test_pytorch_onnx_no_runtime.py

# Owner(s): ["module: onnx"]

"""Tests for onnx export that don't run the exported model."""

from __future__ import annotations

import contextlib
import io
import itertools
import unittest
import unittest.mock
import warnings
from typing import Callable, Dict, Iterable, List, Optional, Tuple, Union

import numpy as np

import onnx
import onnx.numpy_helper
import pytorch_test_common

import torch
import torch.nn.functional as F
from torch import Tensor
from torch.onnx import symbolic_helper, utils
from torch.onnx._internal import registration
from torch.testing._internal import common_quantization, common_utils, jit_utils


def export_to_onnx(
    model: Union[torch.nn.Module, torch.jit.ScriptFunction],
    input: Union[torch.Tensor, Tuple[torch.Tensor]],
    custom_ops: Optional[
        Iterable[Union[contextlib.AbstractContextManager, contextlib.ContextDecorator]]
    ] = None,
    mocks: Optional[Iterable] = None,
    operator_export_type: torch.onnx.OperatorExportTypes = torch.onnx.OperatorExportTypes.ONNX,
    opset_version: int = 17,
    **torch_onnx_export_kwargs,
) -> onnx.ModelProto:
    """Exports `model(input)` to ONNX and returns it.

    Custom operators and/or unittest patches can be used help reproducing specific behaviors.

    Args:
        model: model to export
        input: model input with same format as `torch.onnx.export(..,args,...)`
        custom_ops: list of custom operators to use during export
        mocks: list of mocks to use during export
        operator_export_type: export type as described by `torch.onnx.export(...operator_export_type,...)`
        opset_version: ONNX opset version as described by `torch.onnx.export(...opset_version,...)`
        torch_onnx_export_kwargs: extra torch.onnx.export kwargs arguments
    Returns:
        A valid ONNX model (`onnx.ModelProto`)
    """
    custom_ops = custom_ops or []
    mocks = mocks or []
    with contextlib.ExitStack() as stack:
        for ctx in itertools.chain(custom_ops, mocks):
            stack.enter_context(ctx)

        f = io.BytesIO()
        torch.onnx.export(
            model,
            input,
            f,
            operator_export_type=operator_export_type,
            opset_version=opset_version,
            **torch_onnx_export_kwargs,
        )

    # Validate ONNX graph before returning it
    onnx_model = onnx.load_from_string(f.getvalue())
    onnx.checker.check_model(onnx_model)
    return onnx_model


@common_utils.instantiate_parametrized_tests
class TestONNXExport(pytorch_test_common.ExportTestCase):
    def test_fuse_addmm(self):
        class AddmmModel(torch.nn.Module):
            def forward(self, x):
                return torch.mm(x, x) + x

        x = torch.ones(3, 3)
        f = io.BytesIO()
        torch.onnx.export(AddmmModel(), x, f)

    def test_onnx_transpose_incomplete_tensor_type(self):
        # Smoke test to get us into the state where we are attempting to export
        # a transpose op, where the input is a TensorType without size information.
        # This would previously not work, since we would
        # take the size of the input and use the length of its sizes as the
        # number of dimensions in the permutation.
        class Foo(torch.jit.ScriptModule):
            @torch.jit.script_method
            def forward(self, x):
                return x.contiguous().transpose(0, 1).sum()

        class TraceMe(torch.nn.Module):
            def __init__(self) -> None:
                super().__init__()
                self.foo = Foo()

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

        tm = TraceMe()
        tm = torch.jit.trace(tm, torch.rand(3, 4))
        f = io.BytesIO()
        torch.onnx.export(tm, (torch.rand(3, 4),), f)

    def test_export_tensoroption_to(self):
        def foo(x):
            return x[0].detach().clone().cpu() + x

        traced = torch.jit.trace(foo, (torch.rand([2])))

        f = io.BytesIO()
        torch.onnx.export(traced, (torch.rand([2]),), f)

    def test_onnx_export_script_module(self):
        class ModuleToExport(torch.jit.ScriptModule):
            @torch.jit.script_method
            def forward(self, x):
                y = x - x
                return x + x

        mte = ModuleToExport()
        f = io.BytesIO()
        torch.onnx.export(mte, (torch.zeros(1, 2, 3),), f)

    @common_utils.suppress_warnings
    def test_onnx_export_func_with_warnings(self):
        @torch.jit.script
        def func_with_warning(inp):
            return torch.nn.functional.sigmoid(inp)  # triggers a deprecation warning

        class WarningTest(torch.nn.Module):
            def forward(self, x):
                return func_with_warning(x)

        # no exception
        f = io.BytesIO()
        torch.onnx.export(WarningTest(), torch.randn(42), f)

    def test_onnx_export_script_python_fail(self):
        class PythonModule(torch.jit.ScriptModule):
            @torch.jit.ignore
            def forward(self, x):
                return torch.neg(x)

        class ModuleToExport(torch.jit.ScriptModule):
            def __init__(self) -> None:
                super().__init__()
                self.mod = PythonModule()

