Google Git
Sign in
android / platform / external / pytorch / 492de213e27eb268ec62b9ee7be5718f472ab73f / . / torch / _dynamo / variables / builder.py
blob: 7f0e9d7e083b57c6a2c49a221096b217d71812d7 [file] [log] [blame]
# mypy: ignore-errors
import abc
import collections
import contextlib
import dataclasses
import enum
import functools
import inspect
import itertools
import logging
import math
import operator
import re
import sys
import types
import weakref
from typing import Any, List, NamedTuple, Optional, Union
from torch._utils_internal import justknobs_check
from torch.utils._sympy.value_ranges import ValueRanges
try:
import numpy as np
except ModuleNotFoundError:
np = None
import torch
from torch import SymInt
from torch._guards import GuardSource, TracingContext
from torch._higher_order_ops.torchbind import call_torchbind
from torch._ops import HigherOrderOperator
from torch._streambase import _EventBase, _StreamBase
from torch._subclasses.fake_tensor import FakeTensor, is_fake, maybe_get_fake_mode
from torch._subclasses.meta_utils import is_sparse_any
from torch.fx.experimental._backward_state import BackwardState
from torch.fx.experimental.symbolic_shapes import (
_constrain_range_for_size,
DimDynamic,
RelaxedUnspecConstraint,
StatefulSymbolicContext,
SubclassSymbolicContext,
SymbolicContext,
)
from torch.fx.immutable_collections import immutable_dict, immutable_list
from torch.utils._python_dispatch import is_traceable_wrapper_subclass
from torch.utils.weak import TensorWeakRef
from .. import config, mutation_guard, replay_record, trace_rules
from ..device_interface import get_registered_device_interfaces
from ..exc import InternalTorchDynamoError, unimplemented
from ..guards import GuardBuilder, install_guard, make_dupe_guard
from ..side_effects import SideEffects
from ..source import (
AttrSource,
CallMethodItemSource,
ConstantSource,
ConstDictKeySource,
ConvertIntSource,
FloatTensorSource,
GetItemSource,
GradSource,
is_cell_contents,
is_constant_source,
is_from_defaults,
is_from_optimizer_source,
is_unspecialized_builtin_nnmodule_attr,
LocalSource,
NumpyTensorSource,
OptimizerSource,
RandomValueSource,
Source,
SubclassAttrListSource,
TupleIteratorGetItemSource,
)
from ..trace_rules import (
is_callable_allowed,
is_numpy,
is_numpy_dtype,
is_numpy_type_info,
)
from ..utils import (
build_checkpoint_variable,
clone_input,
common_constant_types,
get_fake_value,
get_locals_to_steal,
get_static_address_type,
is_function_or_wrapper,
is_lru_cache_wrapped_function,
is_namedtuple,
is_typing,
is_utils_checkpoint,
is_wrapper_or_member_descriptor,
istype,
odict_values,
proxy_args_kwargs,
set_example_value,
tensor_always_has_static_shape,
tuple_iterator,
tuple_iterator_getitem,
tuple_iterator_len,
unwrap_with_attr_name_if_wrapper,
wrap_fake_exception,
)
from .base import MutableLocal, typestr, VariableTracker, VariableTrackerMeta
from .constant import ConstantVariable, EnumVariable
from .ctx_manager import (
AutocastModeVariable,
EventVariable,
NullContextVariable,
PreserveVersionContextVariable,
StreamContextVariable,
StreamVariable,
)
from .dicts import (
ConstDictVariable,
CustomizedDictVariable,
DefaultDictVariable,
HFPretrainedConfigVariable,
PythonSysModulesVariable,
SetVariable,
)
from .distributed import (
DeviceMeshVariable,
PlacementClassVariable,
PlacementVariable,
ProcessGroupVariable,
WorldMetaClassVariable,
)
from .functions import (
CollectiveFunctionRewriteVariable,
FunctoolsPartialVariable,
TritonKernelVariable,
UserFunctionVariable,
UserMethodVariable,
WrapperUserFunctionVariable,
)
from .higher_order_ops import TorchHigherOrderOperatorVariable
from .iter import ItertoolsVariable
from .lazy import LazyVariableTracker
from .lists import (
BaseListVariable,
ListVariable,
NamedTupleVariable,
RangeVariable,
RestrictedListSubclassVariable,
SizeVariable,
SliceVariable,
TupleIteratorVariable,
TupleVariable,
)
from .misc import (
AutogradEngineVariable,
AutogradFunctionContextVariable,
AutogradFunctionVariable,
ComptimeVariable,
DebuggingVariable,
DelayGraphBreakVariable,
GetAttrVariable,
GetSetDescriptorVariable,
InspectSignatureVariable,
LambdaVariable,
LoggingLoggerVariable,
MethodWrapperVariable,
NumpyDTypeVariable,
NumpyTypeInfoVariable,
NumpyVariable,
PythonModuleVariable,
RegexPatternVariable,
SavedTensorBox,
TorchVersionVariable,
TypingVariable,
)
from .nn_module import (
FSDPManagedNNModuleVariable,
UnspecializedBuiltinNNModuleVariable,
UnspecializedNNModuleVariable,
)
from .optimizer import OptimizerVariable
from .script_object import TorchScriptObjectVariable
from .sdpa import SDPAParamsVariable
from .tensor import (
NumpyNdarrayVariable,
SymNodeVariable,
TensorSubclassVariable,
TensorVariable,
UnspecializedPythonVariable,
)
from .torch import TorchCtxManagerClassVariable, TorchInGraphFunctionVariable
from .torch_function import build_torch_function_fn, TensorWithTFOverrideVariable
from .user_defined import (
KeyedJaggedTensorVariable,
SourcelessGraphModuleVariable,
UserDefinedClassVariable,
UserDefinedObjectVariable,
WeakRefVariable,
)
log = logging.getLogger(__name__)
DimList = List
class _missing:
pass
@dataclasses.dataclass
class GraphArg:
source: Source
# TODO: storing a SymInt here but not a FakeTensor is a pretty strange
# thing to do. Probably should have example (which stores an int) and
# fake_example
_example: Union[TensorWeakRef, torch.SymInt]
# When True, this indicates that this GraphArg is a Python quantity (e.g.,
# a float or int) which we pass to the FX graph as a Tensor. This
# controls how we codegen calls into the Dynamo graph: we will call
# torch.as_tensor on the quantity before passing it in.
#
# Note that we typically do not pass dynamic integers as tensors, because
# they will most frequently just be used for size computation. But this
# is a policy decision that we can change our mind on; in particular, when
# an int comes from a random number generator (e.g., random.randint), we
# DO pass it as a tensor.
#
# It's also worth noting that our current tracing rules for
# pass_arg_as_tensor as subtly broken: we just pun the variable as a
# 0d scalar Tensor and pray that the semantics are the same. Which they
# often are, but not necessarily. ezyang(May 2024) plans to fix this
# soon.
pass_arg_as_tensor: bool
fake_tensor: Optional[torch._subclasses.fake_tensor.FakeTensor]
# UnspecializedPythonVariable often masquerades as a tensor.
# We MUST NOT generate shape guard code
# that actually tries to access tensor properties on these values.
# is_tensor lets us tell if this graph arg actually is a tensor
# or not.
is_tensor: bool = True
# Sometimes, the Tensor we pass to example is freshly allocated (smh).
# Then we cannot only keep a weak reference to it. This lets you
# stash a strong reference too.
example_strong_ref: Optional[torch.Tensor] = None
@property
def example(self):
if isinstance(self._example, TensorWeakRef):
r = self._example()
assert r is not None
return r
else:
return self._example
def __post_init__(self):
if isinstance(self._example, torch.Tensor):
self._example = TensorWeakRef(self._example)
assert is_fake(self.fake_tensor)
def reconstruct(self, codegen):
self.source.reconstruct(codegen)
def erase(self):
self._example = None
self.example_strong_ref = None
def __eq__(self, other):
return self.source.name() == other.source.name()
class BackwardStateGraphArg(GraphArg):
def __init__(self):
super().__init__(
source=None,
_example=BackwardState(),
pass_arg_as_tensor=False,
fake_tensor=None,
is_tensor=False,
)
def reconstruct(self, codegen):
assert codegen.tx.output.backward_state_var
codegen.add_push_null(
lambda: codegen.load_import_from(BackwardState.__module__, "BackwardState")
)
codegen.call_function(0, False)
codegen.dup_top()
codegen.store(codegen.tx.output.backward_state_var)
@dataclasses.dataclass
class FrameStateSizeEntry:
scalar: Optional[int]
size: Optional[List[int]]
class VariableBuilder:
"""Wrap a python value in a VariableTracker() instance"""
def __init__(
self,
tx,
source: Source,
):
assert (
source is not None
), "Consider SourcelessBuilder for ephemeral objects, usually objects created locally."
