test_foreach.py

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

from contextlib import nullcontext
from numbers import Number
import random
import re
import torch
import unittest
import itertools
import weakref

from torch.testing import make_tensor
from torch.testing._comparison import default_tolerances
from torch.testing._internal.common_cuda import TEST_MULTIGPU
from torch.testing._internal.common_utils import \
    TestCase, run_tests, TEST_WITH_ROCM, skipIfTorchDynamo, parametrize, gradcheck
from torch.testing._internal.common_device_type import \
    (instantiate_device_type_tests, dtypes, onlyCUDA, ops, OpDTypes)
from torch.testing._internal.common_methods_invocations import (
    foreach_unary_op_db, foreach_binary_op_db, foreach_pointwise_op_db,
    foreach_reduce_op_db, foreach_other_op_db)
from torch.testing._internal.common_dtype import (
    all_types_and_complex_and, floating_types_and, floating_types, integral_types_and,
)


_BOOL_SUB_ERR_MSG = "Subtraction, the `-` operator"


class RegularFuncWrapper:
    def __init__(self, func):
        self.func = func

    def __call__(self, inputs, scalars=None, **kwargs):
        if scalars is not None:
            assert len(inputs) == 3
            # We need to distribute each scalar to the regular func and it needs
            # special consideration as it is a keyword only argument to the
            # regular func. (Strangely, it is not a keyword only argument to the
            # foreach func)
            return [self.func(*i, value=scalars[idx], **kwargs) for idx, i in enumerate(zip(*inputs))]
        if len(inputs) == 2 and isinstance(inputs[1], (Number, torch.Tensor)):
            # binary op with tensorlist and scalar.
            inputs[1] = [inputs[1] for _ in range(len(inputs[0]))]
        return [self.func(*i, **kwargs) for i in zip(*inputs)]


class ForeachFuncWrapper:
    def __init__(self, func):
        self.func = func
        # Some foreach functions don't have in-place implementations.
        self.is_inplace = False if func is None else func.__name__.endswith('_')

    def __call__(self, inputs, is_cuda, expect_fastpath, **kwargs):
        actual = None
        zero_size = kwargs.pop("zero_size", False)
        if (
            is_cuda and
            torch.autograd.kineto_available() and
            torch.profiler.ProfilerActivity.CUDA in torch.profiler.supported_activities()
        ):
            with torch.profiler.profile() as p:
                actual = self.func(*inputs, **kwargs)
            keys = tuple([e.key for e in p.key_averages()])
            mta_called = any("multi_tensor_apply_kernel" in k for k in keys)
            assert mta_called == (expect_fastpath and (not zero_size))
        else:
            actual = self.func(*inputs, **kwargs)
        # note(mkozuki): inplace foreach functions are void functions.
        return inputs[0] if self.is_inplace else actual


class InplaceForeachVersionBumpCheck:

    def __init__(self, testcase: TestCase, tensorlist: "List[torch.Tensor]") -> None:  # noqa: F821
        self._testcase = testcase
        self._tensorlist = tensorlist
        self._orig_version_counts = [t._version for t in tensorlist]

    def __enter__(self):
        pass

    def __exit__(self, exc_type, exc_value, traceback):
        # note(crcrpar): some methods e.g. `_binary_test` could call the given inplace function multiple times
        self._testcase.assertGreaterEqual([t._version for t in self._tensorlist], self._orig_version_counts)


def get_transform_func(num_tensors, dtype, device, is_fastpath):
    def transform(t):
        if not torch.is_tensor(t):
            return t
        if torch.is_tensor(t) and t.ndim == 0:
            return t
        return make_tensor(
            (num_tensors, num_tensors), dtype=dtype, device=device,
            requires_grad=True, noncontiguous=not is_fastpath,
        )

    return transform


# note(crcrpar): `zero_size` is `False` unless (dtype, device) == (torch.float32, "cuda")
# as the pair would go through `multi_tensor_apply_kernel` if inputs are not zero size.
class TestForeach(TestCase):
    @property
    def is_cuda(self):
        return self.device_type == 'cuda'

    def _get_funcs(self, op):
        return (
            ForeachFuncWrapper(op.method_variant),
            RegularFuncWrapper(op.ref),
            ForeachFuncWrapper(op.inplace_variant),
            RegularFuncWrapper(op.ref_inplace),
        )

    # note(crcrpar): Make sure 0-size tensors are appropriately ignored by `multi_tensor_apply`
    # which is originally reported in https://github.com/pytorch/pytorch/issues/94865.
    # rel:
    #   - https://github.com/pytorch/pytorch/pull/94655
    #   - https://github.com/pytorch/pytorch/issues/100701
    #   - https://github.com/pytorch/pytorch/pull/100811
    @onlyCUDA
    @ops(
        foreach_unary_op_db + foreach_binary_op_db + foreach_pointwise_op_db + foreach_reduce_op_db + foreach_other_op_db,
        dtypes=(torch.float32,)
    )
    def test_all_zero_size_tensors_do_not_launch_kernel(self, device, dtype, op):
        wrapped_op, _, inplace_op, _ = self._get_funcs(op)

        for sample in op.sample_zero_size_inputs(device, dtype):
            if op.supports_out:
                wrapped_op((sample.input, *sample.args), is_cuda=self.is_cuda, expect_fastpath=True, zero_size=True)
            with InplaceForeachVersionBumpCheck(self, sample.input):
                inplace_op((sample.input, *sample.args), is_cuda=self.is_cuda, expect_fastpath=True, zero_size=True)

