diff --git a/python/sglang/test/kits/mmmu_vlm_kit.py b/python/sglang/test/kits/mmmu_vlm_kit.py
index ad82b25d2..8c587037a 100644
--- a/python/sglang/test/kits/mmmu_vlm_kit.py
+++ b/python/sglang/test/kits/mmmu_vlm_kit.py
@@ -7,7 +7,7 @@ import tempfile
 from pathlib import Path
 from types import SimpleNamespace
 
-from sglang.srt.utils import kill_process_tree
+from sglang.srt.utils import is_cuda, kill_process_tree
 from sglang.srt.utils.common import temp_set_env
 from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -349,8 +349,11 @@ class MMMUMultiModelTestBase(CustomTestCase):
             process_env = os.environ.copy()
             if custom_env:
                 process_env.update(custom_env)
-            # if test vlm with cuda_ipc feature, open this env_var
-            process_env["SGLANG_USE_CUDA_IPC_TRANSPORT"] = "1"
+            # CUDA IPC multimodal-feature transport is NVIDIA-only; enabling it
+            # on ROCm/HIP makes ServerArgs raise "requires NVIDIA CUDA" and the
+            # server exits. Only opt in on CUDA so AMD exercises CPU transport.
+            if is_cuda():
+                process_env["SGLANG_USE_CUDA_IPC_TRANSPORT"] = "1"
 
             # Prepare stdout/stderr redirection if needed
             stdout_file = None