assert TracingContext.try_get() is not None, "Expected active TracingContext"
super().__init__()
self.tx = tx
self.source = source
self.name = source.name()
def __call__(self, value):
if value in self.tx.output.side_effects:
side_effect_result = self.tx.output.side_effects[value]
dup_guard = make_dupe_guard(self.source, side_effect_result.source)
if dup_guard:
self.install_guards(dup_guard)
return side_effect_result
cached_vt = self.tx.output.variable_tracker_cache.lookup(value, self.source)
if cached_vt:
return cached_vt
vt = self._wrap(value)
vt.source = self.source
if (
self._can_lift_attrs_to_inputs(vt)
and value not in self.tx.output.side_effects
and not is_wrapper_or_member_descriptor(value)
):
vt = self.tx.output.side_effects.track_object_existing(value, vt)
self.tx.output.variable_tracker_cache.add(value, self.source, vt)
return vt
def _can_lift_attrs_to_inputs(self, vt):
if type(vt) in [
TensorVariable,
TensorWithTFOverrideVariable,
UserDefinedObjectVariable,
NumpyNdarrayVariable,
]:
return True
return False
@staticmethod
@functools.lru_cache(None)
def _common_constants():
return {
# We zero-one specialize shapes, so specialize these constants
# too
0,
1,
# NB: There used to be more constants here, but honestly it was
# pretty confusing. Note we specialize floats by default, and
# DON'T specialize ints by default. This all only matters with
# dynamic_shapes
}
def get_source(self):
return self.source
def install_guards(self, *guards):
source = self.get_source()
if (
isinstance(source, ConstantSource)
or source.guard_source() == GuardSource.CONSTANT
):
return None
install_guard(*[source.make_guard(guard) for guard in guards], skip=1)
return {}
def set_source_and_track_mutable(self, value, var):
assert isinstance(var, VariableTracker)
var.source = self.source
return self.tx.output.side_effects.track_mutable(value, var)
@classmethod
@functools.lru_cache(None)
def _type_dispatch(cls):
# NB: Careful not to close over self to avoid ref cycle from lru_cache
entries = [
(
(
torch.Tensor,
torch.nn.Parameter,
torch._subclasses.FakeTensor,
torch._subclasses.functional_tensor.FunctionalTensor,
),
cls.wrap_tensor,
),
(
(tuple, list, odict_values, collections.deque, torch.Size),
cls.wrap_listlike,
),
(tuple_iterator, cls.wrap_tuple_iterator),
((slice, range), cls.wrap_slice_range),
(tuple(common_constant_types), cls.wrap_literal),
(re.Pattern, cls.wrap_regex_pattern),
(weakref.ReferenceType, cls.wrap_weakref),
(torch.utils.hooks.RemovableHandle, cls.wrap_removable_handle),
(torch.jit.ScriptFunction, cls.wrap_jit_function),
]
if config.trace_numpy and np:
entries.append((np.ndarray, cls.wrap_numpy_ndarray))
result = {}
for ts, fn in entries:
for t in ts if isinstance(ts, tuple) else (ts,):
assert t not in result
result[t] = fn
return result
def wrap_regex_pattern(self, value: re.Pattern):
# TODO(jansel): something like a REPR_MATCH might be more robust here
self.install_guards(GuardBuilder.ID_MATCH)
return RegexPatternVariable(value)
def wrap_weakref(self, value: weakref.ReferenceType):
self.install_guards(GuardBuilder.TYPE_MATCH)
return WeakRefVariable(value, source=self.source)
def wrap_removable_handle(self, value):
# This means that the removable handle was created in some other frame.
# Our current infra requires the hook to be registered and removed in
# the same frame. So graph break.
# Related test - PYTORCH_TEST_WITH_DYNAMO=1 python test/test_autograd.py -k TestAutograd.test_hooks
unimplemented("unregistered hook removable handle")
def wrap_jit_function(self, value):
self.install_guards(GuardBuilder.TYPE_MATCH)
return WrapperUserFunctionVariable(
value, "_torchdynamo_inline", source=self.source
)
@classmethod
@functools.lru_cache(None)
def _id_dispatch(cls):
from ..comptime import comptime
entries = [
(
inspect.signature,
lambda self, value: LambdaVariable(
InspectSignatureVariable.create,
source=self.source,
**self.install_guards(GuardBuilder.CLOSURE_MATCH),
),
),
(comptime, lambda self, value: ComptimeVariable()),
(
dataclasses.fields,
lambda self, value: LambdaVariable(
_dataclasses_fields_lambda,
source=self.source,
**self.install_guards(GuardBuilder.FUNCTION_MATCH),
),
),
(torch.__version__, lambda self, value: TorchVersionVariable()),
]
result = {}
for ts, fn in entries:
for t in ts if isinstance(ts, (tuple, list)) else (ts,):
assert t not in result
result[id(t)] = fn
return result
def _wrap(self, value):
# import here to avoid circular dependencies
from torch.utils._triton import has_triton
if has_triton():
from triton.runtime.autotuner import Autotuner
from triton.runtime.jit import JITFunction
else:
class JITFunction:
pass
class Autotuner:
pass
# Handle exact type() match
type_dispatch = self._type_dispatch().get(type(value))
if type_dispatch is not None:
return type_dispatch(self, value)
# Handle exact id() match
id_dispatch = self._id_dispatch().get(id(value))
if id_dispatch is not None:
return id_dispatch(self, value)
# Note - There are some nested values where types mismatch!
# We want to get those out and wrap those.
value = inspect.getattr_static(value, "_torchdynamo_inline", value)
# Everything else (NB: order matters!)
if is_traceable_wrapper_subclass(value) or istype(
value, config.traceable_tensor_subclasses
):
return self.wrap_tensor(value)
elif is_namedtuple(value):
return self.wrap_listlike(value)
elif value is torch.utils._pytree.SUPPORTED_NODES:
# For SUPPORTED_NODES, we guard on the dictionary version (PEP509)
# under the assumption that the values themselves don't change.
self.install_guards(GuardBuilder.DICT_VERSION)
# The keys on the SUPPORTED_NODES can be arbitrary, so save on the
# key order.
self.tx.output.guard_on_key_order.add(self.source.name())
result = {
ConstantVariable.create(k): UserDefinedObjectVariable(
v,
source=GetItemSource(
self.get_source(), ConstDictKeySource(self.get_source(), i)
),
)
for i, (k, v) in enumerate(value.items())
}
return ConstDictVariable(result, type(value))
elif value is sys.modules:
self.install_guards(GuardBuilder.FUNCTION_MATCH)
return PythonSysModulesVariable(source=self.source)
elif CustomizedDictVariable.is_matching_cls_hf(type(value)):
self.install_guards(GuardBuilder.TYPE_MATCH)
result = CustomizedDictVariable.wrap(self, value)
result.source = self.source
return self.tx.output.side_effects.track_object_existing(value, result)
elif istype(value, (dict, collections.defaultdict, collections.OrderedDict)):
self.install_guards(GuardBuilder.SEQUENCE_LENGTH)
# Optimisation for the common case strings, ints, etc
all_const = all(ConstantVariable.is_literal(k) for k in value.keys())
if all_const:
# TODO(anijain2305) - Do we have to guard on all the keys? Can
# keys be guarded lazily, similar to values?
self.install_guards(GuardBuilder.DICT_CONST_KEYS)
else:
# Guard on the key order
# This is not ideal, i.e., there is no need to guard on the key
# order. But we guard on the key order because of the complexity
#
# 1) For non-constant objects, we can't save the key in the
# guard context because it can be memory heavy. We can add
# weakrefs but this complicates the accesses.
#
# 2) For non-constant objects, we also have to guard on the keys
# (like TENSOR_MATCH on tensor). We might also have guards on
# the attributes of the keys (like tensor.grad). To make this
# work in tree strucutre is complicated.
#
# So, instead we guard on the key order. While guarding on key
# order, we just save the indices and use it to access keys and
# values. Indices are cheap to save.
self.tx.output.guard_on_key_order.add(self.source.name())
# We need all the keys to be hashable. We do this within the
# _HashableTracker class in dicts.py
def build_key_value(i, k, v):
if all_const:
key = ConstantVariable.create(k)
source_key = k
else:
source_key = ConstDictKeySource(self.get_source(), i)
key = LazyVariableTracker.create(k, source_key)
source_value = GetItemSource(self.get_source(), source_key)
value = LazyVariableTracker.create(v, source_value)
return key, value
result = dict(
build_key_value(i, k, v) for i, (k, v) in enumerate(value.items())
)
if istype(value, collections.defaultdict):
factory_source = AttrSource(self.source, "default_factory")
result = DefaultDictVariable(
result,
type(value),
default_factory=VariableBuilder(self.tx, factory_source)(
value.default_factory
),
source=self.source,
)
else:
result = ConstDictVariable(result, type(value), source=self.source)
return self.set_source_and_track_mutable(value, result)
elif isinstance(value, torch.nn.Module):
return self.wrap_module(value)
elif ConstantVariable.is_literal(value): # non-atomic literals
return self.wrap_literal(value)
elif istype(value, frozenset) and (
ConstantVariable.is_literal(x) for x in value
):
# For frozenset, we can guard by object ID instead of value
# equality, this allows us to handle non-literal values
self.install_guards(GuardBuilder.ID_MATCH)
return ConstantVariable.create(value=value, source=self.source)
elif isinstance(value, enum.Enum):
self.install_guards(GuardBuilder.ID_MATCH)
return EnumVariable(value=value, source=self.source)
elif DebuggingVariable.is_reorderable_logging_function(value):
# Put this above builtin_callable so that print() can be handled
# along with other builtin debugging functions
self.install_guards(GuardBuilder.BUILTIN_MATCH)
return DebuggingVariable(value, source=self.source)
elif isinstance(value, logging.Logger):
self.install_guards(GuardBuilder.FUNCTION_MATCH)
return LoggingLoggerVariable(value, source=self.source)
elif is_utils_checkpoint(value):
return build_checkpoint_variable(source=self.source)
elif isinstance(value, functools.partial):
func_src = AttrSource(self.get_source(), "func")
func_obj = VariableBuilder(self.tx, func_src)(value.func)
args = []
args_source = AttrSource(self.get_source(), "args")
for i, arg in enumerate(value.args):
args.append(
VariableBuilder(self.tx, GetItemSource(args_source, i))(arg)
)
keywords = {}
keywords_source = AttrSource(self.get_source(), "keywords")
for k, v in value.keywords.items():
if not ConstantVariable.is_literal(k):
unimplemented("functools.partial with non-literal keyword")
keywords[k] = VariableBuilder(
self.tx, GetItemSource(keywords_source, k)
)(v)
install_guard(
self.get_source().make_guard(GuardBuilder.TYPE_MATCH),
keywords_source.make_guard(GuardBuilder.DICT_KEYS),
args_source.make_guard(GuardBuilder.SEQUENCE_LENGTH),
)
return FunctoolsPartialVariable(func_obj, args, keywords)
elif is_typing(value):
# typing.List, typing.Mapping, etc.
self.install_guards(GuardBuilder.ID_MATCH)
return TypingVariable(
value,
source=self.source,
)
elif np is not None and isinstance(value, np.generic):
# numpy array scalars: convert to 0D arrays
return self.wrap_numpy_ndarray(np.asarray(value))
elif is_numpy(value):
assert np
self.install_guards(
GuardBuilder.FUNCTION_MATCH
if callable(value)
else GuardBuilder.TYPE_MATCH
)
return NumpyVariable(value, source=self.source)
elif is_numpy_dtype(value):
self.install_guards(GuardBuilder.ID_MATCH)
return NumpyDTypeVariable(value, source=self.source)
elif is_numpy_type_info(value):
if isinstance(value, np.iinfo):
self.install_guards(GuardBuilder.TYPE_MATCH)
dt_source = AttrSource(self.source, "dtype")
install_guard(dt_source.make_guard(GuardBuilder.ID_MATCH))
else:
self.install_guards(GuardBuilder.ID_MATCH)
return NumpyTypeInfoVariable(value, source=self.source)
# NB: These can't be put in type_dispatch, they have to run later
elif CollectiveFunctionRewriteVariable.can_rewrite(value):
self.install_guards(GuardBuilder.FUNCTION_MATCH)
return CollectiveFunctionRewriteVariable.create(
self.tx,
value,
source=self.source,
)
elif istype(value, torch.autograd.function.FunctionMeta):
self.install_guards(GuardBuilder.FUNCTION_MATCH)
return AutogradFunctionVariable(
value,
source=self.source,
)
elif isinstance(value, torch.autograd.function.FunctionCtx):
actual_saved_tensors = None
try:
actual_saved_tensors = value.saved_tensors
except RuntimeError:
pass
saved_tensors = []
guards = [self.source.make_guard(GuardBuilder.TYPE_MATCH)]
if isinstance(actual_saved_tensors, tuple):
saved_tensors_source = AttrSource(self.source, "saved_tensors")
guards.append(
saved_tensors_source.make_guard(GuardBuilder.SEQUENCE_LENGTH)
)
for i, v in enumerate(actual_saved_tensors):
saved_tensors.append(
VariableBuilder(
self.tx, GetItemSource(saved_tensors_source, i)
)(v)
)
install_guard(*guards)
return self.tx.output.side_effects.track_object_existing(
value,
AutogradFunctionContextVariable(
value,
source=self.source,
saved_tensors=SavedTensorBox(saved_tensors),
),
)
elif (
isinstance(value, types.MethodType)
and istype(
getattr(value, "__self__", None), torch.autograd.function.FunctionMeta
)
and getattr(value, "__name__", "") == "apply"
and value == getattr(value.__self__, "apply", None)
):
# handle aliased autograd function `apply` calls
self.install_guards(GuardBuilder.FUNCTION_MATCH)
return GetAttrVariable(
AutogradFunctionVariable(
value.__self__, source=AttrSource(self.source, member="__self__")
),
"apply",
)
elif isinstance(value, torch._C._ImperativeEngine):
self.install_guards(GuardBuilder.ID_MATCH)
return AutogradEngineVariable(value, source=self.source)
elif (
value
is torch._dynamo.external_utils.FakeCompiledAutogradEngine._exec_final_callbacks_stub
):
self.install_guards(GuardBuilder.FUNCTION_MATCH)
return LambdaVariable(
lambda: UserFunctionVariable(
torch._dynamo.external_utils.FakeCompiledAutogradEngine.exec_final_callbacks,
).call_function(
self.tx,
(self.tx.output.side_effects.get_ca_final_callbacks_var(),),
