test_fx_to_onnx_with_onnxruntime.py

# Owner(s): ["module: onnx"]
from __future__ import annotations

import itertools
import os
import tempfile
import unittest

from typing import Any, Callable, Dict, List, Mapping, Optional, Tuple, Type

import onnx_test_common
import onnxruntime  # type: ignore[import]
import parameterized
import pytorch_test_common
import torch
import torch.onnx
import transformers  # type: ignore[import]
from torch import nn

from torch._subclasses import fake_tensor
from torch.onnx._internal import _beartype
from torch.onnx._internal.fx import (
    fx_symbolic_graph_extractor,
    patcher,
    serialization as fx_serialization,
)
from torch.testing._internal import common_utils

try:
    import torchvision

    HAS_TORCHVISION = True
except ImportError:
    HAS_TORCHVISION = False
except RuntimeError:
    HAS_TORCHVISION = False
skip_if_no_torchvision = unittest.skipIf(not HAS_TORCHVISION, "no torchvision")


def _parameterized_class_attrs_and_values():
    input_values = []
    input_values.extend(
        itertools.product(
            (True, False),
            (True, False),
        )
    )
    return {
        "attrs": ["op_level_debug", "dynamic_shapes"],
        "input_values": input_values,
    }


def _parameterize_class_name(cls: Type, idx: int, input_dicts: Mapping[Any, Any]):
    """Combine class name with the parameterized arguments.

    This function is passed to `parameterized.parameterized_class` as the
    `class_name_func` argument.
    """
    suffixes = []
    for k, v in input_dicts.items():
        suffixes.append(f"{k}_{v}")
    return f"{cls.__name__}_{'_'.join(suffixes)}"


@parameterized.parameterized_class(
    **_parameterized_class_attrs_and_values(),
    class_name_func=_parameterize_class_name,
)
class TestFxToOnnxWithOnnxRuntime(onnx_test_common._TestONNXRuntime):
    op_level_debug: bool
    dynamic_shapes: bool

    def setUp(self):
        super().setUp()
        self.opset_version = 18
        self.ort_version = onnxruntime.__version__

    @pytorch_test_common.skip_min_ort_version(
        reason="ORT doesn't support dynamic fx exporter yet making SegFault flaky test",
        version="1.15",
        dynamic_only=True,
    )
    def test_simple_function(self):
        def func(x):
            # TODO(justinchuby): Replicate torch's type casting policy
            # in the exporter for type promotion support
            y = x + 1.0
            z = y.relu()
            return (y, z)

        tensor_x = torch.randn(1, 1, 2, dtype=torch.float32)

        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(func, (tensor_x,))

    @pytorch_test_common.xfail(
        "AssertionError: Dynamo input/output is not consistent with traced input/output. "
        "Ref: https://github.com/pytorch/pytorch/issues/96379"
    )
    @pytorch_test_common.skip_min_ort_version(
        reason="ORT doesn't support dynamic fx exporter yet making SegFault flaky test",
        version="1.15",
        dynamic_only=True,
    )
    def test_func_with_args_and_tensor_kwargs(self):
        # Non-tensor optional kwargs are always folded into constant and
        # removed from input list in Dynamo-traced graph, if its value is not provided
        # to tracer. So for a function like
        #   def func(x, b=1.0)
        # here. E.g., if you first Dynamo-trace the model with arguments (x,),
        # and then call the traced graph with arguments (x, b=2.0), it will complain
        # somewhere that model is called with extra args because the modified
        # function is traced into
        #   def forward(self, x : torch.Tensor):
        #     add = x + 1.0;  x = None
        #     relu = add.relu()
        #     return (add, relu)
        # To summarize, in order to be traced as graph input, the value of optional kwarg
        # must be provided. Otherwise, they are treated as in-graph constants in Dynamo.
        # Tensor optional kwargs are an exception. It is always traced as input.
        # It is unclear if this behavior is intended or not. But in general it is bad
        # practice to set mutable default values.
        # `DynamoOptimizeExporter` applies a workaround by binding args and kwargs to
        # model signature and fill in the default values of unprovided optional arguments.
        def func(x, b=torch.tensor(1.0)):
            y = x + b
            z = y.relu()
            return (y, z)

        tensor_x = torch.randn(1, 2, 3, dtype=torch.float32)

