diff --git a/python/sglang/bench_one_batch.py b/python/sglang/bench_one_batch.py
index 863bc5839..9bbe9b0f1 100644
--- a/python/sglang/bench_one_batch.py
+++ b/python/sglang/bench_one_batch.py
@@ -47,6 +47,7 @@ import itertools
 import json
 import logging
 import multiprocessing
+import os
 import time
 from typing import Tuple
 
@@ -62,11 +63,7 @@ from sglang.srt.model_executor.model_runner import ModelRunner
 from sglang.srt.sampling.sampling_params import SamplingParams
 from sglang.srt.server import _set_envs_and_config
 from sglang.srt.server_args import PortArgs, ServerArgs
-from sglang.srt.utils import (
-    configure_logger,
-    kill_child_process,
-    suppress_other_loggers,
-)
+from sglang.srt.utils import configure_logger, kill_process_tree, suppress_other_loggers
 
 
 @dataclasses.dataclass
@@ -468,4 +465,4 @@ if __name__ == "__main__":
         main(server_args, bench_args)
     finally:
         if server_args.tp_size != 1:
-            kill_child_process()
+            kill_process_tree(os.getpid(), include_parent=False)
diff --git a/python/sglang/bench_one_batch_server.py b/python/sglang/bench_one_batch_server.py
index 9d6048bc1..01cc561e1 100644
--- a/python/sglang/bench_one_batch_server.py
+++ b/python/sglang/bench_one_batch_server.py
@@ -15,6 +15,7 @@ import dataclasses
 import itertools
 import json
 import multiprocessing
+import os
 import time
 from typing import Tuple
 
@@ -23,7 +24,7 @@ import requests
 
 from sglang.srt.server import launch_server
 from sglang.srt.server_args import ServerArgs
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 
 
 @dataclasses.dataclass
@@ -69,7 +70,7 @@ def launch_server_internal(server_args):
     except Exception as e:
         raise e
     finally:
-        kill_child_process()
+        kill_process_tree(os.getpid(), include_parent=False)
 
 
 def launch_server_process(server_args: ServerArgs):
@@ -175,7 +176,7 @@ def run_benchmark(server_args: ServerArgs, bench_args: BenchArgs):
             )
     finally:
         if proc:
-            kill_child_process(proc.pid, include_self=True)
+            kill_process_tree(proc.pid)
 
     print(f"\nResults are saved to {bench_args.result_filename}")
 
diff --git a/python/sglang/launch_server.py b/python/sglang/launch_server.py
index 3e2cd4a97..b2ad1b320 100644
--- a/python/sglang/launch_server.py
+++ b/python/sglang/launch_server.py
@@ -4,7 +4,7 @@ import sys
 
 from sglang.srt.server import launch_server
 from sglang.srt.server_args import prepare_server_args
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 
 if __name__ == "__main__":
     server_args = prepare_server_args(sys.argv[1:])
@@ -12,4 +12,4 @@ if __name__ == "__main__":
     try:
         launch_server(server_args)
     finally:
-        kill_child_process()
+        kill_process_tree(os.getpid(), include_parent=False)
diff --git a/python/sglang/srt/managers/data_parallel_controller.py b/python/sglang/srt/managers/data_parallel_controller.py
index d4730e3f7..8edb79417 100644
--- a/python/sglang/srt/managers/data_parallel_controller.py
+++ b/python/sglang/srt/managers/data_parallel_controller.py
@@ -15,9 +15,11 @@
 
 import logging
 import multiprocessing as mp
+import signal
 import threading
 from enum import Enum, auto
 
+import psutil
 import zmq
 
 from sglang.srt.managers.io_struct import (
@@ -26,13 +28,7 @@ from sglang.srt.managers.io_struct import (
 )
 from sglang.srt.managers.scheduler import run_scheduler_process
 from sglang.srt.server_args import PortArgs, ServerArgs
-from sglang.srt.utils import (
-    bind_port,
-    configure_logger,
-    get_zmq_socket,
-    kill_parent_process,
-    suppress_other_loggers,
-)
+from sglang.srt.utils import bind_port, configure_logger, get_zmq_socket
 from sglang.utils import get_exception_traceback
 
 logger = logging.getLogger(__name__)
@@ -235,7 +231,7 @@ def run_data_parallel_controller_process(
     pipe_writer,
 ):
     configure_logger(server_args)
-    suppress_other_loggers()
+    parent_process = psutil.Process().parent()
 
     try:
         controller = DataParallelController(server_args, port_args)
@@ -244,6 +240,6 @@ def run_data_parallel_controller_process(
         )
         controller.event_loop()
     except Exception:
-        msg = get_exception_traceback()
-        logger.error(msg)
-        kill_parent_process()
+        traceback = get_exception_traceback()
+        logger.error(f"DataParallelController hit an exception: {traceback}")
+        parent_process.send_signal(signal.SIGQUIT)
diff --git a/python/sglang/srt/managers/detokenizer_manager.py b/python/sglang/srt/managers/detokenizer_manager.py
index 18f77424d..e74ba5026 100644
--- a/python/sglang/srt/managers/detokenizer_manager.py
+++ b/python/sglang/srt/managers/detokenizer_manager.py
@@ -15,9 +15,11 @@
 
 import dataclasses
 import logging
+import signal
 from collections import OrderedDict
 from typing import List, Union
 
