pytorch/torch/_subclasses/fake_utils.py

# mypy: ignore-errors

import functools
import warnings
from typing import Any, Callable, List, Union

import torch
import torch.utils._pytree as pytree
from torch._ops import OpOverload
from torch._subclasses.fake_tensor import (
    FakeTensor,
    FakeTensorMode,
    MetadataMismatchError,
    tree_flatten_only,
    UnsupportedFakeTensorException,
)
from torch.utils._python_dispatch import TorchDispatchMode


aten = torch._ops.ops.aten


def outputs_alias_inputs(outputs, inputs):
    input_storages = {
        inp._typed_storage()._cdata
        for inp in tree_flatten_only(torch.Tensor, inputs)
        if torch._C._has_storage(inp)
    }
    return any(
        torch._C._has_storage(out) and out._typed_storage()._cdata in input_storages
        for out in tree_flatten_only(torch.Tensor, outputs)
    )


def outputs_are_inputs(outputs, inputs):
    input_ids = {id(inp) for inp in tree_flatten_only(torch.Tensor, inputs)}
    return any(id(out) in input_ids for out in tree_flatten_only(torch.Tensor, outputs))


def output_alias_each_other(outputs):
    storages = set()
    for out in tree_flatten_only(torch.Tensor, outputs):
        if not torch._C._has_storage(out):
            continue
        stor = out._typed_storage()._cdata
        if stor in storages:
            return True
        storages.add(stor)
    return False


def _check_alias_info(context, real_out, real_in, fake_out, fake_in):
    r_aliasing = outputs_alias_inputs(real_out, real_in)
    f_aliasing = outputs_alias_inputs(fake_out, fake_in)
    if r_aliasing != f_aliasing:
        raise MetadataMismatchError(
            f"{context} mismatch in outputs_alias_inputs check {f_aliasing} != {r_aliasing}"
        )

    r_identity_eq = outputs_are_inputs(real_out, real_in)
    f_identity_eq = outputs_are_inputs(fake_out, fake_in)
    if r_identity_eq != f_identity_eq:
        raise MetadataMismatchError(
            f"{context} mismatch in outputs_are_inputs check {f_identity_eq} != {r_identity_eq}"
        )

    r_output_alias_each_other = output_alias_each_other(real_out)
    f_output_alias_each_other = output_alias_each_other(fake_out)
    if r_output_alias_each_other != f_output_alias_each_other:
        raise MetadataMismatchError(
            f"{context} mismatch in outputs_alias_each_other check "
            f"{f_output_alias_each_other} != {r_output_alias_each_other}"
        )


def is_sdpa_error(func, idx, e):
    if (
        (
            func is aten._scaled_dot_product_flash_attention.default
            or func is aten._flash_attention_forward.default
        )
        and idx in (6, 7)
        and "Devices" in repr(e)
    ):
        return True
    if (
        (
            func is aten._scaled_dot_product_efficient_attention.default
            or func is aten._efficient_attention_forward.default
        )
        and idx in (2, 3)
        and "Devices" in repr(e)
    ):
        return True
    if (
        func is aten._scaled_dot_product_cudnn_attention.default
        and idx in (6, 7)
        and "Devices" in repr(e)
    ):
        return True
    return False


def try_convert_fake_to_real(
    ten_list: List[Union[FakeTensor, Any]]
) -> List[Union[FakeTensor, torch.Tensor, Any]]:
    """
    Attempt to convert fake tensors to a corresponding real tensor with the correct underlying storage by looking up
    the FakeTensorMode meta to real storage mapping. On failure to find the storage mapping, the FakeTensor will
    remain in the list.

    Note: this is not currently optimized (makes copies of the meta converter internal dictionaries)
    """

    fake_tensor = next(
        (item for item in ten_list if isinstance(item, FakeTensor)), None
    )
    if fake_tensor is None:
        return ten_list

    fake_mode = fake_tensor.fake_mode
    meta_converter = fake_mode.fake_tensor_converter.meta_converter
    desc = meta_converter.describer

    storage_to_key = {v: k for k, v in meta_converter.storage_memo.items()}
    key_to_real_storage = {v: k for k, v in desc.lookup_storage.items()}
    out = []
    for t in ten_list:
        if not isinstance(t, FakeTensor) or not t.layout == torch.strided:
            out.append(t)
            continue

        key = storage_to_key.get(t.untyped_storage())
        real_storage = None if key is None else key_to_real_storage.get(key)
        if real_storage is None:
            out.append(t)
            continue

        unhinted = False

        def map_symint(s):
            nonlocal unhinted
            if not isinstance(s, torch.SymInt):
                return s
            unhinted = unhinted if not unhinted else s.node.has_hint()
            return s.node.hint

        stor_offset = map_symint(t.storage_offset())
        size = [map_symint(s) for s in t.shape]
        stride = [map_symint(s) for s in t.stride()]