        # Test without providing optional kwarg.
        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(func, (tensor_x,))
        # Test with only positional args.
        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            func, (tensor_x, torch.tensor(8.0))
        )
        # Test while specifying optional kwarg.
        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            func, (tensor_x,), {"b": torch.tensor(5.0)}
        )

    @pytorch_test_common.xfail(
        "https://github.com/pytorch/pytorch/issues/99534"
        "Non-tensor input is not traceable in dynamo."
    )
    @pytorch_test_common.skip_min_ort_version(
        reason="ORT doesn't support dynamic fx exporter yet making SegFault flaky test",
        version="1.15",
        dynamic_only=True,
    )
    def test_xfail_func_with_non_tensor_args(self):
        def func(x, b=1.0):
            y = x + b
            z = y.relu()
            return (y, z)

        tensor_x = torch.randn(1, 1, 2, dtype=torch.float32)

        export_output = torch.onnx.dynamo_export(
            func,
            tensor_x,
            8.0,
            export_options=torch.onnx.ExportOptions(
                opset_version=self.opset_version,
                op_level_debug=self.op_level_debug,
                dynamic_shapes=self.dynamic_shapes,
            ),
        )
        onnx_format_args = export_output.adapt_torch_inputs_to_onnx(tensor_x, 8.0)
        ref_outputs = export_output.adapt_torch_outputs_to_onnx(func(tensor_x, 8.0))
        ort_outputs = onnx_test_common.run_ort(export_output, onnx_format_args)
        for ref_output, ort_output in zip(ref_outputs, ort_outputs):
            torch.testing.assert_close(ref_output, torch.tensor(ort_output))

        # test on different non-tensor input - xfail
        onnx_format_args = export_output.adapt_torch_inputs_to_onnx(tensor_x, 9.0)
        ref_outputs = export_output.adapt_torch_outputs_to_onnx(func(tensor_x, 9.0))
        _ = onnx_test_common.run_ort(export_output, onnx_format_args)
        for ref_output, ort_output in zip(ref_outputs, ort_outputs):
            torch.testing.assert_close(ref_output, torch.tensor(ort_output))

    @pytorch_test_common.skip_min_ort_version(
        reason="ORT doesn't support dynamic fx exporter yet making SegFault flaky test",
        version="1.15",
        dynamic_only=True,
    )
    def test_func_with_nested_input_structure(self):
        def func(
            x_dict: Dict[str, torch.Tensor],
            y_tuple: Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]],
            z_list: List[List[torch.Tensor]],
        ):
            if "a" in x_dict:
                x = x_dict["a"]
            elif "b" in x_dict:
                x = x_dict["b"]
            else:
                x = torch.randn(3)

            y1, (y2, y3) = y_tuple

            z = x + y1 + y2 + y3
            for z_sub_list in z_list:
                z = z + torch.stack(z_sub_list).sum()

            return z

        x_dict = {"a": torch.randn(3), "c": torch.randn(3)}
        y_tuple = (torch.randn(3), (torch.randn(3), torch.randn(3)))
        z_list = [
            [torch.randn(3), torch.randn(3)],
            [torch.randn(3), torch.randn(3), torch.randn(3)],
        ]
        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            func, (x_dict, y_tuple, z_list)
        )

    @pytorch_test_common.skip_min_ort_version(
        reason="ORT doesn't support dynamic fx exporter yet making SegFault flaky test",
        version="1.15",
        dynamic_only=True,
    )
    def test_func_with_nested_output_structure(self):
        def func(x, y, z):
            x = x + y
            y = y + z
            z = x + y
            out1 = (x, (y, z))
            out2 = [[x, y], [y, z]]
            out3 = {"z": z, "x": x}
            return out1, out2, out3

        x = torch.randn(3)
        y = torch.randn(3)
        z = torch.randn(3)
        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(func, (x, y, z))