+import psutil
 import zmq
 
 from sglang.srt.hf_transformers_utils import get_tokenizer
@@ -28,7 +30,7 @@ from sglang.srt.managers.io_struct import (
 )
 from sglang.srt.managers.schedule_batch import FINISH_MATCHED_STR, FINISH_MATCHED_TOKEN
 from sglang.srt.server_args import PortArgs, ServerArgs
-from sglang.srt.utils import configure_logger, get_zmq_socket, kill_parent_process
+from sglang.srt.utils import configure_logger, get_zmq_socket
 from sglang.utils import find_printable_text, get_exception_traceback
 
 logger = logging.getLogger(__name__)
@@ -193,11 +195,12 @@ def run_detokenizer_process(
     port_args: PortArgs,
 ):
     configure_logger(server_args)
+    parent_process = psutil.Process().parent()
 
     try:
         manager = DetokenizerManager(server_args, port_args)
         manager.event_loop()
     except Exception:
-        msg = get_exception_traceback()
-        logger.error(msg)
-        kill_parent_process()
+        traceback = get_exception_traceback()
+        logger.error(f"DetokenizerManager hit an exception: {traceback}")
+        parent_process.send_signal(signal.SIGQUIT)
diff --git a/python/sglang/srt/managers/scheduler.py b/python/sglang/srt/managers/scheduler.py
index 2563bb559..a327f37a2 100644
--- a/python/sglang/srt/managers/scheduler.py
+++ b/python/sglang/srt/managers/scheduler.py
@@ -15,6 +15,7 @@
 
 import logging
 import os
+import signal
 import threading
 import time
 import warnings
@@ -23,6 +24,7 @@ from concurrent import futures
 from types import SimpleNamespace
 from typing import List, Optional
 
+import psutil
 import torch
 import zmq
 
@@ -73,7 +75,6 @@ from sglang.srt.utils import (
     crash_on_warnings,
     get_bool_env_var,
     get_zmq_socket,
-    kill_parent_process,
     set_gpu_proc_affinity,
     set_random_seed,
     suppress_other_loggers,
@@ -316,6 +317,7 @@ class Scheduler:
         self.watchdog_timeout = server_args.watchdog_timeout
         t = threading.Thread(target=self.watchdog_thread, daemon=True)
         t.start()
+        self.parent_process = psutil.Process().parent()
 
         # Init profiler
         if os.getenv("SGLANG_TORCH_PROFILER_DIR", "") == "":
@@ -359,7 +361,7 @@ class Scheduler:
                     self.watchdog_last_time = time.time()
             time.sleep(self.watchdog_timeout / 2)
 
-        kill_parent_process()
+        self.parent_process.send_signal(signal.SIGQUIT)
 
     @torch.no_grad()
     def event_loop_normal(self):
@@ -1423,6 +1425,7 @@ def run_scheduler_process(
         configure_logger(server_args, prefix=f" DP{dp_rank} TP{tp_rank}")
 
     suppress_other_loggers()
+    parent_process = psutil.Process().parent()
 
     try:
         scheduler = Scheduler(server_args, port_args, gpu_id, tp_rank, dp_rank)
@@ -1434,6 +1437,6 @@ def run_scheduler_process(
         else:
             scheduler.event_loop_normal()
     except Exception:
-        msg = get_exception_traceback()
-        logger.error(msg)
-        kill_parent_process()
+        traceback = get_exception_traceback()
+        logger.error(f"Scheduler hit an exception: {traceback}")
+        parent_process.send_signal(signal.SIGQUIT)
diff --git a/python/sglang/srt/managers/tokenizer_manager.py b/python/sglang/srt/managers/tokenizer_manager.py
index 001ecc1eb..15518e9e5 100644
--- a/python/sglang/srt/managers/tokenizer_manager.py
+++ b/python/sglang/srt/managers/tokenizer_manager.py
@@ -58,7 +58,7 @@ from sglang.srt.managers.io_struct import (
 from sglang.srt.metrics.collector import TokenizerMetricsCollector
 from sglang.srt.sampling.sampling_params import SamplingParams
 from sglang.srt.server_args import PortArgs, ServerArgs
-from sglang.srt.utils import get_zmq_socket, kill_child_process
+from sglang.srt.utils import get_zmq_socket, kill_process_tree
 
 asyncio.set_event_loop_policy(uvloop.EventLoopPolicy())
 