        if unhinted:
            out.append(t)
            continue

        new_tensor = torch.empty(
            [],
            dtype=t.dtype,
            device=t.device,
        )
        new_tensor.set_(
            real_storage,
            storage_offset=stor_offset,
            size=size,
            stride=stride,
        )
        out.append(new_tensor.clone())

    return out


def _check_fake_real_tensors(
    real_out: torch.Tensor,
    fake_out: FakeTensor,
    context="",
    sizes=True,
    strides=False,
    storage_offset=True,
    requires_grad=True,
):
    if requires_grad:
        if real_out.requires_grad != fake_out.requires_grad:
            raise MetadataMismatchError(
                f"{context} mismatched requires_grad-ness of outputs. "
                f"This usually means that you have added autograd support "
                f"for your operator at a dispatch key other than Autograd, "
                f"which will lead to problems"
            )

    if torch._C._has_storage(real_out):
        r_offset = real_out.storage_offset()
        f_offset = fake_out.storage_offset()
        if r_offset != f_offset:
            raise MetadataMismatchError(f"{context} mismatched storage offset")

    torch._prims.utils.compare_tensor_meta(
        real_out,
        fake_out,
        check_sizes=sizes,
        check_strides=strides,
        allow_rhs_unbacked=True,
    )


class CrossRefFakeMode(TorchDispatchMode):
    def __init__(
        self,
        ignore_op_fn: Union[Callable[[OpOverload], bool], None] = None,
        *,
        check_strides=True,
        check_aliasing=True,
        only_check_ops_with_meta=True,
    ):
        super().__init__()
        self.ignore_op_fn = (
            ignore_op_fn if ignore_op_fn is not None else lambda fn: False
        )
        self.check_strides = check_strides
        self.check_aliasing = check_aliasing
        self.only_check_ops_with_meta = only_check_ops_with_meta

    def __torch_dispatch__(self, func, types, args=(), kwargs=None):
        kwargs = kwargs or {}

        fake_r = None

        # empty_like excluded for now due to sparse complex
        # aten._to_dense.default this one is getting called with csc
        if (
            func
            not in (
                aten.lift_fresh.default,
                aten.lift_fresh_copy.default,
                aten.set_.source_Storage_storage_offset,
            )
            and not self.ignore_op_fn(func)
            and (
                not self.only_check_ops_with_meta
                or torch._subclasses.fake_impls.has_meta(func)
            )
            and torch.Tag.dynamic_output_shape not in func.tags
            and torch.Tag.inplace_view not in func.tags
            and torch.Tag.data_dependent_output not in func.tags
        ):
            # Do not import symbolic_shapes at the top of the module as it imports sympy and that's slow
            from torch.fx.experimental.symbolic_shapes import ShapeEnv

            try:
                # TODO: enable_python_dispatcher() here
                with FakeTensorMode(shape_env=ShapeEnv()) as fake_mode:
                    fake_args, fake_kwargs = pytree.tree_map_only(
                        torch.Tensor,
                        functools.partial(fake_mode.from_tensor, static_shapes=True),
                        (args, kwargs),
                    )
                    with warnings.catch_warnings():
                        fake_r = func(*fake_args, **fake_kwargs)
            except UnsupportedFakeTensorException:
                pass

        context = (
            f"When comparing the output of {func} on FakeTensor and concrete Tensors, "
            f"found"
        )
        r = func(*args, **kwargs)
        if fake_r is not None:
            r_flat = pytree.tree_leaves(r)
            f_flat = pytree.tree_leaves(fake_r)
            assert len(f_flat) == len(
                r_flat
            ), f"{context} mismatch in number of returns {len(f_flat)} != {len(r_flat)}"

            if self.check_aliasing:
                _check_alias_info(
                    context, r, (args, kwargs), fake_r, (fake_args, fake_kwargs)
                )

            for idx, (r_out, f_out) in enumerate(
                zip(pytree.tree_leaves(r), pytree.tree_leaves(fake_r))
            ):
                r_is_ten = isinstance(r_out, torch.Tensor)
                assert r_is_ten == isinstance(
                    f_out, torch.Tensor
                ), f"{context} mismatched number of tensor outputs"
                if r_is_ten:
                    try:
                        _check_fake_real_tensors(
                            r_out,
                            f_out,
                            sizes=True,
                            strides=self.check_strides,
                            storage_offset=True,
                            requires_grad=True,
                        )
                    except Exception as e:
                        if is_sdpa_error(func, idx, e):
                            continue
                        error_message = (
                            f"{context} mismatched tensor metadata: {e}"
                            if len(r_flat) == 1
                            else f"{context} mismatched tensor metadata for output[{idx}]: {e}"
                        )
                        raise MetadataMismatchError(error_message) from e
        return r