diff --git a/tests/benchmarks/test_bench_startup.py b/tests/benchmarks/test_bench_startup.py
new file mode 100644
index 000000000..44c9bc9b7
--- /dev/null
+++ b/tests/benchmarks/test_bench_startup.py
@@ -0,0 +1,19 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import subprocess
+
+import pytest
+
+
+@pytest.mark.benchmark
+def test_bench_startup():
+    command = [
+        "vllm",
+        "bench",
+        "startup",
+    ]
+    result = subprocess.run(command, capture_output=True, text=True)
+    print(result.stdout)
+    print(result.stderr)
+
+    assert result.returncode == 0, f"Benchmark failed: {result.stderr}"
diff --git a/vllm/benchmarks/startup.py b/vllm/benchmarks/startup.py
index 086f7bf62..d0d46a963 100644
--- a/vllm/benchmarks/startup.py
+++ b/vllm/benchmarks/startup.py
@@ -55,7 +55,7 @@ def cold_startup():
             os.environ.pop("VLLM_CACHE_ROOT", None)
 
 
-def run_startup_in_subprocess(engine_args_dict, result_queue):
+def run_startup_in_subprocess(engine_args, result_queue):
     """
     Run LLM startup in a subprocess and return timing metrics via a queue.
     This ensures complete isolation between iterations.
@@ -63,9 +63,6 @@ def run_startup_in_subprocess(engine_args_dict, result_queue):
     try:
         # Import inside the subprocess to avoid issues with forking
         from vllm import LLM
-        from vllm.engine.arg_utils import EngineArgs
-
-        engine_args = EngineArgs(**engine_args_dict)
 
         # Measure total startup time
         start_time = time.perf_counter()
@@ -200,15 +197,13 @@ def main(args: argparse.Namespace):
         Create LLM instance in a subprocess and measure startup time.
         Returns timing metrics, using subprocess for complete isolation.
         """
-        # Convert engine_args to dictionary for pickling
-        engine_args_dict = dataclasses.asdict(engine_args)
 
         # Create a queue for inter-process communication
         result_queue = multiprocessing.Queue()
         process = multiprocessing.Process(
             target=run_startup_in_subprocess,
             args=(
-                engine_args_dict,
+                engine_args,
                 result_queue,
             ),
         )