@@ -532,7 +532,7 @@ class TokenizerManager:
             else:
                 break
 
-        kill_child_process(include_self=True)
+        kill_process_tree(os.getpid(), include_parent=True)
         sys.exit(0)
 
     async def handle_loop(self):
diff --git a/python/sglang/srt/managers/tp_worker_overlap_thread.py b/python/sglang/srt/managers/tp_worker_overlap_thread.py
index 3b53759a7..a5412094c 100644
--- a/python/sglang/srt/managers/tp_worker_overlap_thread.py
+++ b/python/sglang/srt/managers/tp_worker_overlap_thread.py
@@ -15,16 +15,19 @@
 
 import dataclasses
 import logging
+import signal
 import threading
 from queue import Queue
 from typing import Optional
 
+import psutil
 import torch
 
 from sglang.srt.managers.io_struct import UpdateWeightReqInput
 from sglang.srt.managers.schedule_batch import ModelWorkerBatch
 from sglang.srt.managers.tp_worker import TpModelWorker
 from sglang.srt.server_args import ServerArgs
+from sglang.utils import get_exception_traceback
 
 logger = logging.getLogger(__name__)
 
@@ -70,6 +73,7 @@ class TpModelWorkerClient:
             target=self.forward_thread_func,
         )
         self.forward_thread.start()
+        self.parent_process = psutil.Process().parent()
 
     def get_worker_info(self):
         return self.worker.get_worker_info()
@@ -87,8 +91,13 @@ class TpModelWorkerClient:
         )
 
     def forward_thread_func(self):
-        with torch.cuda.stream(self.forward_stream):
-            self.forward_thread_func_()
+        try:
+            with torch.cuda.stream(self.forward_stream):
+                self.forward_thread_func_()
+        except Exception:
+            traceback = get_exception_traceback()
+            logger.error(f"TpModelWorkerClient hit an exception: {traceback}")
+            self.parent_process.send_signal(signal.SIGQUIT)
 
     @torch.no_grad()
     def forward_thread_func_(self):
diff --git a/python/sglang/srt/server.py b/python/sglang/srt/server.py
index a4753a134..c95893067 100644
--- a/python/sglang/srt/server.py
+++ b/python/sglang/srt/server.py
@@ -23,6 +23,8 @@ import json
 import logging
 import multiprocessing as mp
 import os
+import signal
+import sys
 import threading
 import time
 from http import HTTPStatus
@@ -79,7 +81,7 @@ from sglang.srt.utils import (
     configure_logger,
     delete_directory,
     is_port_available,
-    kill_child_process,
+    kill_process_tree,
     maybe_set_triton_cache_manager,
     prepare_model_and_tokenizer,
     set_prometheus_multiproc_dir,
@@ -572,6 +574,15 @@ def _set_envs_and_config(server_args: ServerArgs):
             "at https://docs.flashinfer.ai/installation.html.",
         )
 
+    # Register the signal handler.
+    # The child processes will send SIGQUIT to this process when any error happens
+    # This process then clean up the whole process tree
+    def sigquit_handler(signum, frame):
+        kill_process_tree(os.getpid())
+
+    signal.signal(signal.SIGQUIT, sigquit_handler)
+
+    # Set mp start method
     mp.set_start_method("spawn", force=True)
 
 
@@ -598,7 +609,7 @@ def _wait_and_warmup(server_args, pipe_finish_writer):
         if pipe_finish_writer is not None:
             pipe_finish_writer.send(last_traceback)
         logger.error(f"Initialization failed. warmup error: {last_traceback}")
-        kill_child_process(include_self=True)
+        kill_process_tree(os.getpid())
         return
 
     model_info = res.json()
@@ -631,7 +642,7 @@ def _wait_and_warmup(server_args, pipe_finish_writer):
         if pipe_finish_writer is not None:
             pipe_finish_writer.send(last_traceback)
         logger.error(f"Initialization failed. warmup error: {last_traceback}")
-        kill_child_process(include_self=True)
+        kill_process_tree(os.getpid())
         return
 
     # logger.info(f"{res.json()=}")
@@ -700,7 +711,7 @@ class Runtime:
 
     def shutdown(self):
         if self.pid is not None:
-            kill_child_process(self.pid, include_self=True)
+            kill_process_tree(self.pid)
             self.pid = None
 
     def cache_prefix(self, prefix: str):
@@ -924,7 +935,7 @@ class Engine:
             return ret
 
     def shutdown(self):
-        kill_child_process()
+        kill_process_tree(os.getpid(), include_parent=False)
 
     def get_tokenizer(self):
         global tokenizer_manager
diff --git a/python/sglang/srt/utils.py b/python/sglang/srt/utils.py
index 4a974e2e7..46b4db8e8 100644
--- a/python/sglang/srt/utils.py
+++ b/python/sglang/srt/utils.py
@@ -443,26 +443,14 @@ def assert_pkg_version(pkg: str, min_version: str, message: str):
         )
 