{},
)
)
elif callable(value) and trace_rules.lookup_callable(value) is not None:
if is_callable_allowed(value):
self.tx.output.has_user_defined_allowed_in_graph = True
return trace_rules.lookup_callable(value).create_with_source(
value, source=self.source
)
elif np and isinstance(value, np.number):
return self.wrap_unspecialized_primitive(value)
elif HFPretrainedConfigVariable.is_matching_object(value):
self.install_guards(GuardBuilder.TYPE_MATCH)
return HFPretrainedConfigVariable(value)
elif isinstance(value, HigherOrderOperator):
self.install_guards(GuardBuilder.TYPE_MATCH, GuardBuilder.NAME_MATCH)
return TorchHigherOrderOperatorVariable.make(value, source=self.source)
elif isinstance(value, torch.cuda.StreamContext):
self.install_guards(GuardBuilder.ID_MATCH)
stream_source = AttrSource(self.source, "stream")
stream_var = VariableBuilder(self.tx, stream_source)(value.stream)
return StreamContextVariable.create(self.tx, stream_var)
elif isinstance(value, _StreamBase):
self.install_guards(GuardBuilder.ID_MATCH)
stream_proxy = self.tx.output.create_proxy(
"call_function",
torch.cuda.Stream,
(),
{
"stream_id": value.stream_id,
"device_index": value.device_index,
"device_type": value.device_type,
},
)
set_example_value(stream_proxy.node, value)
return StreamVariable(
stream_proxy,
value,
value.device,
source=self.source,
)
elif isinstance(value, (torch._C._SDPAParams)):
self.install_guards(GuardBuilder.TYPE_MATCH)
return SDPAParamsVariable.create(self.tx, value, self.source)
elif isinstance(value, _EventBase):
self.install_guards(GuardBuilder.ID_MATCH)
return EventVariable(
None,
value,
source=self.source,
)
elif (
isinstance(value, torch._C._TensorMeta)
and value in config.traceable_tensor_subclasses
):
return TensorSubclassVariable(value, source=self.source)
elif (
istype(value, contextlib.nullcontext)
and inspect.getattr_static(value, "enter_result", None) is None
):
self.install_guards(GuardBuilder.TYPE_MATCH)
return NullContextVariable(source=self.source)
elif KeyedJaggedTensorVariable.is_matching_object(value):
self.install_guards(GuardBuilder.TYPE_MATCH)
result = KeyedJaggedTensorVariable(value, source=self.source)
# TODO: this doing it manually is bad
return self.tx.output.side_effects.track_object_existing(value, result)
elif isinstance(value, torch.optim.Optimizer):
self.install_guards(GuardBuilder.ID_MATCH)
self.source = OptimizerSource(self.source)
return OptimizerVariable(value, source=self.source)
elif WorldMetaClassVariable.is_group_member_type(value):
return WorldMetaClassVariable(value, source=self.source)
elif ProcessGroupVariable.is_process_group(value):
self.install_guards(GuardBuilder.ID_MATCH)
return ProcessGroupVariable(value, source=self.source)
elif DeviceMeshVariable.is_device_mesh(value):
# TODO: see if we need to add custom guard instead of a simple ID_MATCH
self.install_guards(GuardBuilder.ID_MATCH)
return DeviceMeshVariable(value, source=self.source)
elif PlacementClassVariable.is_placement_type(value):
# TODO: see if we need to add custom guard instead of a simple ID_MATCH
self.install_guards(GuardBuilder.ID_MATCH)
return PlacementClassVariable(value, source=self.source)
elif PlacementVariable.is_placement(value):
# TODO: see if we need to add custom guard instead of a simple ID_MATCH
self.install_guards(GuardBuilder.ID_MATCH)
return PlacementVariable(
value,
source=self.source,
)
elif istype(value, type) and value in itertools.__dict__.values():
self.install_guards(GuardBuilder.FUNCTION_MATCH)
return ItertoolsVariable(value, source=self.source)
elif isinstance(value, torch.SymBool):
# Note: the idea here is to re-use the infra we've built for SymInt by simulating the
# user provided SymBool with a SymInt in dynamo.
# Concretely,
# 1. We create a SymInt in dynamo's shape_env, whose source is constructed as ConvertIntSource(self.source).
# so that guards on the SymInts can be effectively applied on the original SymBool in user program.
# 2. We create a SymBool based on the SymInt in dynamo's ShapeEnv. Because the original user program
# depends on the value being a SymBool. This allows dynamo to interpret the user's program correctly.
value_hint = value.node.require_hint()
new_source = ConvertIntSource(self.source)
new_symint = self.tx.output.shape_env.create_unspecified_symint_and_symbol(
int(value_hint),
new_source,
dynamic_dim=DimDynamic.DYNAMIC,
)
sym_node_proxy = self.tx.output.root_tracer.create_graph_input(
re.sub(r"[^a-zA-Z0-9]+", "_", self.name),
type(new_symint),
source=new_source,
)
sym_node_proxy.node.meta["grapharg"] = GraphArg(
new_source,
new_symint,
False,
None,
is_tensor=False,
example_strong_ref=new_symint,
)
self.tx.output.bound_symbols.add(new_symint.node.expr)
self.tx.output.tracked_fakes.append(
TrackedFake(new_symint, new_source, None)
)
return SymNodeVariable(
sym_node_proxy,
new_symint == 1,
)
elif isinstance(value, (JITFunction, Autotuner)):
self.install_guards(GuardBuilder.ID_MATCH)
return TritonKernelVariable(
value,
None, # No kernel idx provided
None, # No grid provided
source=self.source,
)
elif isinstance(value, torch.amp.autocast_mode.autocast):
self.install_guards(GuardBuilder.ID_MATCH)
return AutocastModeVariable(
target_values=[
value.device,
value.fast_dtype,
value._enabled,
value._cache_enabled,
],
source=self.source,
)
elif TorchCtxManagerClassVariable.is_matching_cls(value):
self.install_guards(GuardBuilder.FUNCTION_MATCH)
return TorchCtxManagerClassVariable(value, source=self.source)
elif inspect.getattr_static(value, "__script_if_tracing_wrapper", False):
self.install_guards(GuardBuilder.TYPE_MATCH)
return WrapperUserFunctionVariable(
value, "__original_fn", source=self.source
)
elif is_lru_cache_wrapped_function(value):
self.install_guards(GuardBuilder.TYPE_MATCH)
return WrapperUserFunctionVariable(value, "__wrapped__", source=self.source)
elif is_function_or_wrapper(value):
value, attr_name = unwrap_with_attr_name_if_wrapper(value)
# For these wrappers, Dynamo points to the wrapped function,
# so source needs to be updated as well.
if attr_name is not None:
self.source = AttrSource(self.source, attr_name)
return trace_rules.lookup(value).create_with_source(
value, source=self.source
)
# Don't use istype, since some python modules are not subclasses of types.ModuleType directly.
# E.g, type(torch.ops) -> <class 'torch._ops._Ops'>,
# type(torch.backends.cudnn) -> <class 'torch.backends.cudnn.CudnnModule'>
elif isinstance(value, (types.ModuleType, replay_record.DummyModule)):
self.install_guards(GuardBuilder.FUNCTION_MATCH)
return PythonModuleVariable(
value,
source=self.source,
)
elif isinstance(value, types.MethodType) and isinstance(
value.__self__, (torch.nn.Module, torch.utils._pytree.TreeSpec)
):