    @pytorch_test_common.skip_min_ort_version(
        reason="ORT doesn't support dynamic fx exporter yet making SegFault flaky test",
        version="1.15",
        dynamic_only=True,
    )
    def test_mnist(self):
        class MNISTModel(nn.Module):
            def __init__(self):
                super().__init__()
                self.conv1 = nn.Conv2d(1, 32, 3, 1, bias=True)
                self.conv2 = nn.Conv2d(32, 64, 3, 1, bias=True)
                self.fc1 = nn.Linear(9216, 128, bias=True)
                self.fc2 = nn.Linear(128, 10, bias=True)

            def forward(self, tensor_x: torch.Tensor):
                tensor_x = self.conv1(tensor_x)
                tensor_x = torch.sigmoid(tensor_x)
                tensor_x = self.conv2(tensor_x)
                tensor_x = torch.sigmoid(tensor_x)
                tensor_x = torch.max_pool2d(tensor_x, 2)
                tensor_x = torch.flatten(tensor_x, 1)
                tensor_x = self.fc1(tensor_x)
                tensor_x = torch.sigmoid(tensor_x)
                tensor_x = self.fc2(tensor_x)
                output = torch.log_softmax(tensor_x, dim=1)
                return output

        tensor_x = torch.rand((64, 1, 28, 28), dtype=torch.float32)
        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            MNISTModel(), (tensor_x,)
        )

    @pytorch_test_common.skip_min_ort_version(
        reason="ORT doesn't support dynamic fx exporter yet making SegFault flaky test",
        version="1.15",
        dynamic_only=True,
    )
    def test_log_sigmoid(self):
        # This produces op as `torch.ops.aten.log_sigmoid_forward`, instead of the more
        # conventional `torch.ops.aten.log_sigmoid`.
        class Model(torch.nn.Module):
            def __init__(self):
                super().__init__()
                self.m = torch.nn.LogSigmoid()

            def forward(self, x):
                return self.m(x)

        input = torch.randn(2)
        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(Model(), (input,))

    @pytorch_test_common.skip_min_ort_version(
        reason="ORT doesn't support dynamic fx exporter yet making SegFault flaky test",
        version="1.15",
        dynamic_only=True,
    )
    @skip_if_no_torchvision
    def test_resnet18(self):
        # TODO(bowbao): Note [training vs eval in dynamo_export]
        # So we are effectively exporting all models in traning mode by
        # default. But for the sake of this export we are only interested in eval mode.
        # The question is, should we call `model.eval()` in `dynamo_export`?
        # This particular test fails 'functionalization' in training mode.
        # So we are explicitly calling `model.eval()` for any model that contains
        # batch norm.
        # Ref: https://github.com/pytorch/pytorch/issues/99662#issuecomment-1528178221
        model = torchvision.models.resnet18(pretrained=False).eval()
        dummy_input = torch.randn(1, 3, 224, 224)

        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            model,
            (dummy_input,),
        )

    @pytorch_test_common.xfail(
        "RuntimeError: Unknown call_function target: aten.mean.dim"
    )
    @pytorch_test_common.skip_min_ort_version(
        reason="ORT doesn't support dynamic fx exporter yet making SegFault flaky test",
        version="1.15",
        dynamic_only=True,
    )
    @skip_if_no_torchvision
    def test_shufflenet_v2(self):
        # TODO(bowbao): see Note [training vs eval in dynamo_export]
        model = torchvision.models.shufflenet_v2_x0_5(pretrained=False).eval()
        dummy_input = torch.randn(1, 3, 224, 224, requires_grad=True)
        test_inputs = torch.randn(3, 3, 224, 224, requires_grad=True)

        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            model,
            (dummy_input,),
            additional_test_inputs=[((test_inputs,),)],
            rtol=1e-3,
            atol=1e-5,
        )

    @pytorch_test_common.skip_min_ort_version(
        reason="ORT doesn't support dynamic fx exporter yet making SegFault flaky test",
        version="1.15",
        dynamic_only=True,
    )
    def test_add(self):
        class DynamicAdd(torch.nn.Module):
            def forward(self, x, y):
                return torch.ops.aten.add(x, y)

        x = torch.randn(2, 3)
        y = torch.randn(2, 3)
        another_x = torch.randn(3, 4)
        another_y = torch.randn(3, 4)

