import contextlib import functools import inspect import itertools import logging import math import operator import types from collections import defaultdict, OrderedDict from typing import Dict, List import torch from torch import sym_float, sym_int from .. import config, polyfill, variables from ..exc import ( AttributeMutationError, unimplemented, Unsupported, UserError, UserErrorType, ) from ..guards import GuardBuilder, install_guard from ..replay_record import DummyModule from ..source import AttrSource, GetItemSource, is_constant_source, TypeSource from ..utils import ( build_checkpoint_variable, check_constant_args, check_numpy_ndarray_args, check_unspec_python_args, extract_fake_example_value, get_fake_value, guard_if_dyn, is_utils_checkpoint, istype, numpy_operator_wrapper, proxy_args_kwargs, tensortype_to_dtype, ) from .base import MutableLocal, typestr, VariableTracker from .constant import ConstantVariable from .ctx_manager import EventVariable, StreamVariable from .dicts import ConstDictVariable, DefaultDictVariable, SetVariable from .lists import ( BaseListVariable, ListIteratorVariable, ListVariable, SizeVariable, TupleIteratorVariable, TupleVariable, ) from .tensor import FakeItemVariable, SymNodeVariable, UnspecializedPythonVariable from .user_defined import UserDefinedVariable log = logging.getLogger(__name__) IN_PLACE_DESUGARING_MAP = { operator.iadd: operator.add, operator.isub: operator.sub, operator.imul: operator.mul, operator.ifloordiv: operator.floordiv, operator.itruediv: operator.truediv, operator.imod: operator.mod, operator.imatmul: operator.imatmul, operator.ilshift: operator.lshift, operator.irshift: operator.rshift, operator.ipow: operator.pow, operator.iand: operator.and_, operator.ior: operator.or_, operator.ixor: operator.xor, } def _polyfill_call_impl(name): """Create a BuiltinVariable.call_{name} method that inlines through polyfill.{name}""" def call_fn(self, tx, *args, **kwargs): return tx.inline_user_function_return( variables.UserFunctionVariable(fn), args, kwargs ) fn = getattr(polyfill, name) call_fn.__name__ = f"call_{name}" return call_fn class BuiltinVariable(VariableTracker): @staticmethod @functools.lru_cache(None) def _constant_fold_functions(): fns = { abs, all, any, bool, callable, chr, divmod, float, int, len, max, min, ord, pow, repr, round, str, str.format, sum, type, operator.pos, operator.neg, operator.not_, operator.invert, operator.pow, operator.mul, operator.matmul, operator.floordiv, operator.truediv, operator.mod, operator.add, operator.sub, operator.getitem, operator.lshift, operator.rshift, operator.and_, operator.or_, operator.xor, operator.ipow, operator.imul, operator.imatmul, operator.ifloordiv, operator.itruediv, operator.imod, operator.iadd, operator.isub, operator.ilshift, operator.irshift, operator.iand, operator.ixor, operator.ior, operator.index, } fns.update(x for x in math.__dict__.values() if isinstance(x, type(math.sqrt))) return fns def can_constant_fold_through(self): return self.fn in self._constant_fold_functions() @staticmethod @functools.lru_cache(None) def _fx_graph_functions(): fns = { operator.pos, operator.neg, operator.not_, operator.invert, operator.pow, operator.mul, operator.matmul, operator.floordiv, operator.truediv, operator.mod, operator.add, operator.lt, operator.gt, operator.ge, operator.le, operator.ne, operator.eq, operator.sub, operator.getitem, operator.lshift, operator.rshift, operator.and_, operator.or_, operator.xor, operator.ipow, operator.imul, operator.imatmul, operator.ifloordiv, operator.itruediv, operator.imod, operator.iadd, operator.isub, operator.ilshift, operator.irshift, operator.iand, operator.ixor, operator.ior, } return fns @staticmethod @functools.lru_cache(None) def _binops(): # function -> ([forward name, reverse name, in-place name], in-place op) fns = { operator.add: (["__add__", "__radd__", "__iadd__"], operator.iadd), operator.sub: (["__sub__", "__rsub__", "__isub__"], operator.isub), operator.mul: (["__mul__", "__rmul__", "__imul__"], operator.imul), operator.truediv: ( ["__truediv__", "__rtruediv__", "__itruediv__"], operator.itruediv, ), operator.floordiv: ( ["__floordiv__", "__rfloordiv__", "__ifloordiv__"], operator.ifloordiv, ), operator.mod: (["__mod__", "__rmod__", "__imod__"], operator.imod), pow: (["__pow__", "__rpow__", "__ipow__"], operator.ipow), operator.pow: (["__pow__", "__rpow__", "__ipow__"], operator.ipow), operator.lshift: ( ["__lshift__", "__rlshift__", "__ilshift__"], operator.ilshift, ), operator.rshift: ( ["__rshift__", "__rrshift__", "__irshift__"], operator.irshift, ), # NB: The follow binary operators are not supported for now, since the # corresponding magic methods aren't defined on SymInt / SymFloat: # operator.matmul # divmod # operator.and_ # operator.or_ # operator.xor } return fns @staticmethod @functools.lru_cache(None) def _binop_handlers(): # Multiple dispatch mechanism defining custom binop behavior for certain type # combinations. Handlers are attempted in order, and will be used if the type checks # match. They are expected to have the signature: # fn(tx, arg0: VariableTracker, arg1: VariableTracker, options) -> VariableTracker # Override table contains: op_fn -> [list of handlers] op_handlers = {} for ( op, (magic_method_names, in_place_op), ) in BuiltinVariable._binops().items(): op_handlers[op] = [] op_handlers[in_place_op] = [] forward_name, reverse_name, inplace_name = magic_method_names # User-defined args (highest precedence) def user_defined_handler( tx, a, b, options, forward_name=forward_name, reverse_name=reverse_name, ): # Manually handle reversing logic if needed (e.g. call __radd__) # TODO: If we expand this to handle tensor args, we need to manually # handle cases like this: # # class A(int): # def __radd__(self, other): # print("woof") # torch.randn(3) + A(3) # # In this example, A.