    @unittest.skipIf(TEST_WITH_ROCM, "Skipped on ROCm, since it is failing on ROCm 5.7")
    @ops(
        foreach_unary_op_db + foreach_binary_op_db + foreach_pointwise_op_db + foreach_reduce_op_db + foreach_other_op_db,
    )
    @parametrize(
        "noncontiguous,inplace",
        [(False, False), (False, True), (True, False), (True, True)],
        name_fn=lambda x, y: '{}_{}'.format(
            'fastpath' if not x else 'slowpath', 'inplace' if y else 'outplace'
        )
    )
    def test_parity(self, device, dtype, op, noncontiguous, inplace):
        if inplace:
            _, _, func, ref = self._get_funcs(op)
        else:
            func, ref, _, _ = self._get_funcs(op)
        for sample in op.sample_inputs(device, dtype, noncontiguous=noncontiguous):
            ref_kwargs = sample.kwargs
            # div promotes ints to floats, so we cannot go on the fastpath there
            div_slowpath = dtype in integral_types_and(torch.bool) and op.name == '_foreach_div'
            expect_fastpath = not (noncontiguous or sample.disable_fastpath or div_slowpath)
            ref_input, ctxmgr = sample.input, nullcontext()
            if inplace:
                with torch.no_grad():
                    ref_input = [t.clone().detach() for t in sample.input]
                ctxmgr = InplaceForeachVersionBumpCheck(self, sample.input)
            try:
                with ctxmgr:
                    actual = func([sample.input, *sample.args], self.is_cuda, expect_fastpath, **sample.kwargs)
            except Exception as e:
                with (
                    self.assertRaisesRegex(type(e), re.escape(str(e)))
                    if not (op.has_no_in_place or not op.supports_out)
                    else self.assertRaises(type(e))
                ):
                    ref([ref_input, *sample.ref_args], **ref_kwargs)
            else:
                expected = ref([ref_input, *sample.ref_args], **ref_kwargs)
                self.assertEqual(expected, actual)

    def _binary_test(
        self,
        dtype, op, ref, inputs, is_fastpath, is_inplace,
        *,
        alpha, scalar_self_arg: bool,
    ):
        ref_inputs = [[t.clone().detach() for t in inputs[0]], inputs[1]] if is_inplace else inputs
        try:
            with InplaceForeachVersionBumpCheck(self, inputs[0]) if op.is_inplace else nullcontext():
                actual = op(inputs, self.is_cuda, is_fastpath)
        except RuntimeError as e:
            with self.assertRaisesRegex(type(e), re.escape(str(e))):
                if not scalar_self_arg:
                    ref(ref_inputs)
                else:
                    [ref.func(ref_inputs[0], t) for t in ref_inputs[1]]
        else:
            expected = ref(ref_inputs) if not scalar_self_arg else [ref.func(ref_inputs[0], t) for t in ref_inputs[1]]
            self.assertEqual(actual, expected)
        if alpha is not None and not scalar_self_arg:
            kwargs = {'alpha': alpha}
            ref_inputs = inputs
            try:
                op_kwargs = {}
                op_kwargs.update(kwargs)
                with InplaceForeachVersionBumpCheck(self, inputs[0]) if op.is_inplace else nullcontext():
                    actual = op(inputs, self.is_cuda, is_fastpath, **op_kwargs)
            except RuntimeError as e:
                with self.assertRaisesRegex(type(e), re.escape(str(e))):
                    ref(ref_inputs, **kwargs)
            else:
                expected = ref(ref_inputs, **kwargs)
                if dtype in (torch.float16, torch.bfloat16) and TEST_WITH_ROCM:
                    self.assertEqual(expected, actual, atol=1.e-3, rtol=default_tolerances(dtype)[0])
                else:
                    self.assertEqual(expected, actual)

    @ops(filter(lambda op: op.supports_scalar_self_arg, foreach_binary_op_db))
    @parametrize("is_fastpath", (True, False))
    def test_binary_op_with_scalar_self_support(self, device, dtype, op, is_fastpath):

        def clone(arg):
            if isinstance(arg, (list, tuple)):
                return [clone(a) for a in arg]
            if torch.is_tensor(arg):
                return arg.clone().detach().requires_grad_()
            else:
                return arg

        scalar_self_arg_test_complete = False
        for i, sample in enumerate(op.sample_inputs(device, dtype, noncontiguous=not is_fastpath)):
            (rhs_arg,) = sample.args
            kwargs = {} or sample.kwargs
            alpha = kwargs.pop("alpha", None)
            wrapped_op, ref, inplace_op, inplace_ref = self._get_funcs(op)
            if isinstance(rhs_arg, Number) and not scalar_self_arg_test_complete:
                scalar_self_arg_test_complete = True
                self._binary_test(
                    dtype, wrapped_op, ref, [rhs_arg, sample.input], is_fastpath, False,
                    alpha=alpha, scalar_self_arg=True,
                )
                if op.supports_autograd and dtype == torch.float32:
                    transformed_sample = sample.transform(
                        get_transform_func(len(sample.input), dtype, device, is_fastpath))
                    tensors = transformed_sample.input
                    (rhs_arg,) = transformed_sample.args
                    ref_tensors, ref_rhs_arg = clone(tensors), clone(rhs_arg)
                    sum(wrapped_op(
                        [rhs_arg, tensors], is_cuda=False, expect_fastpath=False
                    )).mean().backward()
                    sum([ref.func(ref_rhs_arg, t) for t in ref_tensors]).mean().backward()
                    self.assertEqual([t.grad for t in tensors], [t.grad for t in ref_tensors])