            @torch.jit.script_method
            def forward(self, x):
                y = self.mod(x)
                return y + y

        mte = ModuleToExport()
        f = io.BytesIO()
        with self.assertRaisesRegex(RuntimeError, "Couldn't export Python"):
            torch.onnx.export(mte, (torch.zeros(1, 2, 3),), f)

    def test_onnx_export_script_inline_trace(self):
        class ModuleToInline(torch.nn.Module):
            def forward(self, x):
                return torch.neg(x)

        class ModuleToExport(torch.jit.ScriptModule):
            def __init__(self) -> None:
                super().__init__()
                self.mod = torch.jit.trace(ModuleToInline(), torch.zeros(1, 2, 3))

            @torch.jit.script_method
            def forward(self, x):
                y = self.mod(x)
                return y + y

        mte = ModuleToExport()
        f = io.BytesIO()
        torch.onnx.export(mte, (torch.zeros(1, 2, 3),), f)

    def test_onnx_export_script_inline_script(self):
        class ModuleToInline(torch.jit.ScriptModule):
            @torch.jit.script_method
            def forward(self, x):
                return torch.neg(x)

        class ModuleToExport(torch.jit.ScriptModule):
            def __init__(self) -> None:
                super().__init__()
                self.mod = ModuleToInline()

            @torch.jit.script_method
            def forward(self, x):
                y = self.mod(x)
                return y + y

        mte = ModuleToExport()
        f = io.BytesIO()
        torch.onnx.export(mte, (torch.zeros(1, 2, 3),), f)

    def test_onnx_export_script_module_loop(self):
        class ModuleToExport(torch.jit.ScriptModule):
            @torch.jit.script_method
            def forward(self, x):
                # test if we support end to end onnx export on loop and
                # nested loops with and without loop index
                for _ in range(5):
                    for i in range(3):
                        x = x + i
                return x

        mte = ModuleToExport()
        f = io.BytesIO()
        torch.onnx.export(mte, (torch.zeros(1, 2, 3),), f)

    @common_utils.suppress_warnings
    def test_onnx_export_script_truediv(self):
        class ModuleToExport(torch.jit.ScriptModule):
            @torch.jit.script_method
            def forward(self, x):
                z = x.size(0) / 2
                return x + z

        mte = ModuleToExport()

        f = io.BytesIO()
        torch.onnx.export(mte, (torch.zeros(1, 2, 3, dtype=torch.float),), f)

    def test_onnx_export_script_non_alpha_add_sub(self):
        class ModuleToExport(torch.jit.ScriptModule):
            @torch.jit.script_method
            def forward(self, x):
                bs = x.size(0) + 1
                return bs - 1

        mte = ModuleToExport()
        f = io.BytesIO()
        torch.onnx.export(mte, (torch.rand(3, 4),), f)

    def test_onnx_export_script_module_if(self):
        class ModuleToExport(torch.jit.ScriptModule):
            @torch.jit.script_method
            def forward(self, x):
                if bool(torch.sum(x) > 0):
                    x = torch.neg(x)
                return x

        mte = ModuleToExport()
        f = io.BytesIO()
        torch.onnx.export(mte, (torch.zeros(1, 2, 3),), f)

    def test_onnx_export_script_inline_params(self):
        class ModuleToInline(torch.jit.ScriptModule):
            def __init__(self) -> None:
                super().__init__()
                self.m = torch.nn.Parameter(torch.ones(3, 3))
                self.unused = torch.nn.Parameter(torch.ones(1, 2, 3))

            @torch.jit.script_method
            def forward(self, x):
                return torch.mm(x, self.m)

        class ModuleToExport(torch.jit.ScriptModule):
            def __init__(self) -> None:
                super().__init__()
                self.mod = ModuleToInline()
                self.param = torch.nn.Parameter(torch.ones(3, 4))

            @torch.jit.script_method
            def forward(self, x):
                y = self.mod(x)
                return torch.mm(y, self.param)

        mte = ModuleToExport()
        result = mte(torch.zeros(2, 3))
        reference = torch.mm(
            torch.mm(torch.zeros(2, 3), torch.ones(3, 3)), torch.ones(3, 4)
        )
        self.assertEqual(result, reference)
        f = io.BytesIO()
        torch.onnx.export(mte, (torch.ones(2, 3),), f)

    def test_onnx_export_speculate(self):
        class Foo(torch.jit.ScriptModule):
            def __init__(self, m):
                super().__init__()
                self.m = m

            @torch.jit.script_method
            def forward(self, x):
                x += x
                # because we are testing if we emit `if` statement correctly
                # we cannot use `True` as the condition. Constant prop
                # would remove the `if` statements.
                c = torch.sum(x) > 4
                if bool(c):
                    if bool(c):
                        y = self.m(x)
                    else:
                        y = self.m(x)
                else:
                    y = self.m(x)
                return y

        linear = torch.jit.trace(
            torch.nn.Linear(10, 20).float(), torch.zeros(1, 10, dtype=torch.float)
        )