__radd__() is not called -> nothing is printed, because # Tensor.__add__ only does a subtype test against int, ignoring the subclass. # To be fully correct, we should not call A.__radd__() here, and there may be # other cases to reason about and add exceptions for. if isinstance(a, UserDefinedVariable): return a.call_method(tx, forward_name, [b], {}) else: return b.call_method(tx, reverse_name, [a], {}) op_handlers[op].append( ((UserDefinedVariable, VariableTracker), user_defined_handler) ) op_handlers[op].append( ((VariableTracker, UserDefinedVariable), user_defined_handler) ) def user_defined_inplace_handler( tx, a, b, options, forward_name=inplace_name ): return a.call_method(tx, forward_name, [b], {}) op_handlers[in_place_op].append( ((UserDefinedVariable, VariableTracker), user_defined_inplace_handler) ) op_handlers[in_place_op].append( ((VariableTracker, UserDefinedVariable), user_defined_inplace_handler) ) # Dynamic shape args def dynamic_handler(tx, a, b, options, fn=op): from .builder import wrap_fx_proxy return wrap_fx_proxy( tx, tx.output.create_proxy( "call_function", fn, *proxy_args_kwargs([a, b], {}) ), **options, ) op_handlers[op].append( ((SymNodeVariable, VariableTracker), dynamic_handler) ) op_handlers[op].append( ((VariableTracker, SymNodeVariable), dynamic_handler) ) # NB: Prefer out-of-place op when calling in-place op to generate valid graph op_handlers[in_place_op].append( ((SymNodeVariable, VariableTracker), dynamic_handler) ) op_handlers[in_place_op].append( ((VariableTracker, SymNodeVariable), dynamic_handler) ) # Special cases - lower precedence but still prefer these over constant folding # List-like addition (e.g. [1, 2] + [3, 4]) def tuple_add_handler(tx, a, b, options): return TupleVariable(a.items + list(b.unpack_var_sequence(tx)), **options) def size_add_handler(tx, a, b, options): return SizeVariable(a.items + list(b.unpack_var_sequence(tx)), **options) list_like_addition_handlers = [ # NB: Prefer the tuple-specific logic over base logic because of # some SizeVariable weirdness. Specifically, the tuple-specific logic # drops the subclass type (e.g. SizeVariable) and returns TupleVariables. ( (SizeVariable, SizeVariable), size_add_handler, ), ( (TupleVariable, TupleVariable), tuple_add_handler, ), ( (TupleVariable, ConstantVariable), tuple_add_handler, ), ( (ConstantVariable, TupleVariable), lambda tx, a, b, options: TupleVariable( list(a.unpack_var_sequence(tx)) + b.items, **options ), ), ( (BaseListVariable, BaseListVariable), lambda tx, a, b, options: type(a)(a.items + b.items, **options), ), ] op_handlers[operator.add].extend(list_like_addition_handlers) def list_iadd_handler(tx, a, b, options): if not a.mutable_local or not b.has_unpack_var_sequence(tx): # Handler doesn't apply return None return tx.replace_all( a, ListVariable( list(a.items) + list(b.unpack_var_sequence(tx)), **options, ), ) list_like_iadd_handlers = [ ( (ListVariable, VariableTracker), list_iadd_handler, ), ( (TupleVariable, TupleVariable), tuple_add_handler, ), ( (TupleVariable, ConstantVariable), tuple_add_handler, ), ] op_handlers[operator.iadd].extend(list_like_iadd_handlers) # List-like expansion (e.g. [1, 2, 3] * 3) def expand_list_like(tx, lst, const, options): return lst.__class__( items=lst.items * const.as_python_constant(), mutable_local=MutableLocal(), **options, ) list_like_expansion_handlers = [ ((ListVariable, ConstantVariable), expand_list_like), ((TupleVariable, ConstantVariable), expand_list_like), ( (ConstantVariable, ListVariable), lambda tx, a, b, options: expand_list_like(tx, b, a, options), ), ( (ConstantVariable, TupleVariable), lambda tx, a, b, options: expand_list_like(tx, b, a, options), ), ] op_handlers[operator.mul].extend(list_like_expansion_handlers) return op_handlers @staticmethod def _find_binop_handler(op, a, b): handlers = BuiltinVariable._binop_handlers() if op not in handlers: return None # Return first handler that matches the type checks for (type1, type2), handler in handlers[op]: if isinstance(a, type1) and isinstance(b, type2): return handler return None def can_insert_in_graph(self): return self.fn in self._fx_graph_functions() def __init__(self, fn, **kwargs): super().__init__(**kwargs) self.fn = fn def __str__(self): if self.fn is None: name = "None" else: name = self.fn.__name__ return f"{self.__class__.__name__}({name})" def python_type(self): return type(self.fn) def as_python_constant(self): return self.fn def as_proxy(self): DTYPE = { bool: torch.bool, int: torch.int64, float: torch.float64, } if self.fn in DTYPE: return DTYPE[self.fn] return super().as_proxy() def reconstruct(self, codegen): name = self.fn.__name__ assert self.fn.