    @ops(foreach_pointwise_op_db)
    @parametrize("is_fastpath", (True, False))
    def test_pointwise_op_with_tensor_of_scalarlist_overload(self, device, dtype, op, is_fastpath):
        for sample in op.sample_inputs(device, dtype, noncontiguous=not is_fastpath):
            assert isinstance(sample.args, tuple)
            assert len(sample.args) == 2
            inputs = [sample.input, *sample.args]
            kwargs = sample.kwargs.copy()
            disable_fastpath = sample.disable_fastpath and is_fastpath
            wrapped_op, ref, inplace_op, inplace_ref = self._get_funcs(op)
            scalars = kwargs.pop("scalars", None)

            if is_fastpath and scalars:
                sample = sample.transform(lambda t: t.clone().detach() if torch.is_tensor(t) else t)
                inputs = [sample.input, *sample.args]
                tensor_values = torch.tensor(scalars)
                # 1D Tensor of scalars
                for is_inplace, op_, ref_ in ((False, wrapped_op, ref), (True, inplace_op, inplace_ref)):
                    self._pointwise_test(
                        op_, ref_, inputs, is_fastpath and not disable_fastpath, is_inplace,
                        scalars=tensor_values, **kwargs)
                    self._pointwise_test(
                        op_, ref_, inputs, is_fastpath and not disable_fastpath, is_inplace,
                        scalars=tensor_values[0],
                        custom_values_err="Expected packed scalar Tensor to be of dimension 1. Got 0 instead.",
                        **kwargs,
                    )
                    if self.is_cuda:
                        self._pointwise_test(
                            op_, ref_, inputs, is_fastpath and not disable_fastpath, is_inplace,
                            scalars=tensor_values.cuda(),
                            custom_values_err="Expected scalars to be on CPU, got cuda:0 instead.",
                            **kwargs,
                        )
                    self._pointwise_test(
                        op_, ref_, inputs, is_fastpath and not disable_fastpath, is_inplace,
                        scalars=tensor_values[:2],
                        custom_values_err=f"Expected length of scalars to match input of length {len(scalars)} but got 2 instead.",
                        **kwargs,
                    )
                    self._pointwise_test(
                        op_, ref_, inputs, is_fastpath and not disable_fastpath, is_inplace,
                        scalars=torch.tensor([[0, 1], [2, 3]])[:, 1],
                        custom_values_err="Expected scalars to be contiguous.",
                        **kwargs,
                    )

            # Tests of implicit broadcasting
            N = len(sample.input)
            inputs = [
                [make_tensor((N, N), device=device, dtype=dtype, noncontiguous=not is_fastpath) for _ in range(N)],
                [
                    make_tensor((N - i, 1), device=device, dtype=dtype, noncontiguous=not is_fastpath)
                    for i in range(N)
                ],
                [
                    make_tensor((1, N - i), device=device, dtype=dtype, noncontiguous=not is_fastpath)
                    for i in range(N)
                ],
            ]
            self._pointwise_test(
                wrapped_op, ref, inputs, is_fastpath and disable_fastpath, is_inplace=False,
                scalars=scalars, **kwargs)
            self._pointwise_test(
                inplace_op, inplace_ref, inputs, is_fastpath and disable_fastpath,
                is_inplace=True, scalars=scalars, **kwargs)

    def _pointwise_test(
        self,
        op, ref, inputs, is_fastpath, is_inplace,
        *,
        scalars=None, custom_values_err=None, **kwargs
    ):
        ref_inputs = [[t.clone().detach() for t in inputs[0]], inputs[1], inputs[2]] if is_inplace else inputs
        try:
            with (InplaceForeachVersionBumpCheck(self, inputs[0]) if is_inplace else nullcontext()):
                actual = op(inputs, self.is_cuda, is_fastpath, **kwargs)
        except RuntimeError as e:
            with self.assertRaisesRegex(type(e), re.escape(str(e))):
                ref(ref_inputs, **kwargs)
        else:
            expected = ref(ref_inputs, **kwargs)
            self.assertEqual(expected, actual)
        if scalars is not None:
            kwargs = kwargs.copy()
            kwargs["scalars"] = scalars
            try:
                actual = op(inputs, self.is_cuda, is_fastpath, **kwargs)
            except RuntimeError as e:
                # Match with error messages from regular non-foreach reference if no
                # custom error message was provided.
                if custom_values_err is None:
                    with self.assertRaisesRegex(type(e), re.escape(str(e))):
                        ref(ref_inputs, **kwargs)
                else:
                    self.assertEqual(re.escape(str(e)), re.escape(custom_values_err))
            else:
                expected = ref(ref_inputs, **kwargs)
                self.assertEqual(expected, actual)

    @dtypes(*all_types_and_complex_and(torch.half, torch.bfloat16))
    def test_add_scalar_with_empty_list_and_empty_tensor(self, device, dtype):
        # TODO: enable empty list case
        for tensors in [[torch.randn([0], device=device, dtype=dtype)],
                        [torch.empty_strided((0, 1), (0, 0), dtype=dtype, device=device)]]:
            res = torch._foreach_add(tensors, 1)
            self.assertEqual(res, tensors)

            torch._foreach_add_(tensors, 1)
            self.assertEqual(res, tensors)

            # Regression test for https://github.com/pytorch/pytorch/issues/113156
            torch._foreach_mul_(tensors, 1)

    @ops(
        filter(lambda op: op.supports_out, foreach_binary_op_db),
        dtypes=OpDTypes.supported,
    )
    def test_binary_op_scalar_with_overlapping_tensors(self, device, dtype, op):
        foreach_op, ref = op.method_variant, op.ref
        tensors = [torch.ones(1, 1, device=device, dtype=dtype).expand(2, 1, 3)]

        if ref == torch.sub and dtype == torch.bool:
            with self.assertRaisesRegex(RuntimeError, re.escape(_BOOL_SUB_ERR_MSG)):
                [ref(t, 1) for t in tensors]
            with self.assertRaisesRegex(RuntimeError, re.escape(_BOOL_SUB_ERR_MSG)):
                foreach_op(tensors, 1)
            return

        expected = [ref(t, 1) for t in tensors]
        res = foreach_op(tensors, 1)
        self.assertEqual(res, expected)