        @torch.jit.script
        def transpose(x):
            return x.t()

        f1 = Foo(transpose)
        f2 = Foo(linear)

        f = io.BytesIO()
        torch.onnx.export(f1, (torch.ones(1, 10, dtype=torch.float),), f)
        f = io.BytesIO()
        torch.onnx.export(f2, (torch.ones(1, 10, dtype=torch.float),), f)

    def test_onnx_export_shape_reshape(self):
        class Foo(torch.nn.Module):
            def forward(self, x):
                import torch.onnx.operators

                x = x.repeat(5, 1, 1)
                shape = torch.onnx.operators.shape_as_tensor(x)
                reshaped = torch.onnx.operators.reshape_from_tensor_shape(x, shape)
                return reshaped

        foo = torch.jit.trace(Foo(), torch.zeros(1, 2, 3))
        f = io.BytesIO()
        torch.onnx.export(foo, (torch.zeros(1, 2, 3)), f)

    def test_listconstruct_erasure(self):
        class FooMod(torch.nn.Module):
            def forward(self, x):
                mask = x < 0.0
                return x[mask]

        f = io.BytesIO()
        torch.onnx.export(
            FooMod(),
            (torch.rand(3, 4),),
            f,
            add_node_names=False,
            do_constant_folding=False,
            operator_export_type=torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK,
        )

    def test_export_dynamic_slice(self):
        class DynamicSliceExportMod(torch.jit.ScriptModule):
            @torch.jit.script_method
            def forward(self, x):
                retval = x[0]
                for i in range(x.size(1)):
                    retval += torch.sum(x[0:i], dim=0)
                return retval

        input = torch.rand(3, 4, 5)

        f = io.BytesIO()
        torch.onnx.export(DynamicSliceExportMod(), (input,), f, opset_version=10)

    def test_export_dict(self):
        class DictModule(torch.nn.Module):
            def forward(self, x_in: torch.Tensor) -> Dict[str, torch.Tensor]:
                return {"test_key_out": x_in}

        x_in = torch.tensor(1)
        mod = DictModule()
        mod.train(False)

        f = io.BytesIO()
        torch.onnx.export(mod, (x_in,), f)

        with self.assertRaisesRegex(RuntimeError, r"DictConstruct.+is not supported."):
            f = io.BytesIO()
            torch.onnx.export(torch.jit.script(mod), (x_in,), f)

    def test_source_range_propagation(self):
        class ExpandingModule(torch.nn.Module):
            def __init__(self) -> None:
                super().__init__()
                # Will be expanded during ONNX export
                self.ln = torch.nn.LayerNorm([1])

            def forward(self, input):
                return self.ln(input)

        mod = ExpandingModule()

        graph, _, _ = utils._model_to_graph(
            mod,
            (torch.zeros(1),),
            operator_export_type=torch.onnx.OperatorExportTypes.ONNX,
        )

        # Ensure that every node in the graph has a valid source range
        for node in graph.nodes():
            self.assertTrue(node.sourceRange())

    def test_clip_aten_fallback_due_exception(self):
        def bad_clamp(g, self, min, max):
            return symbolic_helper._onnx_unsupported("Bad boy!")

        class MyClip(torch.nn.Module):
            def forward(self, x):
                return torch.clamp(x, min=-0.5, max=0.5)

        onnx_model = export_to_onnx(
            MyClip(),
            torch.randn(3, 4, requires_grad=True),
            custom_ops=[common_utils.custom_op("aten::clamp", bad_clamp, 17)],
            operator_export_type=torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK,
        )
        self.assertAtenOp(onnx_model, "clamp", "Tensor")

    def test_clip_aten_fallback_explicit_request(self):
        class MyClip(torch.nn.Module):
            def forward(self, x):
                return torch.clamp(x, min=-0.5, max=0.5)

        # Copy of mocked method must be saved to prevent
        # max recursion depth while trying to run original instance method
        original_get_function_group = registration.registry.get_function_group

        def break_is_registered_op_api(name):
            fake_missing_symbolics = {"aten::clamp"}
            if name in fake_missing_symbolics:
                return None
            return original_get_function_group(name)

        # Force missing symbolic for well-known op using a mock
        onnx_model = export_to_onnx(
            MyClip(),
            torch.randn(3, 4, requires_grad=True),
            mocks=[
                unittest.mock.patch(
                    "torch.onnx._internal.registration.registry.get_function_group",
                    side_effect=break_is_registered_op_api,
                    # wraps=registration.registry.get_function_group
                )
            ],
            operator_export_type=torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK,
        )
        self.assertAtenOp(onnx_model, "clamp", "Tensor")

    def _helper_test_to_(self, cast_fn: Callable[[torch.Tensor], torch.Tensor]):
        """Helper to test aten::to(device) variants.

        `cast_fn` is converted into a `torch.jit.script`. It wraps `aten::to`
        during export to preventing the devices to be hard-coded.

        Needed by detectron2 after https://github.com/facebookresearch/detectron2/pull/4132/
        """
        cast_fn = torch.jit.script(cast_fn)
        onnx_model = export_to_onnx(cast_fn, torch.zeros([1, 3, 32, 32]))
        for n in onnx_model.graph.node:
            self.assertNotEqual(n.op_type, "To")
            self.assertNotEqual(n.op_type, "Cast")

    def test_to__cpu_string(self):
        def cast_cpu_string(src: torch.Tensor) -> torch.Tensor:
            return src.to("cpu")

        self._helper_test_to_(cast_cpu_string)

    def test_to__device_cpu_string(self):
        def cast_device_cpu_string(src: torch.Tensor) -> torch.Tensor:
            return src.to(device="cpu")

        self._helper_test_to_(cast_device_cpu_string)

