Optimize uvicorn startup command (#25041)

python/sglang/srt/entrypoints/http_server.py

@@ -146,7 +146,6 @@ from sglang.srt.managers.multi_tokenizer_mixin import (
     MultiTokenizerRouter,
     TokenizerWorker,
     get_main_process_id,
-    monkey_patch_uvicorn_multiprocessing,
     read_from_shared_memory,
     write_data_for_multi_tokenizer,
 )
@@ -2282,7 +2281,6 @@ def _setup_and_run_http_server(
                 "level": "INFO",
                 "propagate": False,
             }
-            monkey_patch_uvicorn_multiprocessing()
 
             if server_args.enable_ssl_refresh:
                 logger.warning(
@@ -2298,6 +2296,7 @@ def _setup_and_run_http_server(
                 root_path=server_args.fastapi_root_path,
                 log_level=server_args.log_level_http or server_args.log_level,
                 timeout_keep_alive=envs.SGLANG_TIMEOUT_KEEP_ALIVE.get(),
+                timeout_worker_healthcheck=envs.SGLANG_UVICORN_WORKER_HEALTHCHECK_TIMEOUT.get(),
                 loop="uvloop",
                 workers=server_args.tokenizer_worker_num,
                 ssl_keyfile=server_args.ssl_keyfile,

python/sglang/srt/environ.py

@@ -519,6 +519,9 @@ class Envs:
 
     # HTTP Server
     SGLANG_TIMEOUT_KEEP_ALIVE = EnvInt(5)
+    # Uvicorn multiprocess supervisor pings each worker on this interval; default 5s is
+    # too short when many workers cold-start and load tokenizers in parallel.
+    SGLANG_UVICORN_WORKER_HEALTHCHECK_TIMEOUT = EnvInt(10)
 
     # HTTP/2 Server
     SGLANG_GRANIAN_PARENT_PID = EnvInt(None)

python/sglang/srt/managers/multi_tokenizer_mixin.py

@@ -26,7 +26,6 @@ import os
 import pickle
 import sys
 import threading
-from functools import partialmethod
 from multiprocessing import shared_memory
 from typing import TYPE_CHECKING, Any, Dict, Optional, Union
 
@@ -591,20 +590,6 @@ def write_data_for_multi_tokenizer(
     return args_shm
 
 
-def monkey_patch_uvicorn_multiprocessing(timeout: float = 10):
-    """Monkey patch uvicorn multiprocessing is_alive timeout"""
-    # from default 5s -> 10s
-    try:
-        from uvicorn.supervisors.multiprocess import Process
-
-        Process.is_alive = partialmethod(Process.is_alive, timeout=timeout)
-
-    except ImportError:
-        logger.warning(
-            "uvicorn.supervisors.multiprocess not found, skipping monkey patch"
-        )
-
-
 class SenderWrapper:
     def __init__(self, port_args: PortArgs, send_to_scheduler: zmq.Socket):
         self.port_args = port_args