    @ops(
        filter(lambda op: op.supports_out, foreach_binary_op_db),
        allowed_dtypes=[torch.float],
    )
    def test_binary_op_scalar_with_different_tensor_dtypes(self, device, dtype, op):
        foreach_op = op.method_variant
        tensors = [
            torch.tensor([1.1], dtype=torch.float, device=device),
            torch.tensor([1], dtype=torch.long, device=device),
        ]
        runtime_error = None
        try:
            foreach_op(tensors, 1)
        except RuntimeError as e:
            runtime_error = e
        self.assertIsNone(runtime_error)

    @skipIfTorchDynamo("Different error msgs, TODO")
    @ops(
        filter(lambda op: op.supports_out, foreach_binary_op_db),
        dtypes=OpDTypes.supported,
    )
    def test_binary_op_list_error_cases(self, device, dtype, op):
        foreach_op, foreach_op_, ref, ref_ = op.method_variant, op.inplace_variant, op.ref, op.ref_inplace
        tensors1 = []
        tensors2 = []
        ops_to_test = [foreach_op, foreach_op_]

        # Empty lists
        for fop in ops_to_test:
            with self.assertRaisesRegex(RuntimeError, "There were no tensor arguments to this function"):
                fop(tensors1, tensors2)

        # One empty list
        tensors1.append(torch.tensor([1], device=device, dtype=dtype))
        for fop in ops_to_test:
            with self.assertRaisesRegex(RuntimeError, "Tensor list must have same number of elements as scalar list."):
                fop(tensors1, tensors2)

        # Lists have different amount of tensors
        tensors2.append(torch.tensor([1], device=device))
        tensors2.append(torch.tensor([1], device=device))
        for fop in ops_to_test:
            with self.assertRaisesRegex(RuntimeError, "Tensor lists must have the same number of tensors, got 1 and 2"):
                fop(tensors1, tensors2)
            with self.assertRaisesRegex(RuntimeError, "Tensor lists must have the same number of tensors, got 2 and 1"):
                fop(tensors2, tensors1)

        # Corresponding tensors with different sizes that aren't compatible with broadcast
        # If sizes are different then foreach chooses slow path, thus error messages are expected
        # to be the same as torch regular function.
        tensors1 = [torch.zeros(10, 10, device=device, dtype=dtype) for _ in range(10)]
        tensors2 = [torch.ones(11, 11, device=device, dtype=dtype) for _ in range(10)]
        try:
            foreach_op(tensors1, tensors2)
        except RuntimeError as e:
            with self.assertRaisesRegex(type(e), re.escape(str(e))):
                [ref(t1, t2) for t1, t2 in zip(tensors1, tensors2)]
        try:
            foreach_op_(tensors1, tensors2)
        except RuntimeError as e:
            with self.assertRaisesRegex(type(e), re.escape(str(e))):
                [ref_(t1, t2) for t1, t2 in zip(tensors1, tensors2)]

        # different devices
        if self.device_type == "cuda" and torch.cuda.device_count() > 1:
            tensor1 = torch.zeros(10, 10, device="cuda:0", dtype=dtype)
            tensor2 = torch.ones(10, 10, device="cuda:1", dtype=dtype)
            if dtype == torch.bool and foreach_op == torch._foreach_sub:
                for fop in ops_to_test:
                    with self.assertRaisesRegex(RuntimeError, re.escape(_BOOL_SUB_ERR_MSG)):
                        fop([tensor1], [tensor2])
                return
            with self.assertRaisesRegex(RuntimeError, "Expected all tensors to be on the same device"):
                foreach_op([tensor1], [tensor2])
            if dtype in integral_types_and(torch.bool) and foreach_op == torch._foreach_div:
                with self.assertRaisesRegex(RuntimeError, "result type"):
                    foreach_op_([tensor1], [tensor2])
            else:
                with self.assertRaisesRegex(RuntimeError, "Expected all tensors to be on the same device"):
                    foreach_op_([tensor1], [tensor2])

    @unittest.skipIf(not torch.cuda.is_available(), "CUDA not found")
    @ops(
        filter(lambda op: op.supports_out, foreach_binary_op_db),
        dtypes=OpDTypes.supported,
    )
    def test_binary_op_list_slow_path(self, device, dtype, op):
        foreach_op, native_op, foreach_op_, native_op_ = self._get_funcs(op)
        # 0-strides
        tensor1 = make_tensor((10, 10), dtype=dtype, device=device)
        tensor2 = make_tensor((1,), device=device, dtype=dtype).expand_as(tensor1)
        inputs = ([tensor1], [tensor2])
        self._binary_test(
            dtype, foreach_op, native_op, inputs, is_fastpath=False, is_inplace=False,
            alpha=None, scalar_self_arg=False)
        self._binary_test(
            dtype, foreach_op_, native_op_, inputs, is_fastpath=False, is_inplace=True,
            alpha=None, scalar_self_arg=False)

        # different strides
        tensor1 = torch.zeros(10, 10, device=device, dtype=dtype)
        tensor2 = torch.ones(10, 10, device=device, dtype=dtype)
        inputs = ([tensor1], [tensor2.t()])
        self._binary_test(
            dtype, foreach_op, native_op, inputs, is_fastpath=False, is_inplace=False,
            alpha=None, scalar_self_arg=False)
        self._binary_test(
            dtype, foreach_op_, native_op_, inputs, is_fastpath=False, is_inplace=True,
            alpha=None, scalar_self_arg=False)