 
-def kill_parent_process():
-    """Kill the parent process and all children of the parent process."""
-    current_process = psutil.Process()
-    parent_process = current_process.parent()
-    kill_child_process(
-        parent_process.pid, include_self=True, skip_pid=current_process.pid
-    )
-    try:
-        current_process.kill()
-    except psutil.NoSuchProcess:
-        pass
-
-
-def kill_child_process(pid=None, include_self=False, skip_pid=None):
-    """Kill the process and all its children process."""
-    if pid is None:
-        pid = os.getpid()
+def kill_process_tree(parent_pid, include_parent: bool = True, skip_pid: int = None):
+    """Kill the process and all its child processes."""
+    if parent_pid is None:
+        parent_pid = os.getpid()
+        include_parent = False
 
     try:
-        itself = psutil.Process(pid)
+        itself = psutil.Process(parent_pid)
     except psutil.NoSuchProcess:
         return
 
@@ -475,13 +463,13 @@ def kill_child_process(pid=None, include_self=False, skip_pid=None):
         except psutil.NoSuchProcess:
             pass
 
-    if include_self:
+    if include_parent:
         try:
             itself.kill()
 
             # Sometime processes cannot be killed with SIGKILL (e.g, PID=1 launched by kubernetes),
             # so we send an additional signal to kill them.
-            itself.send_signal(signal.SIGINT)
+            itself.send_signal(signal.SIGQUIT)
         except psutil.NoSuchProcess:
             pass
 
diff --git a/python/sglang/test/test_utils.py b/python/sglang/test/test_utils.py
index 308966844..3f6cce23d 100644
--- a/python/sglang/test/test_utils.py
+++ b/python/sglang/test/test_utils.py
@@ -22,7 +22,7 @@ from sglang.bench_serving import run_benchmark
 from sglang.global_config import global_config
 from sglang.lang.backend.openai import OpenAI
 from sglang.lang.backend.runtime_endpoint import RuntimeEndpoint
-from sglang.srt.utils import get_bool_env_var, kill_child_process
+from sglang.srt.utils import get_bool_env_var, kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.utils import get_exception_traceback
 
@@ -504,7 +504,7 @@ def run_unittest_files(files: List[str], timeout_per_file: float):
             )
             assert ret_code == 0
         except TimeoutError:
-            kill_child_process(process.pid, include_self=True)
+            kill_process_tree(process.pid)
             time.sleep(5)
             print(
                 f"\nTimeout after {timeout_per_file} seconds when running {filename}\n",
@@ -578,7 +578,7 @@ def run_bench_serving(
             run_benchmark(warmup_args)
         res = run_benchmark(args)
     finally:
-        kill_child_process(process.pid, include_self=True)
+        kill_process_tree(process.pid)
 
     assert res["completed"] == num_prompts
     return res
@@ -611,7 +611,7 @@ def run_bench_one_batch(model, other_args):
         lastline = output.split("\n")[-3]
         output_throughput = float(lastline.split(" ")[-2])
     finally:
-        kill_child_process(process.pid, include_self=True)
+        kill_process_tree(process.pid)
 
     return output_throughput
 
@@ -710,8 +710,8 @@ def run_and_check_memory_leak(
     workload_func(base_url, model)
 
     # Clean up everything
-    kill_child_process(process.pid, include_self=True)
-    kill_child_process(process.pid, include_self=True)
+    kill_process_tree(process.pid)
+    kill_process_tree(process.pid)
     stdout.close()
     stderr.close()
     if os.path.exists(STDOUT_FILENAME):
diff --git a/python/sglang/utils.py b/python/sglang/utils.py
index e694dc198..c1bf62ef9 100644
--- a/python/sglang/utils.py
+++ b/python/sglang/utils.py
@@ -348,9 +348,9 @@ def wait_for_server(base_url: str, timeout: int = None) -> None:
 
 
 def terminate_process(process):
-    from sglang.srt.utils import kill_child_process
+    from sglang.srt.utils import kill_process_tree
 
