diff --git a/onnxscript/rewriter/ort_fusions/_test_utils.py b/onnxscript/rewriter/ort_fusions/_test_utils.py
index 24e9bcce61..35092ddf45 100644
--- a/onnxscript/rewriter/ort_fusions/_test_utils.py
+++ b/onnxscript/rewriter/ort_fusions/_test_utils.py
@@ -2,6 +2,9 @@
 # Licensed under the MIT License.
 from __future__ import annotations
 
+import os
+import tempfile
+
 import numpy as np
 import onnx_ir as ir
 import onnxruntime
@@ -12,13 +15,17 @@
 
 def ort_run(model_name: str, model, inputs):
     providers = ["CPUExecutionProvider"]
-    model_proto = ir.serde.serialize_model(model)
     options = onnxruntime.SessionOptions()
     options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_DISABLE_ALL
-    session = onnxruntime.InferenceSession(
-        model_proto.SerializeToString(), options, providers=providers
-    )
-    return session.run(None, inputs)
+    # Save the model to a temporary file and load it from disk. This uses external data
+    # to store tensors so that large models are not limited by the 2GB protobuf
+    # serialization limit, which can otherwise raise
+    # ``google.protobuf.message.EncodeError: Failed to serialize proto``.
+    with tempfile.TemporaryDirectory() as temp_dir:
+        model_path = os.path.join(temp_dir, f"{model_name}.onnx")
+        ir.save(model, model_path, external_data=f"{model_name}.onnx.data")
+        session = onnxruntime.InferenceSession(model_path, options, providers=providers)
+        return session.run(None, inputs)
 
 
 def assert_allclose(outputs, expected_outputs, rtol=1e-3, atol=1e-3):