        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            DynamicAdd(),
            (x, y),
            additional_test_inputs=[((another_x, another_y),)],
        )

    @pytorch_test_common.skip_min_ort_version(
        reason="ORT doesn't support dynamic fx exporter yet making SegFault flaky test",
        version="1.15",
        dynamic_only=True,
    )
    def test_sigmoid_add(self):
        class DynamicAdd(torch.nn.Module):
            def __init__(self, *args, **kwargs) -> None:
                super().__init__(*args, **kwargs)
                self.sigmoid = torch.nn.Sigmoid()

            def forward(self, x, y):
                z = torch.ops.aten.add(x, y)
                return self.sigmoid(z)

        x = torch.randn(2, 3)
        y = torch.randn(2, 3)
        x = x[1:, :]
        y = y[1:, :]
        input_x = torch.randn(1, 4)
        input_y = torch.randn(1, 4)

        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            DynamicAdd(), (x, y), additional_test_inputs=[((input_x, input_y),)]
        )

    @pytorch_test_common.skip_min_ort_version(
        reason="ORT doesn't support dynamic fx exporter yet making SegFault flaky test",
        version="1.15",
        dynamic_only=True,
    )
    def test_matmul(self):
        class DynamicMatMul(torch.nn.Module):
            def forward(self, x, y):
                return torch.ops.aten.matmul(x, y)

        x = torch.randn(2, 3, 6)
        y = torch.randn(2, 6, 4)
        input_x = torch.randn(2, 3, 4)
        input_y = torch.randn(2, 4, 4)

        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            DynamicMatMul(), (x, y), additional_test_inputs=[((input_x, input_y),)]
        )

    @pytorch_test_common.skip_dynamic_fx_test(
        "fx graph does not capture symbolic value for aten::scalar_tensor."
    )
    def test_scalar_tensor(self):
        class test(torch.nn.Module):
            def forward(self, x):
                return torch.scalar_tensor(x.size(0)), torch.scalar_tensor(
                    x.size(1), dtype=torch.int64
                )

        x = torch.randn(2, 3, 4)
        y = torch.randn(7, 8, 9)
        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            test(),
            (x,),
            additional_test_inputs=[((y,),)],
        )

    def test_transpose_infer_shape(self):
        class TransposeModule(torch.nn.Module):
            def __init__(self):
                super().__init__()
                self.conv = torch.nn.Conv2d(3, 1, 3, stride=2)

            def forward(self, x):
                x = self.conv(x)
                return x.transpose(0, 1)

        x = torch.randn(32, 3, 64, 64)
        y = torch.randn(16, 3, 8, 64)
        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            TransposeModule(),
            (x,),
            additional_test_inputs=[((y,),)],
        )

    @pytorch_test_common.xfail("torch._dynamo.exc.TorchRuntimeError")
    def test_squeeze_runtime_dim(self):
        class Squeeze(torch.nn.Module):
            def forward(self, d1, d2):
                t = torch.zeros(d1[0], d2[0])
                return t.squeeze(0)

        d1 = torch.tensor([1])
        d3 = torch.tensor([3])
        d4 = torch.tensor([4])
        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            Squeeze(), (d1, d4), additional_test_inputs=[((d3, d4),)]
        )
        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            Squeeze(), (d3, d4), additional_test_inputs=[((d1, d3),)]
        )

    @pytorch_test_common.skip_min_ort_version(
        reason="ORT doesn't support dynamic fx exporter yet making SegFault flaky test",
        version="1.15",
        dynamic_only=True,
    )
    def test_slice(self):
        class DynamicSliceExportMod(torch.nn.Module):
            def forward(self, x):
                results = []
                for i in range(4):
                    results.append(x[: x.size(0) - i, i : x.size(2), i:3])
                return tuple(results)

        x = torch.rand(5, 5, 5)
        y = torch.randn(6, 7, 8)
        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            DynamicSliceExportMod(),
            (x,),
            additional_test_inputs=[((y,),)],
        )