# don't let MethodTypes fall through to UserDefinedObject,
# which doesn't support 'CALL_FUNCTION'
# TODO(whc): Why do we limit this to methods on NNModules?
# I don't have a good reason for this, but it preserves the existing behavior
# for MBartForConditionalGeneration, which generates many graph breaks and OOMs otherwise.
# I suspect we probably want to relax this check and dig deeper there.
# In order to construct a MethodVariable in Dynamo, we start with an actual method obj from python,
# but need to separately wrap its underlying `__func__` and its `self` argument. We wrap `self` here
# and then `__func__` gets wrapped inside UserMethodVariable.
self_obj = VariableBuilder(
self.tx, source=AttrSource(self.source, "__self__")
)(value.__self__)
assert self_obj and isinstance(
self_obj, VariableTracker
), "Failed to produce a valid self obj"
self.install_guards(GuardBuilder.FUNCTION_MATCH)
return UserMethodVariable(
value.__func__,
self_obj,
source=self.source,
)
elif isinstance(value, types.GetSetDescriptorType):
# GetSet descriptors are C functions attached to an attribute lookup
# using PyGetSetDef. Python, on attribute lookup, can decide to
# create a new object on the fly, and therefore the `id` of the
# descriptors is not guaranteed to be same for different attribute
# accesses. Since these are unlikely to change during the program
# execution, we can skip guarding on them.
return GetSetDescriptorVariable(value)
elif isinstance(value, types.MethodWrapperType):
# Method-wrappers are written in C, and they are not guaranteed to
# return the same object on attribute lookup. Therefore, we cannot
# insert a FUNCTION_MATCH guard here. method-wrappers are very
# unlikely to change, so its ok to skip the guard here.
return MethodWrapperVariable(value)
elif issubclass(type(value), type):
if value in (torch.utils.hooks.BackwardHook, torch.nn.Parameter):
# TODO(jansel): combine this case with the one above
return trace_rules.lookup(value).create_with_source(
value, source=self.source
)
if value is torch.autograd._unsafe_preserve_version_counter:
self.install_guards(GuardBuilder.FUNCTION_MATCH)
return PreserveVersionContextVariable.constructor(self.tx)
# This is a userdefined class, so install an ID_MATCH even if its a
# global variable.
self.install_guards(GuardBuilder.ID_MATCH)
return UserDefinedClassVariable(
value,
source=self.source,
)
elif RestrictedListSubclassVariable.is_matching_cls(type(value)):
self.install_guards(GuardBuilder.SEQUENCE_LENGTH)
return self.set_source_and_track_mutable(
value,
RestrictedListSubclassVariable(
[
LazyVariableTracker.create(
value=value[i], source=GetItemSource(self.source, i)
)
for i in range(len(value))
],
user_cls=type(value),
user_cls_source=AttrSource(self.source, "__class__"),
),
)
elif TorchScriptObjectVariable.is_matching_cls(type(value)):
from ..source import (
FlattenScriptObjectSource,
ScriptObjectQualifiedNameSource,
)
# This exists to allow a smoother transition.
# The implications are:
# The script objects won't be tracked as proxies.
# Methods on these objects won't show up in the graph.
# The original script object might be mutated.
if not hasattr(value, "__obj_flatten__"):
return self.wrap_user_defined(value)
# Install the guards on the fully qualified name of the script object
LazyVariableTracker.realize_all(
VariableBuilder(self.tx, ScriptObjectQualifiedNameSource(self.source))(
value._type().qualified_name() # type: ignore[attr-defined]
)
)
# Install the guards on the content of the script object by setting the source
# to be FlattenScriptObjectSource, which calls __obj_flatten__() to get the contents.
LazyVariableTracker.realize_all(
VariableBuilder(self.tx, FlattenScriptObjectSource(self.source))(
value.__obj_flatten__()
)
)
fake_script_obj = torch._library.fake_class_registry.to_fake_obj(
self.tx.output.fake_mode, value
)
proxy = self.tx.output.root_tracer.create_graph_input(
re.sub(r"[^a-zA-Z0-9]+", "_", self.name),
type(value),
source=self.source,
)
# setting is_unspecialized=False to not insert a as_tensor call in reconstruct by default
# seting example to be real value because these example values will be used
# as example_inputs for user compiler.
proxy.node.meta["grapharg"] = GraphArg(
self.source, value, False, None, False, fake_script_obj
)
return TorchScriptObjectVariable.create(
proxy,
fake_script_obj,
source=self.source,
)
else:
return self.wrap_user_defined(value)
def wrap_user_defined(self, value: Any):
self.install_guards(GuardBuilder.TYPE_MATCH)
result = UserDefinedObjectVariable(value, source=self.source)
if not SideEffects.cls_supports_mutation_side_effects(type(value)):
# don't allow STORE_ATTR mutation with custom __setattr__
return result
return self.tx.output.side_effects.track_object_existing(value, result)
def wrap_listlike(self, value: Union[tuple, list, odict_values, NamedTuple]):
if config.specialize_int and type(value) is torch.Size:
self.install_guards(GuardBuilder.CONSTANT_MATCH)
return ConstantVariable.create(value=value)
if (
self.source
and is_unspecialized_builtin_nnmodule_attr(self.source)
and type(value) is tuple
and all(ConstantVariable.is_literal(x) for x in value)
):
# Heuristic to speedup up guards coming from conv2d attrs like dilation and padding.
self.install_guards(GuardBuilder.CONSTANT_MATCH)
return TupleVariable([ConstantVariable.create(x) for x in value])
# One can index a tensor with a list/tuple. Therefore, we need to
# have a stricter match.
self.install_guards(GuardBuilder.SEQUENCE_LENGTH)
for item in value:
if item is value:
unimplemented("list elements are pointing to the list itself")
output = [
LazyVariableTracker.create(item, source=GetItemSource(self.get_source(), i))
for i, item in enumerate(value)
]
maybe_gm = self.tx.output.local_scope.get("self")
if isinstance(
self.source, LocalSource
) and self.source.local_name in get_locals_to_steal(maybe_gm):
# The input tensor list to dynamo from compiled autograd may contain activations
# which are freed as they are used in inductor. Dynamo's default behavior is to
# lift all tensors to the graph inputs, but this will cause dynamo to hold an
# extra reference to the activation tensors and increase peak memory usage.
# To allow freeing ASAP, we keep the list as graph argument to the dynamo output
# graph, and unpack it locally.
# e.g. instead of `def forward(self, L_inputs_0_, L_inputs_1_, ...):`, we have
# `def forward(self, L_inputs_):`
source = self.source
assert isinstance(value, list)
tensor_list_proxy = self.tx.output.root_tracer.create_graph_input(
re.sub(r"[^a-zA-Z0-9]+", "_", self.name), type(value), source=source
)
tensor_list_proxy.node.meta["steal_arg"] = True
list_variable = wrap_fx_proxy_cls(
target_cls=TensorVariable,
tx=self.tx,
proxy=tensor_list_proxy,
example_value=value,
subclass_type=None,
source=source,
)
guards = []
for i, tensor_variable in enumerate(list_variable.items):
source_i = GetItemSource(base=source, index=i, index_is_slice=False)
# access unpacked tensor from this list instead of from a lifted arg
self.tx.output.input_source_to_var[source_i] = tensor_variable
guard = functools.partial(
GuardBuilder.TENSOR_MATCH, value=TensorWeakRef(value[i])
)
guards.append(source_i.make_guard(guard))
install_guard(*guards, skip=1)
grapharg = GraphArg(
source,
value,
pass_arg_as_tensor=False,
fake_tensor=None,
is_tensor=False,
)
tensor_list_proxy.node.meta["grapharg"] = grapharg
result = BaseListVariable.cls_for_instance(value)(
output, mutable_local=MutableLocal()
)
if istype(value, list):
return self.set_source_and_track_mutable(value, result)
return result
def wrap_tuple_iterator(self, value: tuple_iterator):
self.install_guards(GuardBuilder.TUPLE_ITERATOR_LEN)
output = [
VariableBuilder(self.tx, TupleIteratorGetItemSource(self.get_source(), i))(
tuple_iterator_getitem(value, i)
)
for i in range(tuple_iterator_len(value))
]
result = TupleIteratorVariable(
output, mutable_local=MutableLocal(), source=self.source
)
return self.set_source_and_track_mutable(value, result)
def wrap_slice_range(self, value: Union[slice, range]):
items = [
VariableBuilder(self.tx, AttrSource(self.get_source(), k))(
getattr(value, k)
)
for k in ("start", "stop", "step")
]
self.install_guards(GuardBuilder.TYPE_MATCH)
if isinstance(value, slice):
return SliceVariable(items, source=self.source)
else:
return RangeVariable(items, source=self.source)
def wrap_module(self, value: torch.nn.Module):
from ..eval_frame import OptimizedModule
if len(value.__dict__) == 0:
unimplemented(f"uninitialized nn.Module: {typestr(value)}")
if istype(value, OptimizedModule):
# Check if the optimized module was disabled
if inspect.getattr_static(value.forward, "_torchdynamo_disable", False):
# This bytecode is mostly of kind LOAD_ATTR or LOAD_METHOD. If
# we graph break here, Dynamo does not know how to create
# continuation functions for such bytecodes. So, we delay the
# graph break to CALL_FUNCTION.
return DelayGraphBreakVariable(source=self.source)
self.install_guards(GuardBuilder.TYPE_MATCH)
self.source = AttrSource(self.source, "_orig_mod")
return self.wrap_module(value._orig_mod)
if (
isinstance(value, (torch.nn.RNN, torch.nn.GRU, torch.nn.LSTM))
and not config.allow_rnn
):
unimplemented("TorchDynamo purposely graph breaks on RNN, GRU, LSTMs")
if getattr(value, "_is_fsdp_managed_module", False):
# See note [Dynamo treats FSDP wrapped modules as UnspecializedNNModule]
# in fully_sharded_data_parallel.py for more information
# we can't do this assert inside FSDP constructor,
# since we don't know yet whether dynamo will be used
assert getattr(
value, "_fsdp_use_orig_params", False
), "Dynamo only supports FSDP with use_orig_params=True"