    def test_script_custom_class_error(self):
        class BoxCoder:
            def __init__(self, bbox_xform_clip: float) -> None:
                self.bbox_xform_clip = bbox_xform_clip

            def decode(self, rel_codes: Tensor, boxes: List[Tensor]) -> Tensor:
                boxes = torch.cat(boxes, dim=0)
                pred_ctr_x = (
                    torch.clamp(rel_codes[:, 0::4], max=self.bbox_xform_clip)
                    * boxes[:, 2]
                )
                return pred_ctr_x

        class MyModule(torch.nn.Module):
            __annotations__ = {
                "box_coder": BoxCoder,
            }

            def __init__(self) -> None:
                super().__init__()
                self.box_coder = BoxCoder(1.4)

            def forward(self, box_regression: Tensor, proposals: List[Tensor]):
                return self.box_coder.decode(box_regression, proposals)

        model = torch.jit.script(MyModule())
        box_regression = torch.randn([4, 4])
        proposal = [torch.randn(2, 4), torch.randn(2, 4)]

        with self.assertRaises(RuntimeError) as cm:
            f = io.BytesIO()
            torch.onnx.export(
                model,
                (box_regression, proposal),
                f,
            )

    def test_initializer_sequence(self):
        class MyModule(torch.nn.Module):
            def __init__(self, input_size, hidden_size, num_classes):
                super().__init__()
                self.fc1 = torch.nn.Linear(input_size, hidden_size)
                self.relu = torch.nn.ReLU()
                self.fc2 = torch.nn.Linear(hidden_size, num_classes)

            def forward(self, x):
                out = self.fc1(x)
                out = self.relu(out)
                out = self.fc2(out)
                return out

        test_model = MyModule(3, 4, 10)
        state_dict_list = [k for (k, v) in test_model.state_dict().items()]
        named_params_list = [k for (k, v) in test_model.named_parameters()]

        x = torch.randn(32, 3)
        f = io.BytesIO()
        torch.onnx.export(test_model, (x,), f, do_constant_folding=False)
        loaded_model = onnx.load_from_string(f.getvalue())

        actual_list = [p.name for p in loaded_model.graph.initializer]
        assert actual_list == state_dict_list, (
            "Initializers' sequence is not as same as state_dict(). Expected: ("
            + ", ".join(state_dict_list)
            + "). Actual:("
            + ", ".join(actual_list)
            + ")."
        )
        assert actual_list == named_params_list, (
            "Initializers' sequence is not as same as named_parameters(). Expected: ("
            + ", ".join(named_params_list)
            + "). Actual:("
            + ", ".join(actual_list)
            + ")."
        )

    def test_initializer_sequence_script_model(self):
        def list_is_expected(short_list, long_list) -> bool:
            if len(short_list) > len(long_list):
                return False

            for i in range(len(short_list)):
                if short_list[i] not in long_list[i]:
                    return False

            return True

        def loop(x, y):
            for i in range(int(y)):
                x = x + i
            return x

        class MyModule(torch.nn.Module):
            def __init__(self, input_size, hidden_size, num_classes):
                super().__init__()
                self.fc1 = torch.nn.Linear(input_size, hidden_size)
                self.relu = torch.nn.ReLU()
                self.fc2 = torch.nn.Linear(hidden_size, num_classes)

            def forward(self, x, y):
                x = loop(x, y)
                out = self.fc1(x)
                out = self.relu(out)
                out = self.fc2(out)
                return out

        test_model = torch.jit.script(MyModule(3, 4, 10))
        state_dict_list = [k for (k, v) in test_model.state_dict().items()]
        named_params_list = [k for (k, v) in test_model.named_parameters()]

        x = torch.ones(2, 3, dtype=torch.float)
        y = torch.tensor(5, dtype=torch.long)
        f = io.BytesIO()

        torch.onnx.export(test_model, (x, y), f, do_constant_folding=False)
        loaded_model = onnx.load_from_string(f.getvalue())

        actual_list = [p.name for p in loaded_model.graph.initializer]
        assert list_is_expected(state_dict_list, actual_list), (
            "ScriptModel - Initializers' sequence is not as same as state_dict(). Expected: ("
            + ", ".join(state_dict_list)
            + "). Actual:("
            + ", ".join(actual_list)
            + ")."
        )
        assert list_is_expected(named_params_list, actual_list), (
            "ScriptModel - Initializers' sequence is not as same as named_parameters(). Expected: ("
            + ", ".join(named_params_list)
            + "). Actual:("
            + ", ".join(actual_list)
            + ")."
        )

    def test_shape_value_map(self):
        class RSoftMax(torch.nn.Module):
            def __init__(self, radix, cardinality):
                super().__init__()
                self.radix = radix
                self.cardinality = cardinality

            def forward(self, x):
                batch = x.size(0)
                x = x.view(batch, self.cardinality, self.radix, -1).transpose(1, 2)
                x = F.softmax(x, dim=1)
                x = x.reshape(batch, -1)
                return x

        radix = 2
        cardinality = 1
        x = torch.randn(10, 1, 128, 1)
        f = io.BytesIO()
        torch.onnx.export(
            RSoftMax(radix, cardinality),
            (x,),
            f,
            input_names=["x"],
            dynamic_axes={"x": [0]},
        )
        loaded_model = onnx.load_from_string(f.getvalue())
        self.assertEqual(
            loaded_model.graph.output[0].type.tensor_type.shape.dim[1].dim_value, 128
        )