    # TODO(titaiwang): This is also detected flaky in static shape:
    # https://github.com/pytorch/pytorch/issues/98622
    @pytorch_test_common.skip_min_ort_version(
        reason="ORT doesn't support dynamic fx exporter yet making SegFault flaky test",
        version="1.15",
        dynamic_only=False,
    )
    def test_mutation(self):
        class MutationModel(torch.nn.Module):
            def forward(self, x):
                x.view(3, 2, -1).add_(2.0)
                return x

        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            MutationModel(), (torch.randn(12),), has_mutation=True
        )

    @pytorch_test_common.skip_min_ort_version(
        reason="ORT doesn't support dynamic fx exporter yet making SegFault flaky test",
        version="1.15",
        dynamic_only=False,
    )
    def test_arange(self):
        class ArangeModel(torch.nn.Module):
            def forward(self, input):
                return (
                    torch.arange(input.shape[0]),
                    torch.arange(12),
                    torch.arange(start=input.shape[0], end=input.shape[0] + 5),
                )

        x = torch.randn(5, 3, 2)
        y = torch.randn(8, 3, 2)
        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            ArangeModel(),
            (x,),
            additional_test_inputs=[((y,),)],
        )

    @pytorch_test_common.xfail(
        "fx.graph: torch._subclasses.fake_tensor.DataDependentOutputException: "
        "aten._local_scalar_dense.default"
    )
    def test_expand_as_fill_zero(self):
        class Model(torch.nn.Module):
            def forward(self, x):
                x[:, x.size(0) :] = 0
                return x

        x = torch.ones(2, 5)
        x2 = torch.randn(3, 4)
        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            Model(),
            (x,),
            additional_test_inputs=[((x2,),)],
        )

    @pytorch_test_common.xfail(
        "RuntimeError: Unknown call_function target: aten.lift_fresh_copy.default"
    )
    def test_expand_as_fill_tensor(self):
        class Model(torch.nn.Module):
            def forward(self, x):
                x[:, x.size(0) :] = torch.tensor([1, 2, 3])
                return x

        x = torch.ones(2, 5, 3)
        x2 = torch.randn(3, 4, 3)
        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            Model(),
            (x,),
            additional_test_inputs=[((x2,),)],
        )

    @pytorch_test_common.xfail(
        "Unknown call_function target: aten.lift_fresh_copy.default"
    )
    def test_expand_as_fill_seperate_tensor(self):
        class Model(torch.nn.Module):
            def forward(self, x):
                aa = torch.tensor([[0], [1], [2]])
                return aa.expand_as(x)

        x = torch.ones(3, 2)
        x2 = torch.randn(3, 5)
        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            Model(),
            (x,),
            additional_test_inputs=[((x2,),)],
        )

    @pytorch_test_common.skip_min_ort_version(
        reason="ORT doesn't support dynamic fx exporter yet making SegFault flaky test",
        version="1.15",
        dynamic_only=True,
    )
    def test_view_dynamic_zero_dim(self):
        class ViewModel(torch.nn.Module):
            def forward(self, input):
                input = input.view(-1, 2)
                return input.view(1, -1)

        x = torch.ones(2)
        another_x = torch.empty((0,))
        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            ViewModel(),
            (x,),
            additional_test_inputs=[((another_x,),)],
        )

    @pytorch_test_common.skip_min_ort_version(
        reason="ORT doesn't support dynamic fx exporter yet making SegFault flaky test",
        version="1.15",
        dynamic_only=True,
    )
    def test_flatten_dynamic_axes(self):
        class MyModule(torch.nn.Module):
            def forward(self, x):
                return torch.flatten(x, start_dim=2, end_dim=3)

        batch_size = 3
        x = torch.randn(batch_size, 5, 4, 5)
        y = torch.randn(5, 5, 4, 5)
        model = MyModule()
        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            model, (x,), additional_test_inputs=[((y,),)]
        )