__module__ == "builtins" assert name not in codegen.tx.f_globals, "shadowed global" return [codegen.create_load_global(name, False, add=True)] def constant_args(self, *args, **kwargs): return check_constant_args(args, kwargs) def tensor_args(self, *args, **kwargs): return any( isinstance(i, variables.TensorVariable) for i in itertools.chain(args, kwargs.values()) ) and not any( isinstance(i, variables.GetAttrVariable) for i in itertools.chain(args, kwargs.values()) ) def unspec_python_args(self, *args, **kwargs): return check_unspec_python_args(args, kwargs) @staticmethod def unwrap_unspec_args_kwargs(args, kwargs): return [x.as_python_constant() for x in args], { k: v.as_python_constant() for k, v in kwargs.items() } def call_function( self, tx, args: "List[VariableTracker]", kwargs: "Dict[str, VariableTracker]" ) -> "VariableTracker": from . import UserFunctionVariable from .builder import wrap_fx_proxy, wrap_fx_proxy_cls args = [v.realize() for v in args] kwargs = {k: v.realize() for k, v in kwargs.items()} constant_args = check_constant_args(args, kwargs) tensor_args = self.tensor_args(*args, **kwargs) unspec_python_args = self.unspec_python_args(*args, **kwargs) has_constant_handler = self.can_constant_fold_through() and ( constant_args or unspec_python_args ) assert isinstance(args, (list, tuple)) assert isinstance(kwargs, dict) # args[0] is list and args[1] is unspec if self.fn is operator.getitem and not isinstance( args[0], variables.TensorVariable ): tensor_args = False if ( self.can_insert_in_graph() and tensor_args and not ( self.fn is operator.getitem and isinstance(args[0], ConstDictVariable) and isinstance(args[1], variables.TensorVariable) ) ): try: fn = self.fn if self.fn in IN_PLACE_DESUGARING_MAP and isinstance( args[0], variables.ConstantVariable ): # In-place operators like += usually mustate tensor # values, but in the edge case of immutable values they # re-bind the variable. # # The easiest way to keep the graph consistent in this # scenario is to de-sugar eagerly. fn, args = IN_PLACE_DESUGARING_MAP[self.fn], [args[0], args[1]] if self.fn is operator.getitem and isinstance(args[1], SymNodeVariable): # Standard indexing will force specialization due to # __index__. Rewrite as a regular torch op which will # trace fine fn, args = torch.select, [ args[0], variables.ConstantVariable.create(0), args[1], ] # Interaction between ndarray and tensors: # We prefer the tensor op whenever there are tensors involved if check_numpy_ndarray_args(args, kwargs) and not any( type(arg) == variables.TensorVariable for arg in args ): proxy = tx.output.create_proxy( "call_function", numpy_operator_wrapper(self.fn), *proxy_args_kwargs(args, kwargs), ) return wrap_fx_proxy_cls(variables.NumpyNdarrayVariable, tx, proxy) proxy = tx.output.create_proxy( "call_function", fn, *proxy_args_kwargs(args, kwargs), ) if any(isinstance(arg, FakeItemVariable) for arg in args): return wrap_fx_proxy_cls( FakeItemVariable, tx, proxy, ) elif self.unspec_python_args(*args, **kwargs): _args, _kwargs = self.unwrap_unspec_args_kwargs(args, kwargs) raw_value = self.fn(*_args, **_kwargs) need_unwrap = any( x.need_unwrap for x in itertools.chain(args, kwargs.values()) if isinstance(x, variables.UnspecializedPythonVariable) ) return wrap_fx_proxy_cls( UnspecializedPythonVariable, tx, proxy, raw_value=raw_value, need_unwrap=need_unwrap, ) elif all(isinstance(x, SymNodeVariable) for x in args): return SymNodeVariable.create(tx, proxy, None) else: # Work around for vision_maskrcnn due to precision difference # specialize the dividend when float divide by tensor if self.fn is operator.truediv and isinstance( args[0], variables.UnspecializedPythonVariable ): args[0] = args[0].convert_to_constant(tx) return wrap_fx_proxy(tx, proxy) except NotImplementedError: unimplemented(f"partial tensor op: {self} {args} {kwargs}") # Handle cases like int(torch.seed()) # Also handle sym_float to sym_int cases if self.fn in (int, float) and isinstance( args[0], (SymNodeVariable, variables.TensorVariable) ): if isinstance(args[0], variables.TensorVariable): item = args[0].call_method(tx, "item", [], {}) else: item = args[0] fn_ = sym_int if self.fn is int else sym_float out = wrap_fx_proxy( tx=tx, proxy=tx.output.create_proxy( "call_function", fn_, (item.as_proxy(),), {}, ), ) return out # Handle `str` on a user defined function if self.fn == str and args and isinstance(args[0], (UserFunctionVariable)): return variables.ConstantVariable.create(value=str(args[0].fn)) # Handle binary ops (e.g. __add__ / __radd__, __iadd__, etc.) # NB: Tensor args are handled above and not here if len(kwargs) == 0 and len(args) == 2: # Try to find a handler for the arg types; otherwise, fall through to constant handler binop_handler = BuiltinVariable._find_binop_handler( self.fn, args[0], args[1] ) if binop_handler: res = binop_handler(tx, args[0], args[1], {}) if res is not None: return res handler = getattr(self, f"call_{self.fn.__name__}", None) if handler: try: inspect.signature(handler).bind(tx, *args, **kwargs) except TypeError as exc: if not has_constant_handler: log.warning( "incorrect arg count %s %s and no constant handler", handler, exc, ) handler = None if handler: try: result = handler(tx, *args, **kwargs) if result is not None: return result except Unsupported as exc: if not has_constant_handler: raise # Actually, we will handle this just fine exc.remove_from_stats() if has_constant_handler: # constant fold return variables.ConstantVariable.create( self.as_python_constant()( *[x.as_python_constant() for x in args], **{k: v.as_python_constant() for k, v in kwargs.items()}, ), ) if self.fn is round: if len(args) > 0 and isinstance(args[0], SymNodeVariable): raise UserError( UserErrorType.STANDARD_LIBRARY, "Calling round() on symbolic value is not supported. " "You can use floor() to implement this functionality", case_name="dynamic_shape_round", ) return super().call_function(tx, args, kwargs) def call_method( self, tx, name, args: "List[VariableTracker]", kwargs: "Dict[str, VariableTracker]", ) -> "VariableTracker": if self.fn == dict and name == "fromkeys": return BuiltinVariable.call_custom_dict_fromkeys(tx, dict, *args, **kwargs) return