3001
+        worker_startup_timeout_secs: Timeout in seconds for worker startup. Default: 300
         cache_threshold: Cache threshold (0.0-1.0) for cache-aware routing. Routes to cached worker
             if the match rate exceeds threshold, otherwise routes to the worker with the smallest
             tree. Default: 0.5
@@ -37,6 +38,7 @@ class Router:
         policy: PolicyType = PolicyType.RoundRobin,
         host: str = "127.0.0.1",
         port: int = 3001,
+        worker_startup_timeout_secs: int = 300,
         cache_threshold: float = 0.50,
         balance_abs_threshold: int = 32,
         balance_rel_threshold: float = 1.0001,
@@ -50,6 +52,7 @@ class Router:
             policy=policy,
             host=host,
             port=port,
+            worker_startup_timeout_secs=worker_startup_timeout_secs,
             cache_threshold=cache_threshold,
             balance_abs_threshold=balance_abs_threshold,
             balance_rel_threshold=balance_rel_threshold,