        # non contiguous
        tensor1 = make_tensor((5, 2, 1, 3), device=device, dtype=dtype, noncontiguous=True)
        tensor2 = make_tensor((5, 2, 1, 3), device=device, dtype=dtype, noncontiguous=True)
        self.assertFalse(tensor1.is_contiguous())
        self.assertFalse(tensor2.is_contiguous())
        inputs = ([tensor1], [tensor2])
        self._binary_test(
            dtype, foreach_op, native_op, inputs, is_fastpath=False, is_inplace=False,
            alpha=None, scalar_self_arg=False)
        self._binary_test(
            dtype, foreach_op_, native_op_, inputs, is_fastpath=False, is_inplace=True,
            alpha=None, scalar_self_arg=False)

        # sliced tensor
        tensor1 = make_tensor((5, 2, 1, 3), device=device, dtype=dtype)
        tensor2 = make_tensor((5, 2, 1, 3 * 7), device=device, dtype=dtype)[:, :, :, ::7]
        inputs = ([tensor1], [tensor2])
        self._binary_test(
            dtype, foreach_op, native_op, inputs, is_fastpath=False, is_inplace=False,
            alpha=None, scalar_self_arg=False)
        self._binary_test(
            dtype, foreach_op_, native_op_, inputs, is_fastpath=False, is_inplace=True,
            alpha=None, scalar_self_arg=False)

    @ops(
        filter(lambda op: op.supports_out, foreach_binary_op_db),
        dtypes=floating_types_and(torch.half, torch.bfloat16),
    )
    def test_binary_op_float_inf_nan(self, device, dtype, op):
        inputs = (
            [
                torch.tensor([float("inf")], device=device, dtype=dtype),
                torch.tensor([-float("inf")], device=device, dtype=dtype),
                torch.tensor([float("nan")], device=device, dtype=dtype),
                torch.tensor([float("nan")], device=device, dtype=dtype),
            ],
            [
                torch.tensor([-float("inf")], device=device, dtype=dtype),
                torch.tensor([float("inf")], device=device, dtype=dtype),
                torch.tensor([float("inf")], device=device, dtype=dtype),
                torch.tensor([float("nan")], device=device, dtype=dtype),
            ],
        )
        op, ref, inplace_op, inplace_ref = self._get_funcs(op)
        self._binary_test(dtype, op, ref, inputs, True, False, alpha=None, scalar_self_arg=False)
        self._binary_test(
            dtype, inplace_op, inplace_ref, inputs, True, True, alpha=None, scalar_self_arg=False
        )

    # note: Below three tests (postfixed with `_tensors_on_different_devices`)
    # checks whether foreach works with lists of tensors on different devices
    # but tensors of the same index are on the same device, e.g., ['cuda', 'cpu].
    @onlyCUDA
    @ops(foreach_unary_op_db)
    def test_unary_op_tensors_on_different_devices(self, device, dtype, op):
        method, ref, inplace_method, ref_inplace = self._get_funcs(op)
        # tensors: ['cuda', 'cpu]
        tensors = next(iter(op.sample_inputs(device, dtype, num_input_tensors=[2]))).input
        tensors[1] = tensors[1].to("cpu")
        if not op.supports_out:
            try:
                actual = method((tensors,), False, False, zero_size=False)
            except RuntimeError as e:
                with self.assertRaisesRegex(type(e), str(e)):
                    ref((tensors,))
            else:
                expected = ref((tensors,))
                self.assertEqual(expected, actual)

        try:
            inplace_method((tensors,), False, False, zero_size=False)
        except RuntimeError as e:
            with self.assertRaisesRegex(type(e), str(e)):
                ref_inplace((tensors,))
        else:
            if not op.supports_out:
                self.assertEqual(expected, tensors)
            else:
                self.assertEqual([torch.zeros_like(t) for t in tensors], tensors)

    @onlyCUDA
    @ops(filter(lambda op: op.supports_out, foreach_binary_op_db))
    def test_binary_op_tensors_on_different_devices(self, device, dtype, op):
        # `tensors1`: ['cuda', 'cpu']
        # `tensors2`: ['cuda', 'cpu']
        _cuda_tensors = next(iter(op.sample_inputs(device, dtype, num_input_tensors=[2], same_size=True))).input
        _cpu_tensors = next(iter(op.sample_inputs("cpu", dtype, num_input_tensors=[2], same_size=True))).input
        tensors1, tensors2 = list(zip(_cuda_tensors, _cpu_tensors))

        foreach_op, foreach_op_ = op.method_variant, op.inplace_variant
        native_op, native_op_ = op.ref, op.ref_inplace
        try:
            actual = foreach_op(tensors1, tensors2)
        except RuntimeError as e:
            with self.assertRaisesRegex(type(e), re.escape(str(e))):
                [native_op(t1, t2) for t1, t2 in zip(tensors1, tensors2)]
        else:
            expected = [native_op(t1, t2) for t1, t2 in zip(tensors1, tensors2)]
            self.assertEqual(expected, actual)
        try:
            foreach_op_(tensors1, tensors2)
        except RuntimeError as e:
            with self.assertRaisesRegex(type(e), re.escape(str(e))):
                [native_op_(t1, t2) for t1, t2 in zip(tensors1, tensors2)]
        else:
            self.assertEqual(actual, tensors1)

    @onlyCUDA
    @ops(foreach_pointwise_op_db, allowed_dtypes=floating_types())
    def test_pointwise_op_tensors_on_different_devices(self, device, dtype, op):
        # tensors1: ['cuda', 'cpu]
        # tensors2: ['cuda', 'cpu]
        # tensors3: ['cuda', 'cpu]
        # first tensorlist is zero-size when float32
        _cuda_tensors = list(
            op.sample_inputs(device, dtype, num_input_tensors=[3], same_size=True)
        )[int(dtype == torch.float32)].input
        _cpu_tensors = next(iter(op.sample_inputs("cpu", dtype, num_input_tensors=[3], same_size=True))).input
        tensors1, tensors2, tensors3 = list(zip(_cuda_tensors, _cpu_tensors))

        foreach_op, foreach_op_, native_op = op.method_variant, op.inplace_variant, op.ref
        actual = foreach_op(tensors1, tensors2, tensors3)
        expected = [native_op(*_cuda_tensors), native_op(*_cpu_tensors)]
        self.assertEqual(expected, actual)