super().call_method(tx, name, args, kwargs) def _call_min_max(self, tx, *args): if len(args) == 1 and args[0].has_unpack_var_sequence(tx): # expand iterable items = args[0].unpack_var_sequence(tx) return self._call_min_max_seq(tx, items) elif len(args) == 2: return self._call_min_max_binary(tx, args[0], args[1]) elif len(args) > 2: return self._call_min_max_seq(tx, args) def _call_min_max_seq(self, tx, items): assert len(items) > 0 if len(items) == 1: return items[0] return functools.reduce(functools.partial(self._call_min_max_binary, tx), items) def _call_min_max_binary(self, tx, a, b): if self.tensor_args(a, b): if not isinstance(a, variables.TensorVariable): a, b = b, a assert isinstance(a, variables.TensorVariable) # result of an item call is a scalar convert to a tensor if isinstance(a, FakeItemVariable): a = variables.TorchVariable(torch.tensor).call_function(tx, [a], {}) # Dynamic input does not get resolved, rather, gets stored as call_function if isinstance(a, SymNodeVariable) or isinstance(b, SymNodeVariable): from .builder import wrap_fx_proxy_cls return wrap_fx_proxy_cls( type(a), tx=tx, proxy=tx.output.create_proxy( "call_function", self.fn, *proxy_args_kwargs([a, b], {}), ), ) # convert min/max to torch ops if b.is_python_constant(): if isinstance(a, variables.NumpyNdarrayVariable): import numpy as np fn = variables.NumpyVariable(np.clip) else: fn = variables.TorchVariable(torch.clamp) kwargs = {"min": b} if (self.fn is max) else {"max": b} result = fn.call_function(tx, [a], kwargs) else: if isinstance(a, variables.NumpyNdarrayVariable): import numpy as np fn = {max: np.maximum, min: np.minimum}[self.fn] fn = variables.NumpyVariable(fn) else: fn = {max: torch.maximum, min: torch.minimum}[self.fn] fn = variables.TorchVariable(fn) result = fn.call_function(tx, [a, b], {}) # return unspec if both a, b are unspec or const if all( isinstance( i, ( variables.UnspecializedPythonVariable, variables.ConstantVariable, ), ) for i in [a, b] ): if any(isinstance(val, FakeItemVariable) for val in [a, b]): return variables.FakeItemVariable.from_tensor_variable(result) if b.is_python_constant(): raw_b = b.as_python_constant() else: raw_b = b.raw_value if self.fn is max: raw_res = max(a.raw_value, raw_b) else: raw_res = min(a.raw_value, raw_b) need_unwrap = any( x.need_unwrap for x in [a, b] if isinstance(x, variables.UnspecializedPythonVariable) ) return variables.UnspecializedPythonVariable.from_tensor_variable( result, raw_res, need_unwrap ) # otherwise return tensor else: return result elif isinstance(a, SymNodeVariable) or isinstance(b, SymNodeVariable): proxy = tx.output.create_proxy( "call_function", self.fn, *proxy_args_kwargs([a, b], {}) ) return SymNodeVariable.create(tx, proxy, None) call_min = _call_min_max call_max = _call_min_max def call_abs(self, tx, arg: "VariableTracker"): # Call arg.__abs__() abs_method = BuiltinVariable(getattr).call_function( tx, [arg, ConstantVariable.create("__abs__")], {} ) return abs_method.call_function(tx, [], {}) def call_range(self, tx, *args): if self.unspec_python_args(*args) or self.constant_args(*args): return variables.RangeVariable(args) elif self._dynamic_args(*args): args = [ variables.ConstantVariable.create(guard_if_dyn(arg)) for arg in args ] return variables.RangeVariable(args) # None no-ops this handler and lets the driving function proceed return None def _dynamic_args(self, *args, **kwargs): return any(isinstance(x, SymNodeVariable) for x in args) or any( isinstance(x, SymNodeVariable) for x in kwargs.values() ) def call_slice(self, tx, *args): return variables.SliceVariable(args) def _dyn_proxy(self, tx, *args, **kwargs): from .builder import wrap_fx_proxy return wrap_fx_proxy( tx, tx.output.create_proxy( "call_function", self.fn, *proxy_args_kwargs(args, kwargs) ), ) def _call_iter_tuple_list(self, tx, obj=None, *args, **kwargs): if self._dynamic_args(*args, **kwargs): return self._dyn_proxy(tx, *args, **kwargs) if isinstance(obj, variables.IteratorVariable): # For non-list iterators, we will guard on vars that # determine the control flow return obj # TODO This should probably be treated as a dict, or dicts should also be treated here if self.fn == set: cls = SetVariable else: cls = variables.BaseListVariable.cls_for(self.fn) if obj is None: if cls is SetVariable: return cls( [], mutable_local=MutableLocal(), ) else: return cls( [], mutable_local=MutableLocal(), ) elif obj.has_unpack_var_sequence(tx): if obj.source and not is_constant_source(obj.source): if isinstance(obj, TupleIteratorVariable): install_guard( obj.source.make_guard(GuardBuilder.TUPLE_ITERATOR_LEN) ) else: install_guard(obj.source.make_guard(GuardBuilder.LIST_LENGTH)) if cls is SetVariable: return cls( list(obj.unpack_var_sequence(tx)), mutable_local=MutableLocal(), ) return cls( list(obj.unpack_var_sequence(tx)), mutable_local=MutableLocal(), ) call_iter = _call_iter_tuple_list call_tuple = _call_iter_tuple_list call_list = _call_iter_tuple_list call_set = _call_iter_tuple_list def call_callable(self, tx, arg): from .functions import BaseUserFunctionVariable if isinstance( arg, (variables.UserDefinedClassVariable, BaseUserFunctionVariable) ): return variables.ConstantVariable.create(True) def call_cast(self, _, *args, **kwargs): if len(args) == 2: return args[1] unimplemented(f"unsupported args to builtin cast(): {args} {kwargs}") def call_dict(self, tx, *args, **kwargs): return BuiltinVariable.call_custom_dict(tx, dict, *args, **kwargs) @staticmethod def call_custom_dict(tx, user_cls, *args, **kwargs): if not kwargs: if not args: args = ({},) assert len(args) == 1 arg = args[0] if isinstance(arg, dict): return ConstDictVariable(arg, user_cls, mutable_local=MutableLocal()) elif isinstance(arg, variables.ConstDictVariable): return arg.clone(user_cls=user_cls, mutable_local=MutableLocal()) elif isinstance( arg, ( ListVariable, TupleVariable, ListIteratorVariable, ), ): items = user_cls() for x in arg.unpack_var_sequence(tx): k, v = x.unpack_var_sequence(tx) k = ConstDictVariable.get_key(k) items.update({k: v}) return ConstDictVariable(items, user_cls, mutable_local=MutableLocal()) elif not args and kwargs: return variables.ConstDictVariable( dict(kwargs), user_cls=user_cls, mutable_local=MutableLocal() ) unimplemented(f"{user_cls.