    def test_onnx_proto_checker(self):
        class Model(torch.nn.Module):
            def forward(self, x):
                return 2 * x

        x = torch.randn(1, 2, 3, requires_grad=True)
        f = io.BytesIO()
        torch.onnx.export(Model(), (x,), f)
        model = onnx.load(f)
        model.ir_version = 0

        def check_proto():
            torch._C._check_onnx_proto(model.SerializeToString())

        self.assertRaises(RuntimeError, check_proto)

    def test_maintain_dynamic_shapes_of_unreliable_nodes(self):
        def symbolic_pythonop(g, *args, **kwargs):
            return g.op("com.microsoft::PythonOp")

        torch.onnx.register_custom_op_symbolic("prim::PythonOp", symbolic_pythonop, 1)
        self.addCleanup(torch.onnx.unregister_custom_op_symbolic, "prim::PythonOp", 1)

        # necessay parameters for transformer embeddings
        hidden_size = 48
        max_position_embeddings = 32
        batch_size = 2

        # issue found that autograd.function making downstream
        # node unreliable but with static shape. The issue was first
        # discovered with using Apex FusedLayerNorm in Transformers
        class CustomLayerNorm(torch.autograd.Function):
            @staticmethod
            def forward(ctx, embedding):
                layer_norm = torch.nn.LayerNorm(hidden_size, eps=1e-12)
                return layer_norm(embedding)

        class EmbeddingModule(torch.nn.Module):
            def forward(
                self,
                embeddings=None,
            ):
                embedding_output = CustomLayerNorm.apply(embeddings)
                query = embedding_output.transpose(0, 1)
                target_len, batch_size, embedding_dim = query.size()
                # Reshape is used for consuming batch_size, and if it is static,
                # this will be a Constant node in the graph
                query = query.reshape(target_len, batch_size, embedding_dim)
                return query

        embeddings = torch.randn(batch_size, max_position_embeddings, hidden_size)

        f = io.BytesIO()
        torch.onnx.export(
            EmbeddingModule().eval(),
            (embeddings,),
            f,
            input_names=["embeddings"],
            dynamic_axes={
                "embeddings": {
                    0: "batch_size",
                    1: "max_position_embeddings",
                    2: "hidden_size",
                }
            },
            custom_opsets={"com.microsoft": 1},
        )
        model = onnx.load(io.BytesIO(f.getvalue()))

        # If there is a constant node with dim=3 and max_position_embeddings,
        # batch_size, hidden_size as shape, it means the shape becomes static.
        # Normally, with dynamic batch size, this constant node should not exist.
        const_node = [n for n in model.graph.node if n.op_type == "Constant"]
        self.assertNotEqual(len(const_node), 0)
        for node in const_node:
            for a in node.attribute:
                if a.name == "value":
                    shape = onnx.numpy_helper.to_array(a.t)
                    self.assertNotEqual(
                        shape.tolist(),
                        [max_position_embeddings, batch_size, hidden_size],
                    )

    def test_is_fp_for_C_TypeList(self):
        class M(torch.nn.Module):
            def forward(self, x):
                x = x.squeeze(1)
                w = x.shape[2]
                pos = x.view(2, -1).argmax(1)
                x_int = pos % w
                y_int = (pos - x_int) // w
                return y_int, x_int

        model = torch.jit.script(M())
        inputs = torch.randn(2, 4, 6)
        f = io.BytesIO()
        torch.onnx.export(
            model, inputs, f, dynamic_axes={"x": [0, 1]}, input_names=["x"]
        )

    def test_dropout_script(self):
        eg = torch.zeros(1, 2, 3, requires_grad=True)

        @jit_utils._trace(eg)
        def foo(x):
            x = torch.neg(x)
            return F.dropout(x)

        class MyDrop(torch.nn.Module):
            def forward(self, x):
                return foo(x)

        f = io.BytesIO()
        with warnings.catch_warnings(record=True):
            torch.onnx.export(MyDrop(), (eg,), f)

    def test_pack_padded_pad_packed_trace(self):
        from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence

        T, B, C = 3, 5, 7

        class PadPackedWrapper(torch.nn.Module):
            def forward(self, x, seq_lens):
                x = pack_padded_sequence(x, seq_lens)
                x, _ = pad_packed_sequence(x)
                return x

        x = np.ones((T, B, C))
        seq_lens = np.array([3, 3, 2, 2, 1], dtype=np.int32)
        # set padding value so we can test equivalence
        for b in range(B):
            if seq_lens[b] < T:
                x[seq_lens[b] :, b, :] = 0
        seq_lens = torch.from_numpy(seq_lens)
        x = torch.autograd.Variable(torch.from_numpy(x), requires_grad=True)

        m = PadPackedWrapper()
        m_traced = torch.jit.trace(
            m,
            (
                x,
                seq_lens,
            ),
        )

        y = m(x, seq_lens)
        loss = torch.sum(y)
        loss.backward()
        grad = x.grad.clone()
        x.grad.zero_()

        y_traced = m_traced(x, seq_lens)
        loss_traced = torch.sum(y_traced)
        loss_traced.backward()
        grad_traced = x.grad.clone()

        self.assertEqual(y_traced, x)
        self.assertEqual(y_traced, y)
        self.assertEqual(grad, grad_traced)

        f = io.BytesIO()
        torch.onnx.export(m, (x, seq_lens), f)