-    kill_child_process(process.pid, include_self=True)
+    kill_process_tree(process.pid)
 
 
 def print_highlight(html_content: str):
diff --git a/rust/py_test/test_launch_server.py b/rust/py_test/test_launch_server.py
index 7fdaea6b1..f39b341df 100644
--- a/rust/py_test/test_launch_server.py
+++ b/rust/py_test/test_launch_server.py
@@ -5,7 +5,7 @@ from types import SimpleNamespace
 
 import requests
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST,
@@ -79,7 +79,7 @@ class TestEvalAccuracyMini(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def test_mmlu(self):
         args = SimpleNamespace(
diff --git a/test/srt/sampling/penaltylib/test_srt_endpoint_with_penalizers.py b/test/srt/sampling/penaltylib/test_srt_endpoint_with_penalizers.py
index 689d52a1c..0eccb3407 100644
--- a/test/srt/sampling/penaltylib/test_srt_endpoint_with_penalizers.py
+++ b/test/srt/sampling/penaltylib/test_srt_endpoint_with_penalizers.py
@@ -4,7 +4,7 @@ from multiprocessing import Process
 
 import requests
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -31,7 +31,7 @@ class TestBatchPenalizerE2E(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def run_decode(
         self,
diff --git a/test/srt/test_cache_report.py b/test/srt/test_cache_report.py
index 5d498ac3f..f128aa147 100644
--- a/test/srt/test_cache_report.py
+++ b/test/srt/test_cache_report.py
@@ -4,7 +4,7 @@ import unittest
 import openai
 import requests
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.test_utils import (
     DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
     DEFAULT_URL_FOR_TEST,
@@ -44,7 +44,7 @@ class TestCacheReport(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def run_decode(self, return_logprob=False, top_logprobs_num=0, n=1):
         response = requests.post(
diff --git a/test/srt/test_data_parallelism.py b/test/srt/test_data_parallelism.py
index f34313ea0..22d000664 100644
--- a/test/srt/test_data_parallelism.py
+++ b/test/srt/test_data_parallelism.py
@@ -4,7 +4,7 @@ from types import SimpleNamespace
 
 import requests
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST,
@@ -28,7 +28,7 @@ class TestDataParallelism(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def test_mmlu(self):
         args = SimpleNamespace(
diff --git a/test/srt/test_double_sparsity.py b/test/srt/test_double_sparsity.py
index 20896aff2..060a7926f 100644
--- a/test/srt/test_double_sparsity.py
+++ b/test/srt/test_double_sparsity.py
@@ -2,7 +2,7 @@ import os
 import unittest
 from types import SimpleNamespace
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST,
@@ -45,7 +45,7 @@ class TestDoubleSparsity(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def test_mmlu(self):
         args = SimpleNamespace(
diff --git a/test/srt/test_dp_attention.py b/test/srt/test_dp_attention.py
index 32fe75a59..31c9cc71b 100644
--- a/test/srt/test_dp_attention.py
+++ b/test/srt/test_dp_attention.py
@@ -1,7 +1,7 @@
 import unittest
 from types import SimpleNamespace
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_MLA_MODEL_NAME_FOR_TEST,
@@ -30,7 +30,7 @@ class TestDPAttention(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def test_mmlu(self):
         args = SimpleNamespace(
diff --git a/test/srt/test_embedding_openai_server.py b/test/srt/test_embedding_openai_server.py
index 666297c65..8097bf42c 100644
--- a/test/srt/test_embedding_openai_server.py
+++ b/test/srt/test_embedding_openai_server.py
@@ -3,7 +3,7 @@ import unittest
 import openai
 
 from sglang.srt.hf_transformers_utils import get_tokenizer
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
@@ -28,7 +28,7 @@ class TestOpenAIServer(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def run_embedding(self, use_list_input, token_input):
         client = openai.Client(api_key=self.api_key, base_url=self.base_url)
diff --git a/test/srt/test_eval_accuracy_large.py b/test/srt/test_eval_accuracy_large.py
index 318390d10..f7fb3cec3 100644
--- a/test/srt/test_eval_accuracy_large.py
+++ b/test/srt/test_eval_accuracy_large.py
@@ -6,7 +6,7 @@ python -m unittest test_eval_accuracy_large.TestEvalAccuracyLarge.test_mmlu
 import unittest
 from types import SimpleNamespace
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST,
@@ -30,7 +30,7 @@ class TestEvalAccuracyLarge(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def test_mmlu(self):
         args = SimpleNamespace(
diff --git a/test/srt/test_eval_accuracy_large_chunked_prefill.py b/test/srt/test_eval_accuracy_large_chunked_prefill.py
index 2e9ff59cd..c8ce5cff2 100644
--- a/test/srt/test_eval_accuracy_large_chunked_prefill.py
+++ b/test/srt/test_eval_accuracy_large_chunked_prefill.py
@@ -1,7 +1,7 @@
 import unittest
 from types import SimpleNamespace
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST,
@@ -25,7 +25,7 @@ class TestEvalAccuracyLargeChunkedPrefill(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def test_mmlu(self):
         args = SimpleNamespace(
diff --git a/test/srt/test_eval_accuracy_large_mixed_chunked_prefill.py b/test/srt/test_eval_accuracy_large_mixed_chunked_prefill.py
index 0fb08e64f..3bc115874 100644
--- a/test/srt/test_eval_accuracy_large_mixed_chunked_prefill.py
+++ b/test/srt/test_eval_accuracy_large_mixed_chunked_prefill.py
@@ -1,7 +1,7 @@
 import unittest
 from types import SimpleNamespace
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST,
@@ -31,7 +31,7 @@ class TestEvalAccuracyLargeChunkedPrefill(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def test_mmlu(self):
         args = SimpleNamespace(
diff --git a/test/srt/test_eval_accuracy_mini.py b/test/srt/test_eval_accuracy_mini.py
index a718feff7..a008c3869 100644
--- a/test/srt/test_eval_accuracy_mini.py
+++ b/test/srt/test_eval_accuracy_mini.py
@@ -1,7 +1,7 @@
 import unittest
 from types import SimpleNamespace
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST,
@@ -22,7 +22,7 @@ class TestEvalAccuracyMini(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def test_mmlu(self):
         args = SimpleNamespace(
diff --git a/test/srt/test_input_embeddings.py b/test/srt/test_input_embeddings.py
index b57b61dad..04d54c6bb 100644
--- a/test/srt/test_input_embeddings.py
+++ b/test/srt/test_input_embeddings.py
@@ -4,7 +4,7 @@ import unittest
 import requests
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.test_utils import (
     DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -107,7 +107,7 @@ class TestInputEmbeds(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
 