__name__}(): {args} {kwargs}") @staticmethod def call_custom_dict_fromkeys(tx, user_cls, *args, **kwargs): assert user_cls in {dict, OrderedDict, defaultdict} if kwargs: # Only `OrderedDict.fromkeys` accepts `value` passed by keyword assert user_cls is OrderedDict assert len(args) == 1 and len(kwargs) == 1 and "value" in kwargs args = (*args, kwargs.pop("value")) if len(args) == 0: raise UserError(TypeError, "fromkeys expected at least 1 argument, got 0") if len(args) == 1: args = (*args, ConstantVariable.create(None)) assert len(args) == 2 arg, value = args DictVariableType = ( ConstDictVariable if user_cls is not defaultdict else DefaultDictVariable ) if isinstance(arg, dict): return DictVariableType( dict.fromkeys(arg, value), user_cls, mutable_local=MutableLocal() ) elif isinstance( arg, ( ConstDictVariable, ListVariable, TupleVariable, ListIteratorVariable, ), ): keys = [DictVariableType.get_key(x) for x in arg.unpack_var_sequence(tx)] return DictVariableType( dict.fromkeys(keys, value), user_cls, mutable_local=MutableLocal() ) unimplemented(f"{user_cls.__name__}.fromkeys(): {args} {kwargs}") def call_zip(self, tx, *args, **kwargs): if kwargs: assert len(kwargs) == 1 and "strict" in kwargs if all(x.has_unpack_var_sequence(tx) for x in args): unpacked = [arg.unpack_var_sequence(tx) for arg in args] if kwargs.pop("strict", False) and len(unpacked) > 0: if not all(len(u) == len(unpacked[0]) for u in unpacked): raise UserError( ValueError, "zip() has one argument of len differing from others", ) items = [variables.TupleVariable(list(item)) for item in zip(*unpacked)] return variables.TupleVariable(items) def call_enumerate(self, tx, *args): if len(args) == 1: start = 0 else: assert len(args) == 2 assert isinstance(args[1], variables.ConstantVariable) start = args[1].as_python_constant() if args[0].has_unpack_var_sequence(tx): items = [ variables.TupleVariable( [variables.ConstantVariable.create(idx), var], ) for idx, var in enumerate(args[0].unpack_var_sequence(tx), start) ] return variables.TupleVariable(items) def call_len(self, tx, *args, **kwargs): return args[0].call_method(tx, "__len__", args[1:], kwargs) def call_getitem(self, tx, *args, **kwargs): return args[0].call_method(tx, "__getitem__", args[1:], kwargs) def call_isinstance(self, tx, arg, isinstance_type): arg_type = arg.python_type() isinstance_type = isinstance_type.as_python_constant() if isinstance(arg, variables.TensorVariable) and arg.dtype is not None: def _tensor_isinstance(tensor_var, tensor_type): def check_type(ty): if ty not in tensortype_to_dtype: return issubclass(arg.python_type(), ty) dtypes = tensortype_to_dtype[ty] return arg.dtype in dtypes if type(tensor_type) is tuple: return any(check_type(ty) for ty in tensor_type) else: return check_type(tensor_type) return variables.ConstantVariable.create( _tensor_isinstance(arg, isinstance_type) ) # UserDefinedObject with C extensions can have torch.Tensor attributes, # so break graph. if isinstance(arg, variables.UserDefinedObjectVariable) and isinstance( arg.value, types.MemberDescriptorType ): unimplemented( f"isinstance called on UserDefinedClass {arg} {isinstance_type}" ) # handle __instancecheck__ defined in user class if ( isinstance(arg, variables.UserDefinedObjectVariable) and "__instancecheck__" in isinstance_type.__class__.__dict__ ): return variables.ConstantVariable.create( isinstance_type.__class__.__instancecheck__(isinstance_type, arg.value) ) try: val = issubclass(arg_type, isinstance_type) except TypeError: val = arg_type is isinstance_type return variables.ConstantVariable.create(val) def call_issubclass(self, tx, left_ty, right_ty): """Checks if first arg is subclass of right arg""" left_ty = left_ty.as_python_constant() right_ty = right_ty.as_python_constant() return variables.ConstantVariable(issubclass(left_ty, right_ty)) def call_super(self, tx, a, b): return variables.SuperVariable(a, b) def call_next(self, tx, arg): if isinstance( arg, (variables.ListIteratorVariable, variables.IteratorVariable) ): val, next_iter = arg.next_variables(tx) return val elif isinstance(arg, variables.BaseListVariable): return arg.items[0] def call_hasattr(self, tx, obj, attr): if attr.is_python_constant(): name = attr.as_python_constant() return obj.call_hasattr(tx, name) def call_map(self, tx, fn, seq): if seq.has_unpack_var_sequence(tx): items = [fn.call_function(tx, [x], {}) for x in seq.unpack_var_sequence(tx)] return variables.TupleVariable(items) def call_sum(self, tx, seq, **kwargs): # Special case for sum on tuple of floats and ints if ( isinstance(seq, (variables.ListVariable, variables.TupleVariable)) and all( isinstance(x, variables.ConstantVariable) and isinstance(x.value, (int, float)) for x in seq.items ) and not kwargs ): new_list = [x.value for x in seq.items] return variables.ConstantVariable.create(sum(new_list)) if seq.has_unpack_var_sequence(tx): start = kwargs.pop( "start", variables.ConstantVariable.create(0) ).as_python_constant() assert not kwargs items = seq.unpack_var_sequence(tx)[start:] return BuiltinVariable(functools.reduce).call_function( tx, [ BuiltinVariable(operator.add), variables.TupleVariable(items), variables.ConstantVariable.create(0), ], {}, ) def call_reduce(self, tx, function, iterable, initializer=None): if iterable.has_unpack_var_sequence(tx): items = iterable.unpack_var_sequence(tx) if initializer is None: value, items = items[0], items[1:] else: value = initializer for element in items: value = function.call_function(tx, [value, element], {}) return value def call_getattr( self, tx, obj: VariableTracker, name_var: VariableTracker, default=None ): from .. import trace_rules from . import ( ConstantVariable, GetAttrVariable, PythonModuleVariable, TorchVariable, UserFunctionVariable, ) from .builder import SourcelessBuilder, VariableBuilder name = name_var.as_python_constant() if not name_var.is_python_constant(): unimplemented("non-const getattr() name") if tx.output.side_effects.is_attribute_mutation(obj): try: # re-read a pending side effect? return tx.output.side_effects.load_attr(obj, name) except KeyError: pass if default is not None: hasattr_var = self.call_hasattr(tx, obj, name_var) assert hasattr_var.as_python_constant() in (True, False) if not hasattr_var.as_python_constant(): return default options = {} if obj.source: source = AttrSource(obj.source, name) options["source"] = source else: source = None if name == "__bases__": try: value = obj.as_python_constant() if isinstance(value, type): bases = value.__bases__ if source is not None: tuple_args = [ VariableBuilder(tx, GetItemSource(source, i))(b) for i, b in enumerate(bases) ] else: tuple_args = [SourcelessBuilder()(tx, b) for b in bases] return variables.TupleVariable(tuple_args, **options) except NotImplementedError: pass if isinstance(obj, variables.NNModuleVariable): return obj.var_getattr(tx, name) elif isinstance(obj, variables.TensorVariable) and name == "grad": if source: # We are going to be raising this tensor as grapharg. So, ensure # that we have real grad value instead of fake tensor value. # Walk through the inputs of the subgraph and find if we already # have the original tensor stored in the graphargs. for grapharg in tx.output.graphargs: if grapharg.source == source.base: old_grad = grapharg.example.grad new_grad = obj.as_proxy().node.meta["example_value"].grad def _grad_changed(old, new): if old is None or new is None: return new is not old try: if old.shape != new.shape: return True if old.stride() != new.stride(): return True return False except TypeError as te: # There is a rare edge case in which # we seem to get symbol mismatches # for jagged tensor comparison. # See PYTORCH_TEST_WITH_DYNAMO=1 python test/test_nestedtensor.py # -k test_dropout_backward_layout_torch_jagged_cpu unimplemented(str(te)) if _grad_changed(old_grad, new_grad): if new_grad is not None: grad_shape_specialized = [ int(x) for x in new_grad.shape ] # We lazily update the grad on the example to its real state as tracked by fake tensor. # This allocation is fine - it is just a hint. It will not make it to runtime, but it coerces # the underlying value to always be correct. grapharg.example.grad = torch.zeros( grad_shape_specialized, device=new_grad.device ) else: grapharg.example.grad = None return VariableBuilder(tx, source)(grapharg.example.grad) unimplemented("tensor grad") else: unimplemented("tensor grad") elif isinstance( obj, ( variables.TensorVariable, variables.NamedTupleVariable, variables.ConstantVariable, variables.UserDefinedClassVariable, variables.UserDefinedObjectVariable, ), ): try: return obj.var_getattr(tx, name).clone(source=source) except NotImplementedError: return GetAttrVariable(obj, name, **options) elif isinstance(obj, TorchVariable): member = getattr(obj.value, name) if is_utils_checkpoint(member): options["source"] = source return build_checkpoint_variable(**options) elif trace_rules.lookup(member) is not None: return trace_rules.lookup(member)(member, **options) elif source is not None: return VariableBuilder(tx, source)(member) else: return SourcelessBuilder()(tx, member) elif isinstance(obj, (PythonModuleVariable, DummyModule)): member = obj.value.__dict__[name] if config.replay_record_enabled: tx.exec_recorder.record_module_access(obj.value, name, member) return VariableBuilder(tx, source)(member) elif istype(obj, UserFunctionVariable) and name in ("__name__", "__module__"): return ConstantVariable.create(getattr(obj.fn, name)) else: try: return obj.var_getattr(tx, name).clone(source=source) except NotImplementedError: return GetAttrVariable(obj, name, **options) def call_setattr( self, tx, obj: VariableTracker, name_var: VariableTracker, val: VariableTracker ): from .distributed import PlacementVariable if isinstance( obj, ( variables.DataClassVariable, variables.CustomizedDictVariable, PlacementVariable, ), ): return obj.call_method(tx, "__setattr__", [name_var, val], {}) elif ( tx.output.side_effects.is_attribute_mutation(obj) and name_var.is_python_constant() ): name = name_var.as_python_constant() if isinstance(obj, variables.TensorVariable): from .builder import wrap_fx_proxy if name == "requires_grad": # TODO(voz): Make it work properly unimplemented( "mutating requires_grad can introduce a new leaf from non-leaf or vice versa in " "the middle of the graph, which aot_autograd does not currently know how to handle. " ) if name == "data": # Remove the old reference in tracked fakes - if we don't do this # new .data value size and shape differences will cause # tracked fakes to produce incorrect guards. This is sound because the TensorVariable # coming out of set_() below will be a new one, and get # installed in tracked fakes. to_remove = [] for tf in tx.output.tracked_fakes: if tf.source == obj.source: to_remove.append(tf) for tf in to_remove: tx.output.tracked_fakes.remove(tf) # Step 1 - disable grads with dynamo_disable_grad(tx), torch.no_grad(): # Step 2 - call `set_` out = wrap_fx_proxy( tx, tx.output.create_proxy( "call_function", torch.Tensor.set_, *proxy_args_kwargs([obj, val], {}), ), ) # Step 3 - drop the version counter - this is a step required to get # .data setting to play correctly with the autograd engine. # Esentially, dynamo is trying to faithful preserve the (absurd) # behavior of .data= from eager mode def _lower_version_count_by_1(x): version = x._version if version > 0: version = version - 1 torch._C._autograd._unsafe_set_version_counter(x, version) return x tx.output.create_proxy( "call_function", _lower_version_count_by_1, (out.as_proxy(),), {}, ) _lower_version_count_by_1(obj.as_proxy().node.meta["example_value"]) # This handles options prop, guards and ends with a clone # Step 4 - replace all reference to the current object with the new one return out tx.output.side_effects.store_attr(obj, name, val) return val elif isinstance(obj, variables.UserDefinedObjectVariable): unimplemented( f"setattr(UserDefinedObjectVariable) {type(obj.value).__setattr__}" ) elif isinstance(obj, variables.NNModuleVariable): if not tx.output.is_root_tracer(): raise AttributeMutationError( "Can't inplace modify module params/buffers inside HigherOrderOp" ) if name_var.is_python_constant() and isinstance( val, variables.TensorVariable ): assigning_fake_val = get_fake_value(val.as_proxy().node, tx) try: getattr_var = obj.var_getattr(tx, name_var.as_python_constant()) except AttributeError: getattr_var = None if isinstance(getattr_var, variables.TensorVariable): # get_fake_val will get the same fake tensor existing_fake_attr = get_fake_value(getattr_var.as_proxy().node, tx) # same tensor identiy, setattr is a no-op mod_setattr = inspect.getattr_static(obj.module_type, "__setattr__") if ( existing_fake_attr is assigning_fake_val and mod_setattr is torch.nn.Module.__setattr__ ): return getattr_var obj.convert_to_unspecialized(tx) # FIXME (tmanlaibaatar) this is utter hack to unblock HuggingFace export # Export generally doesn't want to allow mutations on objects directly, # but we don't have good way to do this rn. For now, we make it an undefined # behaviour and just set attributes directly on the PretrainedConfig object # for now. elif isinstance(obj, variables.dicts.HFPretrainedConfigVariable) and tx.export: if name_var.is_python_constant() and isinstance( val, variables.ConstantVariable ): setattr( obj.obj, name_var.as_python_constant(), val.as_python_constant() ) return ConstantVariable(None) def call_delattr(self, tx, obj: VariableTracker, name_var: VariableTracker): return self.call_setattr(tx, obj, name_var, variables.DeletedVariable()) def call_type(self, tx, obj: VariableTracker): from .builder import SourcelessBuilder, VariableBuilder try: py_type = obj.python_type() except NotImplementedError as error: raise UserError( UserErrorType.INVALID_INPUT, str(error), case_name="unknown_python_type", ) from None if obj.source is None: return SourcelessBuilder()(tx, py_type) else: return VariableBuilder(tx, TypeSource(obj.source))(py_type) def call_reversed(self, tx, obj: VariableTracker): if obj.has_unpack_var_sequence(tx): items = list(reversed(obj.unpack_var_sequence(tx))) return variables.TupleVariable(items) def call_sorted(self, tx, obj: VariableTracker, **kwargs): if ( obj.has_unpack_var_sequence(tx) and not isinstance(obj, variables.TensorVariable) and all(x.is_python_constant() for x in obj.unpack_var_sequence(tx)) ): function = kwargs.pop("key", None) reverse = kwargs.pop( "reverse", ConstantVariable.create(False) ).as_python_constant() assert len(kwargs) == 0 if function: items = sorted( obj.unpack_var_sequence(tx), key=lambda x: function.call_function( tx, [x], {} ).as_python_constant(), reverse=reverse, ) else: items = sorted( obj.unpack_var_sequence(tx), key=lambda x: x.as_python_constant(), reverse=reverse, ) return variables.ListVariable(items) def call_chain(self, tx, *args): if all(obj.has_unpack_var_sequence(tx) for obj in args): items = [] for obj in args: items.extend(obj.unpack_var_sequence(tx)) return variables.TupleVariable(items) def call_islice(self, tx, iterable, *args): if iterable.has_unpack_var_sequence(tx) and all( x.is_python_constant() for x in args ): const_args = [x.as_python_constant() for x in args] items = iterable.unpack_var_sequence(tx) items = list(itertools.islice(items, *const_args)) return variables.TupleVariable(items) # neg is a constant fold function, so we only get here if constant fold is not valid def call_neg(self, tx, a): if isinstance(a, SymNodeVariable): return SymNodeVariable.create( tx, (operator.neg)(a.as_proxy()), sym_num=None, ) # None no-ops this handler and lets the driving function proceed return None def call_format(self, tx, _format_string, *args, **kwargs): format_string = _format_string.as_python_constant() return variables.StringFormatVariable.create(format_string, args, kwargs) def call_id(self, tx, *args): if len(args) > 0 and isinstance(args[0], variables.NNModuleVariable): nn_mod_variable = args[0] mod = tx.output.get_submodule(nn_mod_variable.module_key) return variables.ConstantVariable.create(id(mod)) else: unimplemented(f"call_id with args {args}") def call_deepcopy(self, tx, x): unimplemented(f"copy.deepcopy {repr(x)}") def _comparison(self, tx, left, right): """ Used to implement comparison operators for different types. For example, list1 < list2 is implemented differently from tensor1 < tensor2 """ from . import ( BaseListVariable, ConstantVariable, NNModuleVariable, TensorVariable, UserDefinedObjectVariable, UserFunctionVariable, ) from .lists import SizeVariable from .tensor import ( supported_const_comparison_ops, supported_tensor_comparison_ops, ) op = self.fn def _unimplemented(): unimplemented(f"comparison {typestr(left)} {op} {typestr(right)}") if ( all( isinstance(x, (NNModuleVariable, ConstantVariable)) for x in [left, right] ) and op in supported_const_comparison_ops.values() ): left = ( tx.output.get_submodule(left.module_key) if isinstance(left, NNModuleVariable) else left.as_python_constant() ) right = ( tx.output.get_submodule(right.module_key) if isinstance(right, NNModuleVariable) else right.as_python_constant() ) return ConstantVariable.create(op(left, right)) if isinstance(left, UserFunctionVariable): if op not in supported_const_comparison_ops.values(): _unimplemented() if not isinstance(right, UserFunctionVariable): _unimplemented() return ConstantVariable.create(op(left.fn, right.fn)) # Note, we have a rare BaseListVariable subtype mismatch with valid comparison # x = torch.randn([3, 3]) # x.size() == (3, 3) # True # (3, 3) == x.size() # True if isinstance(left, (SizeVariable, TupleVariable)) and isinstance( right, (TupleVariable, SizeVariable) ): return BaseListVariable.list_compare(tx, op, left, right) if isinstance(left, BaseListVariable): if not type(left) == type(right): # Mismatch in BaseListVariable subclasses _unimplemented() return BaseListVariable.list_compare(tx, op, left, right) if isinstance(left, SetVariable): if not type(left) == type(right): # Mismatch in BaseListVariable subclasses _unimplemented() return ConstantVariable.create( op(left._underlying_items, right._underlying_items) ) if isinstance(left, TensorVariable) or isinstance(right, TensorVariable): from .builder import wrap_fx_proxy_cls if op is operator.is_ and isinstance(right, TensorVariable): return ConstantVariable.create( id(extract_fake_example_value(left.as_proxy().node)) == id(extract_fake_example_value(right.as_proxy().node)) ) if op not in supported_tensor_comparison_ops.values(): _unimplemented() if ( isinstance(left, TensorVariable) and isinstance(right, TensorVariable) and (left.size and right.size) is not None and left.size != right.size ): try: torch.broadcast_shapes(left.size, right.size) except RuntimeError: # not broadcastable, can't be compared _unimplemented() tensor_cls = left if isinstance(left, TensorVariable) else right return wrap_fx_proxy_cls( type(tensor_cls), # handle Ndarrays and Tensors tx, proxy, ) if isinstance(left, SymNodeVariable) or isinstance(right, SymNodeVariable): if op not in supported_tensor_comparison_ops.values(): _unimplemented() proxy = tx.output.create_proxy( "call_function", op, (left.as_proxy(), right.as_proxy()), {} ) return SymNodeVariable.create( tx, proxy, sym_num=None, ) if isinstance(left, UserDefinedObjectVariable) and isinstance( right, UserDefinedObjectVariable ): return ConstantVariable.create(op(left.value, right.value)) if ( (isinstance(left, StreamVariable) and isinstance(right, StreamVariable)) or (isinstance(left, EventVariable) and isinstance(right, EventVariable)) ) and op is operator.eq: return ConstantVariable(op(left.value, right.value)) if op.__name__ == "is_": # If the two objects are of different type, we can safely return False if type(left) is not type(right): return ConstantVariable.create(False) if isinstance(left, BuiltinVariable) and isinstance(right, BuiltinVariable): return ConstantVariable.create(op(left.fn, right.fn)) _unimplemented() def call_and_(self, tx, a, b): # Rely on constant_handler if isinstance(a, ConstantVariable) and isinstance(b, ConstantVariable): return None if isinstance(a, (SymNodeVariable, ConstantVariable)) and isinstance( b, (SymNodeVariable, ConstantVariable) ): return SymNodeVariable.create( tx, tx.output.create_proxy( "call_function", operator.and_, *proxy_args_kwargs([a, b], {}) ), sym_num=None, ) # None no-ops this handler and lets the driving function proceed return None def call_or_(self, tx, a, b): # Rely on constant_handler if isinstance(a, ConstantVariable) and isinstance(b, ConstantVariable): return None if isinstance(a, (SymNodeVariable, ConstantVariable)) and isinstance( b, (SymNodeVariable, ConstantVariable) ): return SymNodeVariable.create( tx, tx.output.create_proxy( "call_function", operator.or_, *proxy_args_kwargs([a, b], {}) ), sym_num=None, ) # None no-ops this handler and lets the driving function proceed return None def call_not_(self, tx, a): if isinstance(a, SymNodeVariable): return SymNodeVariable.create( tx, tx.output.create_proxy( "call_function", operator.not_, *proxy_args_kwargs([a], {}) ), sym_num=None, ) if isinstance(a, ListVariable): return ConstantVariable.create(len(a.items) == 0) return None call_eq = _comparison call_gt = _comparison call_lt = _comparison call_ge = _comparison call_le = _comparison call_ne = _comparison call_is_ = _comparison call_is_not = _comparison call_all = _polyfill_call_impl("all") call_any = _polyfill_call_impl("any") @contextlib.contextmanager def dynamo_disable_grad(tx): from . import GradModeVariable org_value = torch.is_grad_enabled() gmv = GradModeVariable.create(tx, False) try: gmv.enter(tx) yield finally: gmv.exit(tx)