    # Suppression: ONNX warns when exporting RNNs because of potential batch size mismatch.
    @common_utils.suppress_warnings
    def test_rnn_trace_override(self):
        from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence

        num_layers = 3
        T, B, C = 11, 5, 7

        class RNNTraceWrapper(torch.nn.Module):
            def __init__(self, cell_type):
                super().__init__()
                if cell_type == "RNN":
                    self.rnn = torch.nn.RNN(
                        input_size=C, hidden_size=C, num_layers=num_layers
                    )
                elif cell_type == "LSTM":
                    self.rnn = torch.nn.LSTM(
                        input_size=C, hidden_size=C, num_layers=num_layers
                    )
                elif cell_type == "GRU":
                    self.rnn = torch.nn.GRU(
                        input_size=C, hidden_size=C, num_layers=num_layers
                    )

            def forward(self, x, seq_lens):
                x = pack_padded_sequence(x, seq_lens)
                x, _ = self.rnn(x)
                x, _ = pad_packed_sequence(x)
                return x

        for cell_type in ["RNN", "LSTM", "GRU"]:
            x = torch.ones(T, B, C, requires_grad=True)
            seq_lens = torch.from_numpy(np.array([11, 3, 2, 2, 1], dtype=np.int32))

            m = RNNTraceWrapper(cell_type)
            m_traced = torch.jit.trace(
                m,
                (
                    x,
                    seq_lens,
                ),
            )

            y = m(x, seq_lens)
            loss = torch.sum(y)
            loss.backward()
            grad = x.grad.clone()
            x.grad.zero_()

            y_traced = m_traced(x, seq_lens)
            loss_traced = torch.sum(y_traced)
            loss_traced.backward()
            grad_traced = x.grad.clone()

            self.assertEqual(y_traced, y)
            self.assertEqual(grad, grad_traced)

            f = io.BytesIO()
            torch.onnx.export(m, (x, seq_lens), f)

    def test_pushpackingpastrnn_in_peephole_create_own_gather_input(self):
        from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence

        num_layers = 3
        T, B, C = 11, 5, 7
        mask_start_point = 0

        class LSTMTraceWrapper(torch.nn.Module):
            def __init__(self) -> None:
                super().__init__()

                self.rnn = torch.nn.LSTM(
                    input_size=C, hidden_size=C, num_layers=num_layers
                )

            def forward(self, x, seq_lens):
                mask = torch.arange(mask_start_point, x.shape[1])
                seq_lens = seq_lens[mask]
                x = pack_padded_sequence(x, seq_lens)
                # Calculate sizes and prepare views to our zero buffer to pass as hx
                max_batch_size = x.batch_sizes[0]
                hx = torch.randn(num_layers, max_batch_size, C)
                cx = torch.randn(num_layers, max_batch_size, C)
                x, _ = self.rnn(x, (hx, cx))
                x, _ = pad_packed_sequence(x)
                return x

        x = torch.ones(T, B, C)
        # length 5 because of B
        seq_lens = torch.from_numpy(np.array([11, 3, 2, 2, 1], dtype=np.int32))
        m = LSTMTraceWrapper()

        f = io.BytesIO()
        torch.onnx.export(
            m,
            (x, seq_lens),
            f,
            verbose=True,
            input_names=["input", "seq_len"],
            dynamic_axes={"input": {1: "B"}},
        )
        onnx_proto = onnx.load_model_from_string(f.getvalue())
        # the first argument in onnx::Range should be constant node with value 0
        const_node = []
        constant_input_name = None
        for n in onnx_proto.graph.node:
            if n.op_type == "Constant":
                const_node.append(n)
            elif n.op_type == "Range":
                constant_input_name = n.input[0]
        self.assertNotEqual(constant_input_name, None)
        self.assertNotEqual(len(const_node), 0)

        value = None
        for n in const_node:
            if n.output[0] == constant_input_name:
                value = np.frombuffer(n.attribute[0].t.raw_data, dtype=np.int64)
        self.assertEqual(value, 0)

    def test_trace_fork_wait_inline_onnx(self):
        def fork_body(x):
            return torch.neg(x), torch.neg(x)

        class MyMod(torch.nn.Module):
            def forward(self, x):
                fut = torch.jit._fork(fork_body, x)
                val = torch.jit._wait(fut)
                return val[1]

