diff --git a/test/registered/profiling/test_start_profile.py b/test/registered/profiling/test_start_profile.py
index df7fd579a..10b8b31d8 100644
--- a/test/registered/profiling/test_start_profile.py
+++ b/test/registered/profiling/test_start_profile.py
@@ -19,7 +19,7 @@ import unittest
 import requests
 
 from sglang.srt.environ import envs
-from sglang.srt.utils import kill_process_tree
+from sglang.srt.utils import is_hip, kill_process_tree
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import (
     DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
@@ -51,10 +51,17 @@ class TestStartProfile(CustomTestCase):
         envs.SGLANG_TORCH_PROFILER_DIR.set(OUTPUT_DIR)
         cls.model = DEFAULT_SMALL_MODEL_NAME_FOR_TEST
         cls.base_url = DEFAULT_URL_FOR_TEST
+        # ROCm's HIP runtime can deadlock inside hipGraphLaunch while a
+        # torch-profiler (rocprofiler-sdk) session is attached, wedging the
+        # scheduler in at::cuda::CUDAGraph::replay until the watchdog kills the
+        # server. These tests cover the /start_profile request surface rather
+        # than graph replay, so run them eagerly on HIP.
+        other_args = ["--disable-cuda-graph"] if is_hip() else []
         cls.process = popen_launch_server(
             cls.model,
             cls.base_url,
             timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=other_args,
         )
 
     @classmethod