        # note(mkozuki): Limiting dtypes to FP32&FP64, we can safely run inplace ops.
        foreach_op_(tensors1, tensors2, tensors3)
        self.assertEqual(expected, tensors1)

    # note: BFloat16 has the same number of exponent bits as FP32
    # so if squared L2 norm overflows in BF16, then it also overflows in FP32.
    @onlyCUDA
    @ops(foreach_reduce_op_db, allowed_dtypes=(torch.half, torch.bfloat16))
    def test_foreach_l2_large_value_input(self, device, dtype, op):
        ord, N = 2, 10
        max_value = torch.finfo(dtype).max
        scaler = torch.tensor([max_value]).sqrt().to(device=device, dtype=dtype)
        inputs = ([
            t * scaler for t in next(iter(op.sample_inputs(device, dtype, requries_grad=True, num_input_tensors=[N], low=1))).input
        ],)
        # make sure that the min. of squared L2 norm value per tensor is greater than the max value of `dtype`.
        self.assertTrue(scaler * scaler * N > max_value)
        fn, ref_fn, *_ = self._get_funcs(op)
        actual = fn(inputs, is_cuda=True, expect_fastpath=True, ord=ord, zero_size=False)
        expect = ref_fn(inputs, ord=ord)

        if dtype == torch.float16:
            # making sure the reference L2 norm values are in the range of FP16.
            self.assertFalse(any(torch.isinf(e) for e in expect))
        else:
            self.assertTrue(all(
                inputs[0][i].numel() == 0 or torch.isinf(e)
                for i, e in enumerate(expect)))
        self.assertEqual(expect, actual, equal_nan=False)

    @onlyCUDA
    @ops(foreach_reduce_op_db)
    def test_foreach_reduce_large_input(self, device, dtype, op):
        # test inputs larger than kChunkSize = 65536
        ord, N = 2, 65536 * 2
        disable_fastpath = True
        if ord in (1, 2) and dtype in floating_types_and(torch.half, torch.bfloat16):
            disable_fastpath = False
        inputs = ([make_tensor((N,), dtype=dtype, device=device, noncontiguous=False)],)
        wrapped_op, ref, _, _ = self._get_funcs(op)
        self.assertEqual(
            ref(inputs, ord=ord),
            wrapped_op(inputs, self.is_cuda, not disable_fastpath, ord=ord, zero_size=False),
        )

    @onlyCUDA
    @ops(
        foreach_unary_op_db + foreach_binary_op_db + foreach_pointwise_op_db + foreach_other_op_db,
        dtypes=(torch.float,),
    )
    def test_inplace_foreach_leaf_check_and_grad_fn(self, device, dtype, op):
        inplace_op = op.inplace_variant
        if inplace_op is None:
            self.skipTest("no in-place op available")

        sample = next(iter(op.sample_inputs(dtype=dtype, device=device, num_input_tensors=[2], same_size=True)))
        sample.input[0].requires_grad_(True)
        with self.assertRaisesRegex(RuntimeError, "a leaf Variable that requires grad"):
            inplace_op(sample.input, *sample.args)
        sample.input[1].requires_grad_(True)
        with self.assertRaisesRegex(RuntimeError, "a leaf Variable that requires grad"):
            inplace_op(sample.input, *sample.args)

        _tensors = [t.clone().detach().requires_grad_(i == 0) for i, t in enumerate(sample.input)]
        tensors = [t.clone() for t in _tensors]
        inplace_op(tensors, *sample.args)
        self.assertIsNotNone(tensors[0].grad_fn)
        self.assertIsNone(tensors[1].grad_fn)

    @onlyCUDA
    @ops(
        filter(
            lambda op: op.supports_out,
            foreach_unary_op_db + foreach_binary_op_db + foreach_pointwise_op_db + foreach_other_op_db,
        ),
        dtypes=(torch.float,),
    )
    def test_outplace_with_invalid_grads(self, device, dtype, op):
        func, *_ = self._get_funcs(op)
        sample = next(iter(op.sample_inputs(dtype=dtype, device=device, requires_grad=True, num_input_tensors=[2], same_size=True)))
        self.assertTrue(all(t.requires_grad for t in sample.input))
        (out1, out2) = func([sample.input, *sample.args], is_cuda=False, expect_fastpath=False, **sample.kwargs)
        out1.backward(torch.ones_like(out1))
        self.assertIsNotNone(sample.input[0].grad)
        self.assertIsNone(sample.input[1].grad)

    @ops(
        filter(
            lambda op: op.backward_requires_result,
            foreach_unary_op_db + foreach_binary_op_db + foreach_pointwise_op_db + foreach_other_op_db,
        ),
        dtypes=(torch.float32,),
    )
    def test_lifetime_of_grad_fn_when_result_is_saved(self, device, dtype, op):

        def get_ref(func, sample):
            class Foo:
                pass

            out = func((sample.input, *sample.args), is_cuda=False, expect_fastpath=False, **sample.kwargs)
            foo = Foo()
            meta_dict = out[0].grad_fn.metadata
            meta_dict[0] = foo
            ref = weakref.ref(foo)
            return out, ref

        def _test(func, sample):
            out, ref = get_ref(func, sample)
            self.assertIsNotNone(ref())
            del out
            self.assertIsNone(ref())

