diff --git a/tests/v1/worker/test_gpu_model_runner.py b/tests/v1/worker/test_gpu_model_runner.py
index 0439bef1226e3..459abcfdd53cf 100644
--- a/tests/v1/worker/test_gpu_model_runner.py
+++ b/tests/v1/worker/test_gpu_model_runner.py
@@ -761,6 +761,10 @@ def test_init_kv_cache_with_kv_sharing_valid():
     assert kv_cache_config_after_init.kv_cache_groups[0].layer_names[1] == layer_1
 
 
+@pytest.mark.skipif(
+    current_platform.is_rocm(),
+    reason="Attention backend FLASHINFER is not supported on ROCm.",
+)
 def test_hybrid_attention_mamba_tensor_shapes(monkeypatch):
     """
     The GPU model runner creates different views into the