 if __name__ == "__main__":
diff --git a/test/srt/test_json_constrained.py b/test/srt/test_json_constrained.py
index ae27b036f..28acdabd9 100644
--- a/test/srt/test_json_constrained.py
+++ b/test/srt/test_json_constrained.py
@@ -9,7 +9,7 @@ from concurrent.futures import ThreadPoolExecutor
 import openai
 import requests
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST,
     DEFAULT_URL_FOR_TEST,
@@ -46,7 +46,7 @@ class TestJSONConstrainedOutlinesBackend(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def run_decode(self, json_schema, return_logprob=False, top_logprobs_num=0, n=1):
         response = requests.post(
diff --git a/test/srt/test_large_max_new_tokens.py b/test/srt/test_large_max_new_tokens.py
index 5ed2b06fc..dcaeef5aa 100644
--- a/test/srt/test_large_max_new_tokens.py
+++ b/test/srt/test_large_max_new_tokens.py
@@ -10,7 +10,7 @@ from concurrent.futures import ThreadPoolExecutor
 import openai
 
 from sglang.srt.hf_transformers_utils import get_tokenizer
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.test_utils import (
     DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -52,7 +52,7 @@ class TestLargeMaxNewTokens(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
         cls.stdout.close()
         cls.stderr.close()
         os.remove(STDOUT_FILENAME)
diff --git a/test/srt/test_matched_stop.py b/test/srt/test_matched_stop.py
index 81d08b091..7b09a6d35 100644
--- a/test/srt/test_matched_stop.py
+++ b/test/srt/test_matched_stop.py
@@ -3,7 +3,7 @@ import unittest
 
 import requests
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST,
     DEFAULT_URL_FOR_TEST,
@@ -32,7 +32,7 @@ class TestMatchedStop(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def run_completions_generation(
         self,
diff --git a/test/srt/test_metrics.py b/test/srt/test_metrics.py
index 163a7cc0e..3b73e500d 100644
--- a/test/srt/test_metrics.py
+++ b/test/srt/test_metrics.py
@@ -2,7 +2,7 @@ import unittest
 
 import requests
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.test_utils import (
     DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -75,7 +75,7 @@ class TestEnableMetrics(unittest.TestCase):
             self.assertIn("_bucket{", metrics_content)
 
         finally:
-            kill_child_process(process.pid, include_self=True)
+            kill_process_tree(process.pid)
 
 
 if __name__ == "__main__":
diff --git a/test/srt/test_mla.py b/test/srt/test_mla.py
index a11be3950..b8105a84a 100644
--- a/test/srt/test_mla.py
+++ b/test/srt/test_mla.py
@@ -1,7 +1,7 @@
 import unittest
 from types import SimpleNamespace
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_MLA_MODEL_NAME_FOR_TEST,
@@ -25,7 +25,7 @@ class TestMLA(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def test_mmlu(self):
         args = SimpleNamespace(
diff --git a/test/srt/test_mla_fp8.py b/test/srt/test_mla_fp8.py
index 5091759a9..769bdf34d 100644
--- a/test/srt/test_mla_fp8.py
+++ b/test/srt/test_mla_fp8.py
@@ -1,7 +1,7 @@
 import unittest
 from types import SimpleNamespace
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_MLA_FP8_MODEL_NAME_FOR_TEST,
@@ -31,7 +31,7 @@ class TestMLA(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def test_mgsm_en(self):
         args = SimpleNamespace(
diff --git a/test/srt/test_moe_eval_accuracy_large.py b/test/srt/test_moe_eval_accuracy_large.py
index 9880a8162..6f3affbba 100644
--- a/test/srt/test_moe_eval_accuracy_large.py
+++ b/test/srt/test_moe_eval_accuracy_large.py
@@ -6,7 +6,7 @@ python -m unittest test_moe_eval_accuracy_large.TestMoEEvalAccuracyLarge.test_mm
 import unittest
 from types import SimpleNamespace
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_MOE_MODEL_NAME_FOR_TEST,
@@ -35,7 +35,7 @@ class TestMoEEvalAccuracyLarge(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def test_mmlu(self):
         args = SimpleNamespace(
diff --git a/test/srt/test_nightly_gsm8k_eval.py b/test/srt/test_nightly_gsm8k_eval.py
index 7c208e84b..8466c2c64 100644
--- a/test/srt/test_nightly_gsm8k_eval.py
+++ b/test/srt/test_nightly_gsm8k_eval.py
@@ -6,7 +6,7 @@ import warnings
 from datetime import datetime
 from types import SimpleNamespace
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_FP8_TP1,
@@ -132,7 +132,7 @@ class TestEvalAccuracyLarge(unittest.TestCase):
 