        func = self._get_funcs(op)[0]
        for sample in op.sample_inputs(device, dtype, requires_grad=True, num_input_tensors=[1]):
            for key in ("is_fastpath", "disable_fastpath"):
                if key in sample.kwargs:
                    del sample.kwargs[key]
            # note: `_foreach_pow.Scalar` and `_foreach_pow.ScalarList` don't depend on `result`
            # see: https://github.com/pytorch/pytorch/blob/5403c777/tools/autograd/derivatives.yaml#L3048-L3049
            if op.name == "_foreach_pow":
                if (
                    (isinstance(sample.args[0], list) and isinstance(sample.args[0][0], Number))
                    or (isinstance(sample.args[0], Number) and not isinstance(sample.args[0], float))
                ):
                    continue
                if isinstance(sample.args[0], float):
                    new_args = (sample.input,)
                    sample.input = sample.args[0]
                    sample.args = new_args
            _test(func, sample)

    @unittest.skipIf(not TEST_MULTIGPU, "multi-GPU not supported")
    def test_tensors_grouping(self):
        num_tensors_per_list = 10
        num_devices = torch.cuda.device_count()
        dtypes = (torch.float16, torch.float32, torch.float64)
        list1 = [
            torch.tensor(
                i,
                device=torch.device("cuda", random.randint(0, num_devices - 1)),
                dtype=dtypes[random.randint(0, 2)],
            ) for i in range(num_tensors_per_list)
        ]
        list2 = [None for _ in list1]
        list3 = [torch.rand_like(t) for t in list1]
        nested_tensorlists = [list1, list2, list3]
        grouped_tensors = torch.utils._foreach_utils._group_tensors_by_device_and_dtype(nested_tensorlists, with_indices=True)
        num_tensors_seen = 0
        for (device, dtype), ([l1, l2, l3], indices) in grouped_tensors.items():
            for t in itertools.chain(l1, l3):
                self.assertEqual(t.device, device)
                self.assertEqual(t.dtype, dtype)
                num_tensors_seen += 1
            self.assertEqual(len(l1), len(l2))
            self.assertTrue(all(p is None for p in l2))
            for i, index in enumerate(indices):
                self.assertEqual(l1[i], list1[index])
                self.assertEqual(l2[i], list2[index])
                self.assertEqual(l3[i], list3[index])
        self.assertEqual(num_tensors_seen, 2 * num_tensors_per_list)

    @onlyCUDA
    def test_0dim_tensor_overload_cpu_ok(self):
        tensors = [torch.ones((), device="cuda", dtype=torch.float32) for _ in range(2)]
        scalar_cpu_tensor = torch.tensor(4.0, device="cpu")

        # For mul and div, the scalar is allowed to be on CPU too
        actual = torch._foreach_mul(tensors, scalar_cpu_tensor)
        self.assertEqual(actual, [t.mul(scalar_cpu_tensor) for t in tensors])
        actual = torch._foreach_div(tensors, scalar_cpu_tensor)
        self.assertEqual(actual, [t.div(scalar_cpu_tensor) for t in tensors])


    @onlyCUDA
    def test_0dim_tensor_overload_exception(self):
        # check exceptions of fast path
        tensors = [make_tensor((2, 2), dtype=torch.float, device="cuda") for _ in range(2)]
        with self.assertRaisesRegex(RuntimeError, "scalar tensor expected to be on"):
            torch._foreach_add(tensors, torch.tensor(1.0, device="cpu"), alpha=1.0)

        tensors = [make_tensor((2, 2), dtype=torch.float, device=d) for d in ("cpu", "cuda")]
        with self.assertRaisesRegex(RuntimeError, "scalar tensor expected to be 0 dim but"):
            torch._foreach_mul(tensors, torch.tensor([1.0, 1.0], device="cuda"))
        with self.assertRaisesRegex(RuntimeError, "scalar tensor expected to be 0 dim but"):
            torch._foreach_add(tensors, torch.tensor([1.0, 1.0], device="cuda"))

    @onlyCUDA
    @ops(filter(lambda op: op.name == "_foreach_copy", foreach_binary_op_db))
    def test_foreach_copy_with_multi_device_inputs(self, device, dtype, op):
        foreach_copy_ = op.inplace_variant
        copy_ = op.ref_inplace
        for non_blocking in (False, True):
            for sample in op.sample_inputs(device, dtype, noncontiguous=False):
                with torch.no_grad():
                    ref_input = [t.clone().detach() for t in sample.input]
                foreach_copy_(sample.input, sample.args[0], non_blocking)
                for t, s in zip(ref_input, sample.args[0]):
                    copy_(t, s, non_blocking)
                self.assertEqual(sample.input, ref_input)
                if torch.cuda.device_count() > 1:
                    device = torch.device("cuda", 1)
                    rhs_tensors = [t.to(device) for t in sample.args[0]]
                    foreach_copy_(sample.input, rhs_tensors, non_blocking)
                    for t, s in zip(ref_input, rhs_tensors):
                        copy_(t, s, non_blocking)
                    self.assertEqual(ref_input, sample.input)

    # Test reverse-mode & forward-mode AD if supported.
    @onlyCUDA
    @ops(
        foreach_unary_op_db + foreach_binary_op_db + foreach_pointwise_op_db + foreach_reduce_op_db + foreach_other_op_db,
        dtypes=OpDTypes.supported,
        allowed_dtypes=(torch.float64, torch.complex128),
    )
    @parametrize("inplace", (False, True), name_fn=lambda x: "inplace" if x else "outplace")
    def test_autodiff(self, device, dtype, op, inplace):
        if not (op.supports_autograd or op.supports_forward_ad):
            self.skipTest("neither reverse mode nor forward mode supported")
        if (not inplace) and not op.supports_out:
            self.skipTest("out-of-place not implemented")
        if inplace and op.has_no_in_place:
            self.skipTest("in-place not implemented")