    @pytorch_test_common.skip_min_ort_version(
        reason="ORT doesn't support dynamic fx exporter yet making SegFault flaky test",
        version="1.15",
        dynamic_only=True,
    )
    def test_none_input(self):
        class NoneInputModel(torch.nn.Module):
            def forward(
                self, x: torch.Tensor, y: Optional[torch.Tensor], z: torch.Tensor
            ):
                if y is None:
                    return x + z
                return x + y + z

        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            NoneInputModel(), (torch.randn(1, 2), None, torch.randn(1, 2))
        )

    @pytorch_test_common.skip_min_ort_version(
        reason="ORT doesn't support dynamic fx exporter yet making SegFault flaky test",
        version="1.15",
        dynamic_only=True,
    )
    def test_gpt2_tiny(self):
        model_name = "sshleifer/tiny-gpt2"
        # Download pytorch model
        model = transformers.AutoModel.from_pretrained(model_name)
        tokenizer = transformers.AutoTokenizer.from_pretrained(model_name)

        # Transform input tokens
        inputs = tokenizer("Hello world!", return_tensors="pt")
        another_inputs = tokenizer("Another Hello world!", return_tensors="pt")

        self.run_test_with_fx_to_onnx_exporter_and_onnx_runtime(
            model, [], inputs, additional_test_inputs=[((), another_inputs)]
        )

    @_beartype.beartype
    def _test_fx_symbolic_tracer_large_scale_exporter(
        self,
        model_name: str,
        create_model: Callable,
        create_args: Callable,
        create_pytorch_only_kwargs: Callable,
    ):
        """Test helper for large-scale exporter.

        Arguments:
            model_name: Name of the model. It used to name temporary files.
            create_model: A function that creates a model. It should always create the same model.
            create_args: A function that creates random input arguments for the model.
            create_pytorch_only_kwargs: A function that creates kwargs for calling PyTorch model with real tensors.

        This test contains several steps.

        1. Create a toy model.
        2. Save the toy's state (parameters) to a file. This is for simulating a checkpoint file.
        3. Load it back and export it to ONNX with large-scale exporter.
            All operations (including model loading) are done under
            FakeTensorMode so no real tensor is created and no real
            computation happens.
        4. The ONNX model generated in step 3 doesn't contain parameters,
            and this step adds them as external data and save a new ONNX model.
        5. Run PyTorch and ONNX models and compare their results.
        """

        # Create the toy model.
        model = create_model()

        with tempfile.NamedTemporaryFile(
            prefix=model_name, suffix=".pt"
        ) as tmp_file, tempfile.TemporaryDirectory(
            suffix="large_scale_export"
        ) as tmp_folder:
            # Dump state_dict to a file to simulate how HuggingFace model is initialized.
            # The file will be loaded via .load_state_dict(...)
            torch.save(model.state_dict(), tmp_file.name)

            ftm = fake_tensor.FakeTensorMode(
                allow_non_fake_inputs=True, allow_fallback_kernels=False
            )
            ctx = patcher.ONNXTorchPatcher()
            # NOTE: FakeTensorMode disallows symbolic shape of fx graph
            # The following coed block does several things.
            #  1. Create a model whose parameters and buffers are all FakeTensor's.
            #  2. Convert nn.Module into ONNX model without initializers.
            #  3. Record the file paths to find real initializers.
            with ctx, ftm:
                # Toy model with parameters and buffers as FakeTensor's.
                fake_model = create_model()
                fake_model.load_state_dict(torch.load(tmp_file.name))
                # Toy inputs as FakeTensor's.
                fake_args = create_args()
                # Export ONNX model without initializers while ctx.paths records
                # all files that contains real initializers.

                options = torch.onnx.ExportOptions(
                    opset_version=self.opset_version,
                    dynamic_shapes=self.dynamic_shapes,
                    op_level_debug=self.op_level_debug,
                )
                export_options = torch.onnx._internal.exporter.ResolvedExportOptions(
                    options
                )
                export_options.fx_tracer = (
                    fx_symbolic_graph_extractor.FXSymbolicTracer()
                )
                export_output = torch.onnx.dynamo_export(
                    fake_model,
                    *fake_args,
                    export_options=export_options,
                )

                onnx_model = export_output.model_proto

            # Tasks done by the following block.
            #  1. Iterate through all tensors stored in ctx.paths (the file content is loaded torch.load)
            #  2. If a tensor's name matches a "onnx_model"'s input name, an initializer is created and saved to
            #     a seperated folder.
            #  3. A new ONNX model is saved into file with the initializers saved in the previous step.
            #  4. ORT executes the new ONNX model and compares the results with the original GPT model.