# Note on FSDP guarding
# Eager FSDP already assumes (requires, but without enforcement)
# that users don't mutate their model parameters/structure after
# FSDP wrapping, because FSDP wouldn't notice or update its
# FlatParams.
#
# Therefore, torch.compile can skip guarding on params or submodule
# structure of fsdp_managed modules, by using FSDPNNModuleSource as
# the guard source. This behavior is gated on
# config.skip_fsdp_guards.
self.install_guards(GuardBuilder.TYPE_MATCH)
result = FSDPManagedNNModuleVariable(value, source=self.get_source())
if not SideEffects.cls_supports_mutation_side_effects(type(value)):
# don't allow STORE_ATTR mutation with custom __setattr__
return result
return self.tx.output.side_effects.track_object_existing(value, result)
elif mutation_guard.is_dynamic_nn_module(value, self.tx.export):
# created dynamically, don't specialize on it
self.install_guards(GuardBuilder.TYPE_MATCH)
if (
torch._dynamo.config.inline_inbuilt_nn_modules
and torch._inductor.config.freezing
and not torch.is_grad_enabled()
):
from ..decorators import mark_static_address
for p in value.parameters():
mark_static_address(p)
for b in value.buffers():
mark_static_address(b)
# we need to add the module to tracing context
# in order to allow its params to get invalidated
# this will get cleaned up once compile ends
self.tx.output.nn_modules[self.name] = value
if value.__module__.startswith(("torch.nn.", "torch.ao.")):
result = UnspecializedBuiltinNNModuleVariable(value, source=self.source)
else:
result = UnspecializedNNModuleVariable(value, source=self.source)
if not SideEffects.cls_supports_mutation_side_effects(type(value)):
# don't allow STORE_ATTR mutation with custom __setattr__
return result
return self.tx.output.side_effects.track_object_existing(value, result)
elif issubclass(
value.__class__, torch.nn.parallel.distributed.DistributedDataParallel
):
self.install_guards(GuardBuilder.TYPE_MATCH)
return UnspecializedNNModuleVariable(value, source=self.get_source())
else:
return self.tx.output.register_attr_or_module(
value,
self.name,
source=self.get_source(),
# Guards are added inside register_attr_or_module
)
def wrap_literal(self, value):
if not config.specialize_int and type(value) is int:
# unspecializing int by default, but still
# specialize for the following conditions
if not TracingContext.get().force_unspec_int_unbacked_size_like and (
# Assume integers from global variables want to be specialized
not self.source.guard_source().is_local()
# Assume that integers that came from NN modules want to be
# specialized (as we don't expect users to be changing the
# NN modules on the fly)
or self.source.guard_source().is_nn_module()
or is_from_defaults(self.source)
or is_cell_contents(self.source)
# TODO: Delete this condition when rollout is done. NB: this
# condition never evaluates True in open source
or (
not justknobs_check(
"pytorch/dynamo:enable_unspecialize_zero_one_plain_int"
)
and value in self._common_constants()
)
):
self.install_guards(GuardBuilder.CONSTANT_MATCH)
return ConstantVariable.create(value=value, source=self.source)
else:
return self.wrap_symint(value)
elif not config.specialize_float and type(value) is float:
return self.wrap_symfloat(value)
else:
self.install_guards(GuardBuilder.CONSTANT_MATCH)
return ConstantVariable.create(value=value)
def assert_not_wrapped_by_this_graph(self, value: torch.Tensor):
if is_fake(value) and maybe_get_fake_mode(value) is self.tx.fake_mode:
raise InternalTorchDynamoError(
"Cannot wrap a Tensor that has already been",
"wrapped by this instance of Dynamo",
)
def wrap_tensor(self, value: torch.Tensor):
source = self.get_source()
# We cannot already be tracking the tensor, which implies
# it would have already been wrapped
assert value not in self.tx.output.side_effects
if (
source.guard_source().is_nn_module()
or get_static_address_type(value) is not None
) and not source.guard_source().is_fsdp_module():
self.assert_not_wrapped_by_this_graph(value)
return self.tx.output.register_attr_or_module(
value, self.name, source=source
)
if is_constant_source(source):
self.assert_not_wrapped_by_this_graph(value)
return self.tx.output.register_attr_or_module(
value,
re.sub(r"[^a-zA-Z0-9]+", "_", self.name),
source=source,
# Guards are added inside register_attr_or_module
)
if type(value) in config.traceable_tensor_subclasses:
# Ordinarily, we would fakeify a tensor so that it can get dynamic
# shapes and be computed on without triggering actual operations.
# However, how can we fakeify a tensor subclass? Ordinary
# inheritance (nor multiple inheritance) won't work work.
#
# Instead, our plan is to *manually simulate* the tensor subclass
# inheriting from a fake tensor with dynamo. This means our
# data representation for a tensor subclass will be a fake tensor
# + tensor subclass type + any extra data the subclass may have
# been storing on the tensor. Because all Python accesses are
# mediated through TensorWithTFOverrideVariable, we can ensure
# that we dispatch differently, e.g., according to
# __torch_function__
#
# To simplify things for now, the __dict__ tracking bits haven't
# been implemented yet, but they can be added into this design at
# a later point in time.
subclass_type = type(value)
else:
assert type(value) in (
torch.Tensor,
torch.nn.Parameter,
torch._subclasses.fake_tensor.FakeTensor,
torch._subclasses.functional_tensor.FunctionalTensor,
) or is_traceable_wrapper_subclass(value), type(value)
subclass_type = None
# NB: this just says we accessed a tensor from the same source again
# (e.g., a tensor lives in a global foo, and we LOAD_GLOBAL it twice).
# This is distinct from two distinct sources mapping to the same
# Tensor (per id())! No guard is necessary here. See below for the
# other case.
is_duplicate_tensor = source in self.tx.output.input_source_to_var
if is_duplicate_tensor:
return self.tx.output.input_source_to_var[source]
# By this point, we should have deduplicated all tensors
self.assert_not_wrapped_by_this_graph(value)
# tx.output has multiple tracers if we're introspecting HigherOrderOperator.
# When we've discovered an untracked tensor, then we actually need
# to get Dynamo to track the tensor (which is what this function does)
# and put it as a graph input on the root tracer. Later on,
# if the input is actually used in the body of the HigherOrderOperator,
# then the relevant SubgraphTracer will lift it to being an input of
# the subgraph.
# See NOTE [HigherOrderOperator tracing design] for more details.
tensor_proxy = self.tx.output.root_tracer.create_graph_input(
re.sub(r"[^a-zA-Z0-9]+", "_", self.name), type(value), source=source
)
options = {}
if type(value) in config.traceable_tensor_subclasses:
options["torch_function_fn"] = build_torch_function_fn(
self.tx, value, self.source
)
self.install_guards(GuardBuilder.TYPE_MATCH)
if (
isinstance(value, torch.Tensor)
and value.is_nested
and not isinstance(value, torch.nested._internal.nested_tensor.NestedTensor)
):
unimplemented("torch.compile does not support strided NestedTensor")