        # smoke test for ONNX export
        f = io.BytesIO()
        torch.onnx.export(MyMod(), (torch.rand(3, 4),), f)

    def test_trace_detach_onnx_erase(self):
        class Mod(torch.nn.Module):
            def forward(self, x, w):
                return torch.matmul(x, w).detach()

        f = io.BytesIO()
        torch.onnx.export(Mod(), (torch.rand(3, 4), torch.rand(4, 5)), f)

    def test_aten_fallback_must_fallback(self):
        class ModelWithAtenNotONNXOp(torch.nn.Module):
            def forward(self, x, y):
                abcd = x + y
                defg = torch.linalg.qr(abcd)
                return defg

        x = torch.rand(3, 4)
        y = torch.rand(3, 4)
        f = io.BytesIO()
        torch.onnx.export(
            ModelWithAtenNotONNXOp(),
            (x, y),
            f,
            do_constant_folding=False,
            operator_export_type=torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK,
            # support for linalg.qr was added in later op set versions.
            opset_version=9,
        )
        onnx_model = onnx.load(io.BytesIO(f.getvalue()))
        self.assertAtenOp(onnx_model, "linalg_qr")

    def test_onnx_aten(self):
        class ModelWithAtenFmod(torch.nn.Module):
            def forward(self, x, y):
                return torch.fmod(x, y)

        x = torch.randn(3, 4, dtype=torch.float32)
        y = torch.randn(3, 4, dtype=torch.float32)
        f = io.BytesIO()
        torch.onnx.export(
            ModelWithAtenFmod(),
            (x, y),
            f,
            do_constant_folding=False,
            operator_export_type=torch.onnx.OperatorExportTypes.ONNX_ATEN,
        )
        onnx_model = onnx.load(io.BytesIO(f.getvalue()))
        self.assertAtenOp(onnx_model, "fmod", "Tensor")

    def test_onnx_aten_fallback_must_not_fallback(self):
        # For BUILD_CAFFE2=0, aten fallback only when not exportable
        class ONNXExportable(torch.nn.Module):
            def __init__(self) -> None:
                super().__init__()
                self.quant = torch.ao.quantization.QuantStub()
                self.fc1 = torch.nn.Linear(12, 8)
                self.fc2 = torch.nn.Linear(8, 4)
                self.fc3 = torch.nn.Linear(4, 6)
                self.dequant = torch.ao.quantization.DeQuantStub()

            def forward(self, x):
                x = self.quant(x)
                x = x.view((-1, 12))
                h = F.relu(self.fc1(x))
                h = F.relu(self.fc2(h))
                h = F.relu(self.fc3(h))
                h = self.dequant(h)
                return h

        dummy_input = torch.randn(12)
        f = io.BytesIO()
        torch.onnx.export(
            ONNXExportable(),
            (dummy_input,),
            f,
            do_constant_folding=False,
            operator_export_type=torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK,
        )
        onnx_model = onnx.load(io.BytesIO(f.getvalue()))
        all_aten_nodes = [
            p
            for p in onnx_model.graph.node
            if p.op_type == "ATen" and p.domain == "org.pytorch.aten"
        ]
        self.assertEqual(len(all_aten_nodes), 0)

    def test_cat_with_empty_tensor(self):
        class NoopConcat(torch.nn.Module):
            def forward(self, x):
                return torch.cat((torch.Tensor([]), x))

        x = torch.randn(4, 5, 6)
        # TODO: Parametrize this test for opset_version
        for opset_version in {9, 11}:
            f = io.BytesIO()
            torch.onnx.export(NoopConcat(), (x,), f, opset_version=opset_version)
            loaded_model = onnx.load_from_string(f.getvalue())
            self.assertEqual(
                len(loaded_model.graph.output[0].type.tensor_type.shape.dim), 3
            )
            for idx, dim in enumerate(x.shape):
                self.assertEqual(
                    loaded_model.graph.output[0]
                    .type.tensor_type.shape.dim[idx]
                    .dim_value,
                    dim,
                )

    def test_col2im(self):
        # This test can be moved to test/onnx/test_pytorch_onnx_onnxruntime.py when ORT implement ::Col2Im

        # Random batched RGB 32x32 image-shaped input tensor of batch size 64
        original_image_inputs = torch.randn((64, 3, 32, 32))
        output_size = tuple(original_image_inputs.shape[2:])
        kernel_size = (1, 2)
        dilation = 3
        padding = 2
        stride = 1
        model_im2col = torch.nn.Unfold(
            kernel_size, dilation=dilation, padding=padding, stride=stride
        )
        blocks = model_im2col(original_image_inputs)

        model = torch.nn.Fold(
            output_size=output_size,
            kernel_size=kernel_size,
            dilation=dilation,
            padding=padding,
            stride=stride,
        )
        f = io.BytesIO()
        torch.onnx.export(model, (blocks,), f, opset_version=18)

        onnx_model = onnx.load(io.BytesIO(f.getvalue()))
        self.assertEqual(onnx_model.graph.node[-1].op_type, "Col2Im")
        self.assertEqual(onnx_model.graph.node[-1].domain, "")
        self.assertEqual(len(onnx_model.graph.node[-1].input), 3)
        self.assertEqual(onnx_model.graph.node[-1].attribute[0].name, "dilations")
        self.assertEqual(onnx_model.graph.node[-1].attribute[1].name, "pads")
        self.assertEqual(onnx_model.graph.node[-1].attribute[2].name, "strides")