     def tearDown(self):
         if self.process:
-            kill_child_process(self.process.pid, include_self=True)
+            kill_process_tree(self.process.pid)
 
     def test_mgsm_en_all_models(self):
         warnings.filterwarnings(
diff --git a/test/srt/test_nightly_human_eval.py b/test/srt/test_nightly_human_eval.py
index f69bbe132..626e6fb15 100644
--- a/test/srt/test_nightly_human_eval.py
+++ b/test/srt/test_nightly_human_eval.py
@@ -6,7 +6,7 @@ import unittest
 
 from test_nightly_gsm8k_eval import launch_server, parse_models
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_FP8_TP1,
     DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_FP8_TP2,
@@ -32,9 +32,9 @@ class TestEvalAccuracyLarge(unittest.TestCase):
     @classmethod
     def tearDownClass(cls):
         if cls.process:
-            kill_child_process(cls.process.pid)
+            kill_process_tree(cls.process.pid)
         if cls.eval_process:
-            kill_child_process(cls.eval_process.pid)
+            kill_process_tree(cls.eval_process.pid)
 
     def run_evalplus(self, model):
         print("Delete evalplus results")
diff --git a/test/srt/test_openai_server.py b/test/srt/test_openai_server.py
index 1e18e23ef..d007bed31 100644
--- a/test/srt/test_openai_server.py
+++ b/test/srt/test_openai_server.py
@@ -11,7 +11,7 @@ import unittest
 import openai
 
 from sglang.srt.hf_transformers_utils import get_tokenizer
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.test_utils import (
     DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -37,7 +37,7 @@ class TestOpenAIServer(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def run_completion(
         self, echo, logprobs, use_list_input, parallel_sample_num, token_input
diff --git a/test/srt/test_pytorch_sampling_backend.py b/test/srt/test_pytorch_sampling_backend.py
index 9aa6c3300..4f1403e0a 100644
--- a/test/srt/test_pytorch_sampling_backend.py
+++ b/test/srt/test_pytorch_sampling_backend.py
@@ -3,7 +3,7 @@ from types import SimpleNamespace
 
 import requests
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST,
@@ -27,7 +27,7 @@ class TestPyTorchSamplingBackend(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def test_mmlu(self):
         args = SimpleNamespace(
diff --git a/test/srt/test_radix_attention.py b/test/srt/test_radix_attention.py
index cdba7573d..207303c8c 100644
--- a/test/srt/test_radix_attention.py
+++ b/test/srt/test_radix_attention.py
@@ -8,7 +8,7 @@ from sglang.test.test_utils import (
     DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
-    kill_child_process,
+    kill_process_tree,
     popen_launch_server,
 )
 
@@ -80,7 +80,7 @@ class TestRadixCacheFCFS(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def test_radix_attention(self):
         nodes = gen_radix_tree()
diff --git a/test/srt/test_retract_decode.py b/test/srt/test_retract_decode.py
index 834c51f9d..5f169cdb6 100644
--- a/test/srt/test_retract_decode.py
+++ b/test/srt/test_retract_decode.py
@@ -1,7 +1,7 @@
 import unittest
 from types import SimpleNamespace
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST,
@@ -22,7 +22,7 @@ class TestRetractDecode(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def test_mmlu(self):
         args = SimpleNamespace(
diff --git a/test/srt/test_session_control.py b/test/srt/test_session_control.py
index 7396779f6..8558b4249 100644
--- a/test/srt/test_session_control.py
+++ b/test/srt/test_session_control.py
@@ -9,7 +9,7 @@ import unittest
 import requests
 