# Reject sparse, but not coo.
# TODO: remove this altogether when non-coo sparsity propagation is ready
if is_sparse_any(value) and not value.is_sparse:
unimplemented(
f"torch.compile does not support sparse Tensor with {value.layout} layout"
)
# TODO(pearu,sparse-team) - Add the corresponding SPARSE_TENSOR_MATCH guards
if is_sparse_any(value) and value.is_sparse and not self.tx.export:
# A hot fix for sparse tensors + torch.compile. There is some
# support for export + coo tensor. We need to create
# SPARSE_TENSOR_GUARDS for guards to work propertly.
unimplemented("torch.compile does not support sparse Tensors")
tensor_variable = wrap_fx_proxy(
tx=self.tx,
proxy=tensor_proxy,
example_value=value,
subclass_type=subclass_type,
source=source,
**options,
)
guard_type = GuardBuilder.TENSOR_MATCH
if isinstance(source, GradSource) and is_from_optimizer_source(source):
guard_type = GuardBuilder.NOT_NONE_MATCH
self.install_guards(
functools.partial(
guard_type,
value=value
if isinstance(source, NumpyTensorSource)
else TensorWeakRef(value),
)
)
# We install TYPE_MATCH guards for traceable wrapper subclass object,
# and recursively install corresponding guard for each inner attribute.
if is_traceable_wrapper_subclass(value):
self.install_guards(GuardBuilder.TYPE_MATCH)
install_guard(
SubclassAttrListSource(source).make_guard(GuardBuilder.EQUALS_MATCH)
)
attrs, _ = value.__tensor_flatten__()
for attr in attrs:
inner_value = getattr(value, attr)
inner_source = AttrSource(self.source, attr)
LazyVariableTracker.realize_all(
VariableBuilder(self.tx, inner_source)(inner_value)
)
self.tx.output.input_source_to_var[source] = tensor_variable
assert "tensor_dict" not in tensor_proxy.node.meta
tensor_proxy.node.meta["tensor_dict"] = value.__dict__.copy()
# Note: this information is conveyed via subclass_type now
fake_tensor_value = tensor_variable.proxy.node.meta["example_value"]
if maybe_get_fake_mode(fake_tensor_value) is not self.tx.fake_mode:
raise InternalTorchDynamoError("Wrapped Tensor must be this graph's fake")
grapharg = GraphArg(source, value, False, fake_tensor_value)
tensor_proxy.node.meta["grapharg"] = grapharg
self.tx.output.add_symbol_bindings(grapharg)
return tensor_variable
def wrap_numpy_ndarray(self, value):
assert np is not None
assert isinstance(value, np.ndarray)
source = NumpyTensorSource(self.get_source())
from torch._numpy import _util
readonly = not value.flags.writeable
if readonly:
try:
value.flags.writeable = True
except ValueError:
# One can not easily make nditer elements writable,
# but warning is not the end of the world
assert isinstance(value.base, np.nditer)
pass
try:
tensor_value = _util._try_convert_to_tensor(value)
if readonly:
from torch._prims_common import clone_preserve_strides
tensor_value = clone_preserve_strides(tensor_value)
except NotImplementedError as e:
# failed to convert to tensor, graph break
unimplemented(str(e))
# We do this because we want the full behavior of guarding the numpy ndarray as if it were
# a tensor. It's a little annoying to make a VT to throw out, but there's so many side effects here
# that there's not another great way to do this atm.
# This creates the right graphargs, as well as registration for guards in tensor names and shape env.
LazyVariableTracker.realize_all(VariableBuilder(self.tx, source)(tensor_value))
proxy = self.tx.output.root_tracer.create_graph_input(
re.sub(r"[^a-zA-Z0-9]+", "_", self.name), type(tensor_value), source=source
)
options = {"source": source}
numpy_ndarray_variable = wrap_fx_proxy_cls(
target_cls=NumpyNdarrayVariable,
tx=self.tx,
proxy=proxy,
example_value=tensor_value,
**options,
)
self.tx.output.input_source_to_var[source] = numpy_ndarray_variable
example_value = numpy_ndarray_variable.proxy.node.meta["example_value"]
# pass_arg_as_tensor should be true because we are wrapping a np.ndarray as argument input, and it needs to be
# converted to a tensor.
grapharg = GraphArg(
source,
tensor_value,
pass_arg_as_tensor=True,
fake_tensor=example_value,
is_tensor=True,
example_strong_ref=tensor_value,
)
proxy.node.meta["grapharg"] = grapharg
return numpy_ndarray_variable
def wrap_symint(self, value):
assert type(value) is int
if self.name in self.tx.output.unspec_variable_map:
return self.tx.output.unspec_variable_map[self.name]
shape_env = self.tx.output.shape_env
if TracingContext.get().force_unspec_int_unbacked_size_like:
wrapped_value = shape_env.create_unbacked_symint()
_constrain_range_for_size(wrapped_value)
self.tx.output.bound_symbols.add(wrapped_value.node.expr)
self.tx.output.tracked_fakes.append(
TrackedFake(wrapped_value, self.source, None)
)
# NB: We do not do float. For motivation, see
# https://docs.google.com/document/d/1INSCdYu1PxXcr43HrD82OudeEuS-qxQe1yZmLg2wy6A/edit
# but the general idea is that we generate kernels that can
# take unspecialized floats and use them in sizevar computation
elif not is_constant_source(self.get_source()):
if torch._dynamo.config.specialize_int:
# If specialize_int is False, also return
# a constant (but this should have been handled
# in the caller, TBH)
self.install_guards(GuardBuilder.CONSTANT_MATCH)
return ConstantVariable.create(value=value, source=self.source)
name = self.source.name()
if name not in self.tx.output.frame_state:
# Note - this essentially means that if this name gets reused as a tensor,
# it will start fully dynamic. That should always be a safe option, and not awfully inefficient.
# Alternatively, if we want to improve pef here, we can add a third state of unset, but I am not
# sure that is necessary for now.
frame_state_entry = FrameStateSizeEntry(scalar=value, size=None)
else:
frame_state_entry = self.tx.output.frame_state[name]
if frame_state_entry.scalar != value:
log.debug(
"automatic dynamic int %s val %s != %s",
name,
value,
frame_state_entry.scalar,
)
frame_state_entry.scalar = None
self.tx.output.frame_state[name] = frame_state_entry
# TODO: This should be dynamic, as we in general do not
# know if bare integers are actually going to be sizevars
# and it is inappropriate to eagerly duck size them with
# real sizevars
if (
config.automatic_dynamic_shapes and frame_state_entry.scalar is None
) or not config.assume_static_by_default:
dynamic_dim = DimDynamic.DYNAMIC
else: # assume_static_by_default
# TODO: dynamic_dim = DimDynamic.STATIC should work but
# for some reason it doesn't
self.install_guards(GuardBuilder.CONSTANT_MATCH)
return ConstantVariable.create(value=value)
wrapped_value = shape_env.create_unspecified_symint_and_symbol(
value,
source=self.source,
dynamic_dim=dynamic_dim,
)
self.tx.output.bound_symbols.add(wrapped_value.node.expr)
self.tx.output.tracked_fakes.append(
TrackedFake(wrapped_value, self.source, None)
)
else:
assert is_constant_source(self.get_source())
# TODO: Do I actually need guard for constant source?
self.install_guards(GuardBuilder.CONSTANT_MATCH)
return ConstantVariable.create(value=value, source=self.source)
assert not isinstance(self.get_source(), RandomValueSource)
install_guard(self.get_source().make_guard(GuardBuilder.TYPE_MATCH))
options = {"source": self.get_source()}
proxy = self.tx.output.root_tracer.create_graph_input(
re.sub(r"[^a-zA-Z0-9]+", "_", self.name),
type(wrapped_value),
source=self.get_source(),
)
set_example_value(proxy.node, wrapped_value)
unspec_var = SymNodeVariable(proxy, wrapped_value, **options)
self.tx.output.unspec_variable_map[self.name] = unspec_var
if not is_constant_source(self.get_source()):
if self.tx.export and not isinstance(self.get_source(), LocalSource):
raise AssertionError(
f"Dynamo attempts to add additional input during export: value={wrapped_value}, source={self.get_source()}"
)
example_value = unspec_var.proxy.node.meta["example_value"]
proxy.node.meta["grapharg"] = GraphArg(
self.get_source(),
wrapped_value,
pass_arg_as_tensor=False,
fake_tensor=None,
is_tensor=False,
example_strong_ref=wrapped_value,
)
return unspec_var
def wrap_symfloat(self, value):
