[Fix] Fix uvloop get_event_loop() is not suitable for 0.22.x (#13612)

Signed-off-by: lzy <tomlzy213@gmail.com>
Co-authored-by: lzy <tomlzy213@gmail.com>

benchmark/prefill_only/util.py

@@ -240,7 +240,7 @@ async def make_http_call(
         api_name: Name of the API for error messages
     """
     try:
-        start_time = asyncio.get_event_loop().time()
+        start_time = asyncio.get_running_loop().time()
 
         request_json = build_http_request_json(request_data)
         headers = {"Content-Type": "application/json"}
@@ -253,7 +253,7 @@ async def make_http_call(
                     f"[HTTP] {api_name} Request {request_id} failed with status "
                     f"{resp.status}: {resp_text}"
                 )
-                completion_time = asyncio.get_event_loop().time()
+                completion_time = asyncio.get_running_loop().time()
                 await results_queue.put((request_id, 0, False, completion_time))
                 return
 
@@ -271,13 +271,13 @@ async def make_http_call(
                 )
                 success = False
 
-        completion_time = asyncio.get_event_loop().time()
+        completion_time = asyncio.get_running_loop().time()
         elapsed_time = (completion_time - start_time) * 1000
         await results_queue.put((request_id, elapsed_time, success, completion_time))
 
     except Exception as e:
         print(f"[HTTP] {api_name} Error for request {request_id}: {e}")
-        completion_time = asyncio.get_event_loop().time()
+        completion_time = asyncio.get_running_loop().time()
         await results_queue.put((request_id, 0, False, completion_time))
 
 
@@ -738,7 +738,7 @@ async def run_generic_benchmark(
     tasks = []
 
     # Track timing for sending requests
-    send_start_time = asyncio.get_event_loop().time()
+    send_start_time = asyncio.get_running_loop().time()
 
     # HTTP implementation
     async with aiohttp.ClientSession(
@@ -778,7 +778,7 @@ async def run_generic_benchmark(
                 if i < len(all_requests) - 1:
                     await sleep_with_distribution(config.distribution, rps)
 
-        send_end_time = asyncio.get_event_loop().time()
+        send_end_time = asyncio.get_running_loop().time()
         send_duration = send_end_time - send_start_time
 
         # Wait for all requests to complete with progress tracking
@@ -796,7 +796,7 @@ async def run_generic_benchmark(
         if config.profile:
             await send_profile_request("STOP_PROFILE", http_url, session=session)
 
-    completion_end_time = asyncio.get_event_loop().time()
+    completion_end_time = asyncio.get_running_loop().time()
     total_duration = completion_end_time - send_start_time
 
     return await process_results(

python/pyproject.toml

@@ -69,7 +69,7 @@ dependencies = [
   "tqdm",
   "transformers==4.57.1",
   "uvicorn",
-  "uvloop==0.21.0",
+  "uvloop",
   "xgrammar==0.1.27",
   "grpcio==1.75.1", # keep it align with compile_proto.py
   "grpcio-tools==1.75.1", # keep it align with compile_proto.py

python/sglang/srt/entrypoints/engine.py

@@ -533,8 +533,7 @@ class Engine(EngineBase):
             zmq_handles=zmq_handles,
             flush_cache=flush_cache,
         )
-        loop = asyncio.get_event_loop()
+        return self.loop.run_until_complete(
-        return loop.run_until_complete(
             self.tokenizer_manager.update_weights_from_ipc(obj, None)
         )
 

python/sglang/srt/grpc/grpc_request_manager.py

@@ -28,7 +28,7 @@ from sglang.srt.managers.io_struct import (
     TokenizedGenerateReqInput,
 )
 from sglang.srt.server_args import PortArgs, ServerArgs
-from sglang.srt.utils import get_zmq_socket, kill_process_tree
+from sglang.srt.utils import get_or_create_event_loop, get_zmq_socket, kill_process_tree
 from sglang.utils import get_exception_traceback
 
 logger = logging.getLogger(__name__)
@@ -876,7 +876,7 @@ class GrpcRequestManager:
             return
 
         self.no_create_loop = True
-        loop = asyncio.get_event_loop()
+        loop = get_or_create_event_loop()
         self.asyncio_tasks.add(
             loop.create_task(print_exception_wrapper(self.handle_loop))
         )

