fix bug: quantization shape inference (#12983)

model path for onnx.shape_inference.infer_shapes_path and the external
data needs to be under the same directory as doc here:
f4dea9e68b/docs/PythonAPIOverview.md (shape-inference-a-large-onnx-model-2gb)

onnxruntime/python/tools/quantization/quant_utils.py

@@ -545,10 +545,10 @@ def model_has_infer_metadata(model):
 
 
 def load_model_with_shape_infer(model_path: Path):
-    with tempfile.TemporaryDirectory() as temp_dir:
-        inferred_model_path = str(Path(temp_dir) / (model_path.stem + "-inferred.onnx"))
-        onnx.shape_inference.infer_shapes_path(str(model_path), inferred_model_path)
-        model = onnx.load(inferred_model_path)
+    inferred_model_path = generate_identified_filename(model_path, "-inferred")
+    onnx.shape_inference.infer_shapes_path(str(model_path), str(inferred_model_path))
+    model = onnx.load(inferred_model_path.as_posix())
+    inferred_model_path.unlink()
     return model
 
 

onnxruntime/test/python/quantization/test_quant_util.py

@@ -6,11 +6,15 @@
 # license information.
 # --------------------------------------------------------------------------
 
+import tempfile
 import unittest
+from pathlib import Path
 
 import numpy
+import onnx
+from onnx import TensorProto, helper, numpy_helper
 
-from onnxruntime.quantization.quant_utils import compute_scale_zp
+from onnxruntime.quantization.quant_utils import compute_scale_zp, load_model, model_has_infer_metadata
 
 
 class TestQuantUtil(unittest.TestCase):
@@ -30,6 +34,37 @@ class TestQuantUtil(unittest.TestCase):
         self.assertEqual(compute_scale_zp(-tiny_float, tiny_float, 0, 255, symmetric=True), [0, 1.0])
         self.assertEqual(compute_scale_zp(-tiny_float, 0.0, 0, 255, symmetric=False), [0, 1.0])
 
+    def test_load_external_model(self):
+        input_name = "input"
+        output_name = "output"
+        add_shape = [1024, 1024]
+
+        initializers = []
+        weight_name = "weight"
+        weight_data = numpy.random.normal(0, 0.1, add_shape).astype(numpy.float32)
+        initializers.append(numpy_helper.from_array(weight_data, name=weight_name))
+        add_node = helper.make_node("Add", [input_name, weight_name], [output_name], name="add_node")
+
+        # make graph
+        input_tensor = helper.make_tensor_value_info(input_name, TensorProto.FLOAT, add_shape)
+        output_tensor = helper.make_tensor_value_info(output_name, TensorProto.FLOAT, add_shape)
+        graph_name = "test_load_external_model"
+        graph = helper.make_graph(
+            [add_node],
+            graph_name,
+            [input_tensor],
+            [output_tensor],
+            initializer=initializers,
+        )
+        model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
+
+        with tempfile.TemporaryDirectory() as temp_dir:
+            self.assertFalse(model_has_infer_metadata(model))
+            model_file_path = temp_dir + "/test_load_external_model.onnx"
+            onnx.save(model, model_file_path, save_as_external_data=True)
+            model_reloaded = load_model(Path(model_file_path), False)
+            self.assertTrue(model_has_infer_metadata(model_reloaded))
+
 
 if __name__ == "__main__":
     unittest.main()