        # note(crcrpar): without this, some unary functions fail, unlike inplace and/or complex.
        if (not inplace) and dtype == torch.float64 and op.name in (
            "_foreach_acos", "_foreach_asin", "_foreach_log10", "_foreach_log1p", "_foreach_log2",
            "_foreach_log", "_foreach_pow", "_foreach_sqrt",
        ):
            value_range = {"low": 0.5, "high": 1.0}
        else:
            value_range = {}
        for sample in op.sample_inputs(
            device, dtype, requires_grad=True, num_input_tensors=[5], **value_range,
        ):
            # Skip `_foreach_pow.ScalarAndTensor(Scalar, Tensor[])`
            if op.name == "_foreach_pow" and isinstance(sample.input, Number):
                continue

            func = None
            if inplace:
                # Call `clone` to avoid inplace modifications likewise
                # `torch.testing._internal.common_utils.TestGradients._get_safe_inplace`
                def inplace_func(*tensorlist):
                    kwargs = {"alpha": sample.kwargs["alpha"]} if "alpha" in sample.kwargs else {}
                    op.inplace_variant(tuple(t.clone() for t in tensorlist), *sample.args, **kwargs)
                    return tensorlist
                func = inplace_func
            else:
                def outplace_func(*tensorlist):
                    kwargs = {"alpha": sample.kwargs["alpha"]} if "alpha" in sample.kwargs else {}
                    return op.method_variant(tensorlist, *sample.args, **kwargs)
                func = outplace_func

            working_sample, err_msg_pattern = check_autodiff_sample(op, sample, dtype, inplace)

            def call_gradcheck():
                gradcheck(
                    func,
                    sample.input,
                    raise_exception=True,
                    check_forward_ad=op.supports_forward_ad,
                    check_batched_forward_grad=False,
                    check_backward_ad=op.supports_autograd,
                    check_batched_grad=False,
                )

            if not working_sample:
                if not err_msg_pattern:
                    # lhs of float64 and rhs of complex.
                    continue
                with self.assertRaisesRegex(RuntimeError, re.escape(err_msg_pattern)):
                    call_gradcheck()
                continue
            call_gradcheck()

            # Test per-tensor `grad_fn` behavior.
            if inplace and op.supports_inplace_autograd:
                # per-tensor `grad_fn` check.
                hook_buffer = []

                def get_grad_fn_hook(i):

                    def hook(grad_inputs, grad_outputs) -> None:
                        hook_buffer.append(i)

                    return hook

                _inputs = [t.clone().detach().requires_grad_() for t in sample.input]
                inputs = [t.clone() for t in _inputs]
                kwargs = {"alpha": sample.kwargs["alpha"]} if "alpha" in sample.kwargs else {}
                op.inplace_variant(inputs, *sample.args, **kwargs)

                self.assertEqual(len({t.grad_fn for t in inputs}), len(inputs))

                for i, t in enumerate(inputs):
                    t.grad_fn.register_hook(get_grad_fn_hook(i))

                torch.autograd.grad(
                    inputs[0],
                    inputs=(_inputs[0],),
                    grad_outputs=(torch.rand_like(inputs[0]),),
                    retain_graph=True,
                )
                self.assertEqual(hook_buffer, [0])
                hook_buffer.clear()

                # tensors have different shapes.
                sum_of_cloned_tensors = torch.cat([t.view(-1) for t in inputs]).sum()
                grad_output = torch.rand_like(sum_of_cloned_tensors)
                torch.autograd.grad(
                    sum_of_cloned_tensors,
                    inputs=tuple(_inputs),
                    grad_outputs=(grad_output,),
                    retain_graph=False,
                )
                self.assertEqual(hook_buffer, list(reversed(range(len(inputs)))))


# TODO(crcrpar): Hide this inside torch/testing/_internal.
# would end up adding another layer to `foreach_inputs_sample_func.__call__`
# so that we can use this function as something like the first argument of `filter` function.
# Even after moving this function to testing, I personally think it'd be better to check the error message.
def check_autodiff_sample(op, sample, dtype, is_inplace):
    if op.name == "_foreach_abs" and is_inplace and dtype == torch.complex128:
        return False, "In-place abs is not supported for complex tensors."
    if (
        op.name == "_foreach_sub"
        and (
            (isinstance(sample.args[0], list) and any(isinstance(a, bool) for a in sample.args[0]))
            or isinstance(sample.args[0], bool)
        )
    ):
        return False, _BOOL_SUB_ERR_MSG
    if op.name == "_foreach_norm" and (not is_inplace):
        return (
            False,
            "Trying to set a forward gradient that has a different size than that of the original Tensor, "
            "this is not supported. Tensor is of size [] while the given forward gradient is of size [1, 1]."
        )
    rhs_arg_has_complex_number = sample.args and ((
        isinstance(sample.args[0], list)
        and any(isinstance(a, complex) for a in sample.args[0])
    ) or (
        isinstance(sample.args[0], complex)
    ))
    if rhs_arg_has_complex_number and dtype == torch.float64:
        if op.name in ("_foreach_clamp_max", "_foreach_clamp_min", "_foreach_maximum", "_foreach_minimum"):
            return False, "clamp is not supported for complex types"
        if not is_inplace:
            return False, ""
        else:
            if op.name == "_foreach_pow":
                return False, "Found dtype Double but expected ComplexDouble"
            if op.name in ("_foreach_add", "_foreach_sub", "_foreach_mul", "_foreach_div"):
                return False, "result type ComplexDouble can't be cast to the desired output type Double"
    return True, ""


instantiate_device_type_tests(TestForeach, globals())


if __name__ == "__main__":
    run_tests()