python/sglang/srt/managers/tokenizer_manager.py

@@ -98,6 +98,7 @@ from sglang.srt.utils import (
     dataclass_to_string_truncated,
     freeze_gc,
     get_bool_env_var,
+    get_or_create_event_loop,
     get_zmq_socket,
     kill_process_tree,
 )
@@ -1354,12 +1355,13 @@ class TokenizerManager(TokenizerCommunicatorMixin):
 
     def auto_create_handle_loop(self):
         if self._chosen_loop is not None:
+            current_loop = get_or_create_event_loop()
             assert (
-                asyncio.get_event_loop() == self._chosen_loop
+                current_loop == self._chosen_loop
-            ), f"Please ensure only one event loop is ever used with SGLang. Previous loop: {self._chosen_loop}, current loop: {asyncio.get_event_loop()}"
+            ), f"Please ensure only one event loop is ever used with SGLang. Previous loop: {self._chosen_loop}, current loop: {current_loop}"
             return
 
-        loop = asyncio.get_event_loop()
+        loop = get_or_create_event_loop()
         self._chosen_loop = loop
         self.asyncio_tasks.add(
             loop.create_task(print_exception_wrapper(self.handle_loop))

python/sglang/srt/multimodal/processors/llava.py

@@ -89,7 +89,7 @@ class LlavaImageProcessor(BaseMultimodalProcessor):
         grid_pinpoints: str,
     ):
         if self.cpu_executor is not None:
-            loop = asyncio.get_event_loop()
+            loop = asyncio.get_running_loop()
             return await loop.run_in_executor(
                 self.cpu_executor,
                 LlavaImageProcessor._process_single_image_task,

python/sglang/srt/utils/common.py

@@ -104,6 +104,16 @@ show_time_cost = False
 time_infos = {}
 
 
+def get_or_create_event_loop():
+    """Gets the running event loop or creates a new one if it doesn't exist."""
+    try:
+        return asyncio.get_running_loop()
+    except RuntimeError:
+        loop = asyncio.new_event_loop()
+        asyncio.set_event_loop(loop)
+        return loop
+
+
 HIP_FP8_E4M3_FNUZ_MAX = 224.0
 
 

python/sglang/test/few_shot_gsm8k_engine.py

@@ -8,6 +8,7 @@ from typing import Optional
 import numpy as np
 
 import sglang as sgl
+from sglang.srt.utils import get_or_create_event_loop
 from sglang.utils import download_and_cache_file, read_jsonl
 
 INVALID = -9999999
@@ -89,7 +90,7 @@ def run_eval(args):
     # Run requests
     tic = time.perf_counter()
 
-    loop = asyncio.get_event_loop()
+    loop = get_or_create_event_loop()
 
     outputs = loop.run_until_complete(
         concurrent_generate(engine, prompts, sampling_param)

test/srt/openai_server/basic/test_serving_chat.py

@@ -6,7 +6,6 @@ or
     python -m unittest discover -s tests -p "test_*unit.py" -v
 """
 
-import asyncio
 import json
 import unittest
 import uuid
@@ -21,6 +20,7 @@ from sglang.srt.entrypoints.openai.protocol import (
 )
 from sglang.srt.entrypoints.openai.serving_chat import OpenAIServingChat
 from sglang.srt.managers.io_struct import GenerateReqInput
+from sglang.srt.utils import get_or_create_event_loop
 
 
 class _MockTokenizerManager:
@@ -389,7 +389,7 @@ class ServingChatTestCase(unittest.TestCase):
                     break
                 return line
 
-            loop = asyncio.get_event_loop()
+            loop = get_or_create_event_loop()
             line = loop.run_until_complete(collect_first_tool_chunk())
             self.assertIsNotNone(line)
             self.assertTrue(line.startswith("data: "))
@@ -564,7 +564,7 @@ class ServingChatTestCase(unittest.TestCase):
                     break
                 return line
 
-            loop = asyncio.get_event_loop()
+            loop = get_or_create_event_loop()
             line = loop.run_until_complete(collect_first_tool_chunk())
             self.assertIsNotNone(line)
             self.assertTrue(line.startswith("data: "))