add python function def test_shape_reshape

onnxruntime/test/python/onnxruntime_test_python_cudagraph.py

@@ -127,6 +127,31 @@ def run_model_with_cuda_graph(self, providers):
             atol=1e-05,
         )
 
+    def test_shape_reshape(self):
+        providers = [("CUDAExecutionProvider", {"enable_cuda_graph": True})]
+        x = np.random.rand(1,3,60,10).astype(np.float32)
+        y = np.random.rand(1,3,60,10).astype(np.float32)
+        x_ortvalue = onnxrt.OrtValue.ortvalue_from_numpy(x, "cuda", 0)
+        y_ortvalue = onnxrt.OrtValue.ortvalue_from_numpy(y, "cuda", 0)
+
+        onnxrt.set_default_logger_severity(0)
+        session = onnxrt.InferenceSession("/bert_ort/leca/models/prod_model11/prod_model11.onnx", providers=providers)
+        io_binding = session.io_binding()
+
+        # Bind the input and output
+        io_binding.bind_ortvalue_input("data", x_ortvalue)
+        io_binding.bind_ortvalue_output("score", y_ortvalue)
+
+        # One regular run for the necessary memory allocation and cuda graph capturing
+        session.run_with_iobinding(io_binding)
+
+        # After capturing, CUDA graph replay happens from this Run onwards
+        session.run_with_iobinding(io_binding)
+
+        # Update input and then replay CUDA graph
+        x_ortvalue.update_inplace(np.random.rand(1,3,60,10).astype(np.float32))
+        session.run_with_iobinding(io_binding)
+
     def run_model_with_cuda_graph_annotation(self, providers):
         INPUT_SIZE = 1280  # noqa: N806