arg_validator.py

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.

from collections import defaultdict
from typing import Any, Dict, Optional, Sequence, Tuple

import torch
from executorch.exir.dialects.edge._ops import EdgeDialectFunctionSchema, EdgeOpOverload
from executorch.exir.emit._emitter import _Argument, _Target
from executorch.exir.error import ExportError, InternalError
from torch._ops import HigherOrderOperator


class RunHigherOrderOperatorError(Exception):
    """
    Raised when an we try to run delegate or other HigherOrderOperator in a graph module.
    E.g., %executorch_call_delegate : [#users=1] = call_function[
        target=torch.ops.higher_order.executorch_call_delegate](args = (%lowered_module_0, %arg0_1), kwargs = {})
    """

    def __init__(self, message: str) -> None:
        super().__init__(message)


# pyre-ignore[13]: Attribute `node` is never initialized.
class EdgeOpArgValidator(torch.fx.Interpreter):
    """
    Validate whether all the Tensor arguments passed to an operator are valid in terms of allowed dtype.
    Expecting all the operators are EdgeOpOverload which contains the allowed dtype information.
    Violating operators are being kept in self.violating_ops
    """

    node: torch.fx.Node

    def __init__(self, graph_module: torch.fx.GraphModule) -> None:
        super().__init__(graph_module)
        self.violating_ops: Dict[
            EdgeOpOverload, Tuple[Dict[str, Optional[torch.dtype]], torch.fx.Node]
        ] = defaultdict(dict)

    def run_node(self, n: torch.fx.Node) -> None:
        self.node = n
        try:
            ret = super().run_node(n)
        except Exception as e:
            if isinstance(e, (InternalError, ExportError, RunHigherOrderOperatorError)):
                raise e
            else:
                raise InternalError(str(e)) from e
        return ret

    def _get_kernel_arg(self, schema_arg, schema_arg_idx, args, kwargs):
        if schema_arg.name in kwargs:
            kernel_arg = kwargs[schema_arg.name]
        elif not schema_arg.kwarg_only and schema_arg_idx < len(args):
            kernel_arg = args[schema_arg_idx]
        else:
            kernel_arg = schema_arg.default_value

        return kernel_arg

    def call_function(  # noqa: C901  # pyre-fixme[14]
        self, target: _Target, args: Tuple[_Argument, ...], kwargs: Dict[str, _Argument]
    ) -> Any:
        """
        Go through all the node.target and validate their Tensor arguments are having the allowed dtypes.
        """
        if not isinstance(target, EdgeOpOverload) or not isinstance(
            target._schema, EdgeDialectFunctionSchema
        ):
            if isinstance(target, HigherOrderOperator):
                raise RunHigherOrderOperatorError("Can't run delegate")
            return super().call_function(target, args, kwargs)  # pyre-fixme[6]

        # TODO(gasoonjia): Update Optional[torch.dtype] to a concrete class to support mixed dtypes in tensorlist.
        tensor_arg_types: Dict[str, Optional[torch.dtype]] = {}
        for i, schema_arg in enumerate(target._schema.arguments):
            if (
                isinstance(schema_arg.type, torch.TensorType)
                or schema_arg.type == torch.OptionalType.ofTensor()
            ):
                kernel_arg = self._get_kernel_arg(schema_arg, i, args, kwargs)
                if not isinstance(kernel_arg, torch.Tensor):
                    continue
                tensor_arg_types[schema_arg.name] = kernel_arg.dtype
            elif schema_arg.type == torch.ListType.ofTensors():
                kernel_arg = self._get_kernel_arg(schema_arg, i, args, kwargs)
                if not isinstance(kernel_arg, list) or not all(
                    isinstance(kernel_arg[i], torch.Tensor)
                    for i in range(len(kernel_arg))
                ):
                    continue
                if len(kernel_arg):
                    tensor_arg_types[schema_arg.name] = kernel_arg[0].dtype
                else:
                    # If kernel_arg is an empty list, treat its type as None.
                    # FunctionDtypeConstraint.validate will take None as any legal dtype.
                    tensor_arg_types[schema_arg.name] = None

        ret_index = 0
        kernel_rets = self.node.meta["val"]
        ret_iter = iter(
            kernel_rets if isinstance(kernel_rets, Sequence) else [kernel_rets]
        )
        for schema_ret in target._schema.returns:
            name = schema_ret.name if schema_ret.name else f"__ret_{ret_index}"
            kernel_ret = next(ret_iter)
            if isinstance(schema_ret.type, torch.TensorType):
                if isinstance(kernel_ret, torch.Tensor):
                    tensor_arg_types[name] = kernel_ret.dtype
                    ret_index += 1
                # Exceptionally rarely (basically only backwards ops) you might see an OptionalTensor returned.
                # The schema of these ops though is typically -> (Tensor, Tensor ...). So the actual type
                # returned in cpp is empty/undefined tensor. There is no analogy to this in python so it
                # gets crudely mapped to None. To properly fix this core pytorch would have to change the
                # schema to (Tensor?, ...) which is just never going to happen. So we have to handle this case
                # here in the verifier and in memory planning as well.
                elif kernel_ret is None:
                    tensor_arg_types[name] = schema_ret.default_value
                    ret_index += 1
                else:
                    raise InternalError(
                        f"encountered return with type Tensor but value wasnt a tensor or None. schema:{target._schema}, output:{ret_index}"
                    )
            elif schema_ret.type == torch.ListType.ofTensors() and all(
                isinstance(kernel_ret[i], torch.Tensor) for i in range(len(kernel_ret))
            ):
                if len(kernel_ret):
                    tensor_arg_types[name] = kernel_ret[0].dtype
                else:
                    tensor_arg_types[name] = None
                ret_index += 1

        valid = target._schema.dtype_constraint.validate(tensor_arg_types)
        if not valid:
            self.violating_ops[target] = (tensor_arg_types, self.node)
        return super().call_function(target, args, kwargs)  # pyre-fixme[6]