diff --git a/tests/function_libs/torch_lib/ops_test_common.py b/tests/function_libs/torch_lib/ops_test_common.py
index 34f5b5844..ae0578abd 100644
--- a/tests/function_libs/torch_lib/ops_test_common.py
+++ b/tests/function_libs/torch_lib/ops_test_common.py
@@ -8,6 +8,7 @@
 import multiprocessing
 import os
 import pprint
+import sys
 import unittest
 import warnings
 from typing import (
@@ -56,6 +57,7 @@
 )
 
 TEST_OPSET_VERSION = 18
+IS_MACOS = sys.platform.startswith("darwin")
 IS_WINDOWS = os.name == "nt"
 
 
diff --git a/tests/function_libs/torch_lib/ops_test_data.py b/tests/function_libs/torch_lib/ops_test_data.py
index b6fa826bf..cff34897d 100644
--- a/tests/function_libs/torch_lib/ops_test_data.py
+++ b/tests/function_libs/torch_lib/ops_test_data.py
@@ -1932,7 +1932,14 @@ def _where_input_wrangler(
         core_ops.aten_native_layer_norm,
         trace_only=True,
         tolerance={torch.float32: (3.7e-5, 1.8e-4), torch.float16: (1e-1, 7e-4)},
-    ).skip(
+    )
+    .xfail(
+        dtypes=(torch.float32,),
+        matcher=lambda sample: len(sample.input.shape) == 1,
+        enabled_if=ops_test_common.IS_MACOS and version_utils.onnxruntime_older_than("1.18"),
+        reason="fixme: result mismatch. https://github.com/microsoft/onnxruntime/issues/20676",
+    )
+    .skip(
         dtypes=(torch.float16,),
         device_type="cpu",
         reason="native_layer_norm outputs different dtypes on CPU and CUDA. Our implematation is based on that for CUDA",