    @unittest.skipIf(
        not torch.hub._check_module_exists("torch_scatter"),
        "torch_scatter not installed.",
    )
    def test_random_namespace_custom_op_is_onnx_exportable(self):
        from torch_scatter import scatter_max  # type: ignore[import]

        class MyModel(torch.nn.Module):
            def forward(self, src: torch.Tensor, idx: torch.Tensor):
                return scatter_max(src, idx)

        m = MyModel().eval()
        src = torch.ones([3, 10], dtype=torch.float32)
        idx = torch.randint(0, 4, [3, 10], dtype=torch.long)

        def sym_scatter_max(g, src, index, dim, out, dim_size):
            return g.op(
                "torch_scatter::scatter_max", src, index, dim_size_i=-1, outputs=2
            )

        torch.onnx.register_custom_op_symbolic(
            "torch_scatter::scatter_max", sym_scatter_max, 1
        )
        f = io.BytesIO()
        with torch.no_grad():
            torch.onnx.export(
                m,
                (src, idx),
                f,
                opset_version=13,
                custom_opsets={"torch_scatter": 1},
                do_constant_folding=True,
            )

    @common_utils.parametrize("fp8_dtype", [torch.float8_e4m3fn, torch.float8_e5m2])
    def test_fp8_export(self, fp8_dtype: torch.dtype):
        class Model(torch.nn.Module):
            def forward(self, x):
                return x.to(torch.float32)

        x = torch.randn(2, 3).to(fp8_dtype)

        f = io.BytesIO()
        torch.onnx.export(Model(), x, f, opset_version=19)
        onnx.checker.check_model(f.getvalue())

        onnx_type = {
            torch.float8_e4m3fn: 17,
            torch.float8_e5m2: 19,
        }  # From https://github.com/onnx/onnx/blob/main/onnx/onnx.proto3#L512-L521
        loaded_model = onnx.load_from_string(f.getvalue())
        self.assertEqual(
            loaded_model.graph.input[0].type.tensor_type.elem_type, onnx_type[fp8_dtype]
        )


class TestQuantizeEagerONNXExport(common_utils.TestCase):
    def _test_lower_graph_impl(self, model, data):
        model.qconfig = torch.ao.quantization.default_qconfig
        model = torch.ao.quantization.prepare(model)
        model = torch.ao.quantization.convert(model)

        _ = model(data)
        input_names = ["x"]

        def _export_to_onnx(model, input, input_names):
            traced = torch.jit.trace(model, input)
            buf = io.BytesIO()
            torch.jit.save(traced, buf)
            buf.seek(0)

            model = torch.jit.load(buf)
            f = io.BytesIO()
            torch.onnx.export(
                model,
                input,
                f,
                input_names=input_names,
                operator_export_type=torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK,
                opset_version=9,
            )

        _export_to_onnx(model, data, input_names)

    @common_quantization.skipIfNoFBGEMM
    @unittest.skip(
        "onnx opset9 does not support quantize_per_tensor and caffe2 \
    does not support conv3d"
    )
    def test_lower_graph_conv3d(self):
        model = torch.ao.quantization.QuantWrapper(
            torch.nn.Conv3d(3, 5, 2, bias=True)
        ).to(dtype=torch.float)
        data_numpy = np.random.rand(1, 3, 6, 6, 6).astype(np.float32)
        data = torch.from_numpy(data_numpy).to(dtype=torch.float)
        self._test_lower_graph_impl(model, data)

    @pytorch_test_common.skipIfNoCuda
    def test_composed_layer_norm_small_eps_fp16_keep_double(self):
        class Net(torch.nn.Module):
            def __init__(self, C):
                super().__init__()
                self.layer_norm = torch.nn.LayerNorm(C, eps=1e-8)

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

        N, C = 8, 4
        model = Net(C).cuda().half()
        x = torch.randn(N, C).cuda().half()
        f = io.BytesIO()
        torch.onnx.export(model, (x,), f, opset_version=14)
        onnx_model = onnx.load_from_string(f.getvalue())
        const_node = [n for n in onnx_model.graph.node if n.op_type == "Constant"]
        self.assertNotEqual(len(const_node), 0)
        double_type_count = 0
        for node in const_node:
            for a in node.attribute:
                # EPS constant should be in double type
                if a.name == "value" and a.t.data_type == 11:
                    double_type_count += 1
        self.assertNotEqual(double_type_count, 0)

    @pytorch_test_common.skipIfNoCuda
    def test_aten_device_with_index(self):
        from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

        tokenizer = AutoTokenizer.from_pretrained("google/flan-t5-small")
        model = AutoModelForSeq2SeqLM.from_pretrained("google/flan-t5-small")
        model = torch.compile(model, backend="onnxrt")
        model = model.eval()
        device = "cuda:0"
        model = model.to(device)
        ids = tokenizer.batch_encode_plus(["This is a test"], return_tensors="pt").to(
            device
        )

        with torch.no_grad():
            _ = model(
                input_ids=ids["input_ids"],
                attention_mask=ids["attention_mask"],
                decoder_input_ids=ids["input_ids"],
                decoder_attention_mask=ids["attention_mask"],
            )

    def test_aten_linalg_vector_norm_with_reducel2(self):
        class Net(torch.nn.Module):
            def forward(self, x):
                x = F.normalize(x)
                return x

        f = io.BytesIO()
        torch.onnx.export(Net(), (torch.randn(1, 2, 2),), f)
        onnx_model = onnx.load_from_string(f.getvalue())
        onnx_nodes = [n.op_type for n in onnx_model.graph.node]
        self.assertIn("ReduceL2", onnx_nodes)


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