 from sglang.srt.hf_transformers_utils import get_tokenizer
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.test_utils import (
     DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -29,7 +29,7 @@ class TestSessionControl(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def test_session_control(self):
         chunks = [
@@ -191,7 +191,7 @@ class TestSessionControlVision(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def test_session_control(self):
         text_chunks = [
diff --git a/test/srt/test_skip_tokenizer_init.py b/test/srt/test_skip_tokenizer_init.py
index 7ec73b15d..bc99b23ad 100644
--- a/test/srt/test_skip_tokenizer_init.py
+++ b/test/srt/test_skip_tokenizer_init.py
@@ -7,7 +7,7 @@ import unittest
 
 import requests
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.test_utils import (
     DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -30,7 +30,7 @@ class TestSkipTokenizerInit(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def run_decode(self, return_logprob=False, top_logprobs_num=0, n=1):
         max_new_tokens = 32
diff --git a/test/srt/test_srt_endpoint.py b/test/srt/test_srt_endpoint.py
index fb50943f1..006059e03 100644
--- a/test/srt/test_srt_endpoint.py
+++ b/test/srt/test_srt_endpoint.py
@@ -9,7 +9,7 @@ import unittest
 import numpy as np
 import requests
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.test_utils import (
     DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -29,7 +29,7 @@ class TestSRTEndpoint(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def run_decode(
         self,
diff --git a/test/srt/test_torch_compile.py b/test/srt/test_torch_compile.py
index 76945f963..6f3b344b3 100644
--- a/test/srt/test_torch_compile.py
+++ b/test/srt/test_torch_compile.py
@@ -4,7 +4,7 @@ from types import SimpleNamespace
 
 import requests
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST,
@@ -28,7 +28,7 @@ class TestTorchCompile(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def test_mmlu(self):
         args = SimpleNamespace(
diff --git a/test/srt/test_torch_compile_moe.py b/test/srt/test_torch_compile_moe.py
index e744e6686..89d4ed6bd 100644
--- a/test/srt/test_torch_compile_moe.py
+++ b/test/srt/test_torch_compile_moe.py
@@ -4,7 +4,7 @@ from types import SimpleNamespace
 
 import requests
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_SMALL_MOE_MODEL_NAME_FOR_TEST,
@@ -28,7 +28,7 @@ class TestTorchCompile(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def test_mmlu(self):
         args = SimpleNamespace(
diff --git a/test/srt/test_torchao.py b/test/srt/test_torchao.py
index 2a2fcb8df..a6414c60b 100644
--- a/test/srt/test_torchao.py
+++ b/test/srt/test_torchao.py
@@ -3,7 +3,7 @@ from types import SimpleNamespace
 
 import requests
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST,
@@ -27,7 +27,7 @@ class TestTorchAO(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def test_mmlu(self):
         args = SimpleNamespace(
diff --git a/test/srt/test_triton_attention_backend.py b/test/srt/test_triton_attention_backend.py
index a4d19bec0..905590965 100644
--- a/test/srt/test_triton_attention_backend.py
+++ b/test/srt/test_triton_attention_backend.py
@@ -6,7 +6,7 @@ python3 -m unittest test_triton_attention_backend.TestTritonAttnBackend.test_mml
 import unittest
 from types import SimpleNamespace
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST,
@@ -54,7 +54,7 @@ class TestTritonAttnBackend(unittest.TestCase):
             metrics = run_eval(args)
             self.assertGreaterEqual(metrics["score"], 0.65)
         finally:
-            kill_child_process(process.pid, include_self=True)
+            kill_process_tree(process.pid)
 
 
 if __name__ == "__main__":
diff --git a/test/srt/test_update_weights.py b/test/srt/test_update_weights.py
index 327da729a..ddb5a5e08 100644
--- a/test/srt/test_update_weights.py
+++ b/test/srt/test_update_weights.py
@@ -3,7 +3,7 @@ import unittest
 
 import requests
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.test_utils import (
     DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -23,7 +23,7 @@ class TestUpdateWeights(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def run_decode(self):
         response = requests.post(
diff --git a/test/srt/test_vision_openai_server.py b/test/srt/test_vision_openai_server.py
index 95a1624cf..e19e6b01d 100644
--- a/test/srt/test_vision_openai_server.py
+++ b/test/srt/test_vision_openai_server.py
@@ -17,7 +17,7 @@ import requests
 from decord import VideoReader, cpu
 from PIL import Image
 
-from sglang.srt.utils import kill_child_process
+from sglang.srt.utils import kill_process_tree
 from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
@@ -46,7 +46,7 @@ class TestOpenAIVisionServer(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def test_chat_completion(self):
         client = openai.Client(api_key=self.api_key, base_url=self.base_url)
@@ -387,7 +387,7 @@ class TestQWen2VLServerContextLengthIssue(unittest.TestCase):
 
     @classmethod
     def tearDownClass(cls):
-        kill_child_process(cls.process.pid, include_self=True)
+        kill_process_tree(cls.process.pid)
 
     def test_chat_completion(self):
         client = openai.Client(api_key=self.api_key, base_url=self.base_url)