            # Model saved to tmp_folder/onnx_model_location
            # Initializers are saved to tmp_folder/onnx_initializer_location/*.onnx
            onnx_model_location = model_name + "_external_data.onnx"
            onnx_initializer_location = model_name + "_initializers"
            fx_serialization.save_model_with_external_data(
                tmp_folder,
                onnx_model_location,
                onnx_initializer_location,
                tuple(ctx.paths),
                onnx_model,
            )
            # Generate random inputs.
            args = create_args()
            kwargs = create_pytorch_only_kwargs()
            # Original outputs.
            ref_outputs = export_output.adapt_torch_outputs_to_onnx(
                model(*args, **kwargs)
            )
            # ORT outputs.
            args_not_none = export_output.adapt_torch_inputs_to_onnx(*args)

            # Drop Parameters and buffers added by fx_serialization.save_model_with_external_data
            args_not_none = args_not_none[: len(args) - len(kwargs)]

            ort_outputs = onnx_test_common.run_ort(
                os.path.join(tmp_folder, onnx_model_location),
                args_not_none,
            )

            assert len(ref_outputs) == len(ort_outputs)

            for ref_output, ort_output in zip(ref_outputs, ort_outputs):
                torch.testing.assert_close(ref_output, torch.tensor(ort_output))

    @pytorch_test_common.skip_dynamic_fx_test(
        "FakeTensor exporting is not supported by dynamic axes."
    )
    def test_fx_symbolic_tracer_large_scale_exporter_with_toy_mlp(self):
        class MLPModel(nn.Module):
            def __init__(self):
                super().__init__()
                self.fc0 = nn.Linear(8, 8, bias=True)
                self.fc1 = nn.Linear(8, 4, bias=True)
                self.fc2 = nn.Linear(4, 2, bias=True)
                self.fc3 = nn.Linear(2, 2, bias=True)

            def forward(self, tensor_x: torch.Tensor):
                tensor_x = self.fc0(tensor_x)
                tensor_x = torch.sigmoid(tensor_x)
                tensor_x = self.fc1(tensor_x)
                tensor_x = torch.sigmoid(tensor_x)
                tensor_x = self.fc2(tensor_x)
                tensor_x = torch.sigmoid(tensor_x)
                output = self.fc3(tensor_x)
                return output

        def create_model() -> nn.Module:
            return MLPModel()

        def create_args():
            return (torch.rand((97, 8), dtype=torch.float32),)

        def create_pytorch_only_extra_kwargs():
            return {}

        self._test_fx_symbolic_tracer_large_scale_exporter(
            "toy_mlp1",
            create_model,
            create_args,
            create_pytorch_only_extra_kwargs,
        )

    @pytorch_test_common.skip_dynamic_fx_test(
        "FakeTensor exporting is not supported by dynamic axes."
    )
    def test_large_scale_exporter_with_tiny_gpt2(self):
        model_name = "sshleifer/tiny-gpt2"

        def create_model() -> nn.Module:
            return transformers.AutoModel.from_pretrained(model_name)

        def create_args():
            tokenizer = transformers.AutoTokenizer.from_pretrained(model_name)
            kwargs = tokenizer("Hello world!", return_tensors="pt")
            input_ids = kwargs["input_ids"]
            attention_mask = kwargs["attention_mask"]
            return input_ids, None, attention_mask

        def create_pytorch_only_extra_kwargs():
            return {"return_dict": False}

        self._test_fx_symbolic_tracer_large_scale_exporter(
            "tiny_gpt2",
            create_model,
            create_args,
            create_pytorch_only_extra_kwargs,
        )


if __name__ == "__main__":
    common_utils.run_tests()