# SymFloat wrapping is special. We first wrap it in the same way we
# do an unspecialized primitive, and then we item() it into a
# SymFloat. Removal of the item() call is left to a later FX pass,
# mostly because that pass is more easily done after we have lowered
# to ATen ops. (Dynamo doesn't do decomposition right now).
if self.name in self.tx.output.unspec_variable_map:
return self.tx.output.unspec_variable_map[self.name]
# NB: we specialize on nan input, because our guard modeling in
# ShapeEnv cannot deal with nan
if (
torch._dynamo.config.specialize_float
or is_constant_source(self.get_source())
or math.isnan(value)
):
self.install_guards(GuardBuilder.CONSTANT_MATCH)
return ConstantVariable.create(value=value, source=self.source)
# NB: At the point we've gotten here, we don't assume static by
# default. Since we have a guard mechanism, there isn't really any
# downside to trying to be dynamic for float all the time. Unlike
# ints, this won't make codegen perf worse. Modest cost to compile
# time.
wrapped_value = torch.tensor(value, dtype=torch.float64)
# TODO: Switch RandomValueSource over to use this, this is more
# accurate
assert not isinstance(self.get_source(), RandomValueSource)
install_guard(self.get_source().make_guard(GuardBuilder.TYPE_MATCH))
# The FloatTensorSource here is just for pedantic correctness: if you
# guard against an UnspecializedPythonVariable, you need to guard
# against the tensor-ified version of the local, otherwise it's not a
# Tensor. However, we never let the UnspecializedPythonVariable escape
# here, so there should never actually be any guards against this
# source.
options = {"source": FloatTensorSource(self.get_source()), "raw_value": value}
# TODO: Maybe the tensor-ification should be built into the source,
# rather than by special pattern match
proxy = self.tx.output.root_tracer.create_graph_input(
re.sub(r"[^a-zA-Z0-9]+", "_", self.name),
type(wrapped_value),
source=self.get_source(),
)
unspec_var = wrap_fx_proxy_cls(
UnspecializedPythonVariable,
tx=self.tx,
proxy=proxy,
example_value=wrapped_value,
**options,
)
assert isinstance(unspec_var, UnspecializedPythonVariable)
self.tx.output.unspec_variable_map[self.name] = unspec_var
if self.tx.export and not isinstance(self.get_source(), LocalSource):
raise AssertionError(
f"Dynamo attempts to add additional input during export: value={wrapped_value}, source={self.get_source()}"
)
fake_tensor_value = None
example_value = unspec_var.proxy.node.meta["example_value"]
assert is_fake(example_value)
fake_tensor_value = example_value
assert fake_tensor_value.fake_mode is self.tx.fake_mode, (
f"fake mode ({fake_tensor_value.fake_mode}) from fake tensor metadata doesn't match mode"
"({self.tx.fake_mode}) from InstructionTranslator"
)
# There's something a bit incoherent about pass_arg_as_tensor,
# specifically regarding sources.
#
# Specifically, suppose we have "x: float" local argument. We
# eventually end up with an UnspecializedPythonVariable denoting
# torch.as_tensor(x)... but it's source is still L['x'] (which if you
# accessed it directly is a float!) So you gotta be careful when
# setting up your guards, because it's still going to be a float at
# this point, the conversion happens only precisely at the point we're
# actually calling the FX graph. This happens to be what we want for
# shape guard generation, but it's kind of unintuitive.
proxy.node.meta["grapharg"] = GraphArg(
self.get_source(),
wrapped_value,
pass_arg_as_tensor=True,
fake_tensor=fake_tensor_value,
is_tensor=False,
example_strong_ref=wrapped_value,
)
# Directly do item to bypass capture_scalar_outputs
r = wrap_fx_proxy(
self.tx,
self.tx.output.create_proxy(
"call_method",
"item",
*proxy_args_kwargs([unspec_var], {}),
),
)
self.tx.output.tracked_fakes.append(TrackedFake(r.sym_num, self.source, None))
return r
def wrap_unspecialized_primitive(self, value):
if self.name in self.tx.output.unspec_variable_map:
return self.tx.output.unspec_variable_map[self.name]
wrapped_value = torch.tensor(value)
if not isinstance(self.get_source(), RandomValueSource):
install_guard(self.get_source().make_guard(GuardBuilder.TYPE_MATCH))
options = {"source": self.get_source()}
options.update({"raw_value": value})
proxy = self.tx.output.root_tracer.create_graph_input(
re.sub(r"[^a-zA-Z0-9]+", "_", self.name),
type(wrapped_value),
source=self.get_source(),
)
unspec_var = wrap_fx_proxy_cls(
UnspecializedPythonVariable,
tx=self.tx,
proxy=proxy,
example_value=wrapped_value,
**options,
)
self.tx.output.unspec_variable_map[self.name] = unspec_var
if not is_constant_source(self.get_source()):
if self.tx.export and not isinstance(self.get_source(), LocalSource):
raise AssertionError(
f"Dynamo attempts to add additional input during export: value={wrapped_value}, source={self.get_source()}"
)
fake_tensor_value = None
if isinstance(unspec_var, ConstantVariable):
# TODO: when can this happen?
example_value = unspec_var.value
else:
example_value = unspec_var.proxy.node.meta["example_value"]
assert is_fake(example_value)
fake_tensor_value = example_value
assert fake_tensor_value.fake_mode is self.tx.fake_mode, (
f"fake mode ({fake_tensor_value.fake_mode}) from fake tensor metadata doesn't match mode"
"({self.tx.fake_mode}) from InstructionTranslator"
)
proxy.node.meta["grapharg"] = GraphArg(
self.get_source(),
wrapped_value,
pass_arg_as_tensor=True,
fake_tensor=fake_tensor_value,
is_tensor=False,
example_strong_ref=wrapped_value,
)
return unspec_var
def _dataclasses_fields_lambda(obj):
if isinstance(obj, UserDefinedObjectVariable):
value = obj.value
elif isinstance(obj, CustomizedDictVariable):
value = obj.user_cls
else:
unimplemented(f"Dataclass fields handling fails for type {obj}")
items = []
for field in dataclasses.fields(value):
source = None
if obj.source:
source = GetItemSource(
AttrSource(obj.source, "__dataclass_fields__"), field.name
)
items.append(UserDefinedObjectVariable(field, source=source))
return TupleVariable(items)
def wrap_fx_proxy(
tx, proxy, example_value=None, subclass_type=None, **options
) -> VariableTracker:
kwargs = {
"tx": tx,
"proxy": proxy,
"example_value": example_value,
"subclass_type": subclass_type,
**options,
}
if subclass_type is None:
return wrap_fx_proxy_cls(target_cls=TensorVariable, **kwargs)
else:
result = wrap_fx_proxy_cls(target_cls=TensorWithTFOverrideVariable, **kwargs)
result.install_global(tx)
return result
# Note: Unfortunate split due to some gross classes existing that subclass TensorVariable
# Should be compositional instead
#
# This is a horribly complicated function that does too many things, to
# explain what it does, let's first talk about the classic usage wrap_fx_proxy
# for a TensorVariable. There are two primary modes of use:
#
# 1. Wrapping a pre-existing Tensor. In this case, example_value is set
# to the pre-existing Tensor. (Note that this example_value will NOT
# be the final example_value we put into node.meta['example_value'],
# instead it is converted into a fake tensor using
# wrap_to_fake_tensor_and_record and registered as a graph input.)
#
# 2. "Wrapping" the result of some Tensor operation Dynamo traced over. In
# this case, example_value is None (and we are going to figure it out
# ourselves using FakeTensors, via get_fake_value, which will run
# the operation represented by the (singular!) FX node referenced by
# the passed in proxy.)
#
# The expectation is you end up with a Tensor output, and everything is
# straightforwardly traced into the graph.
#
# In all cases, the returned `TensorVariable` subclass will have an `example_value`
# and that `example_value` must be a `FakeTensor` produced by the currently running
# instance of Dynamo.
#
# Upon closer inspection, you may notice that there are a slurry of non-Tensor
# output cases. What gives? Well, we sometimes trace operations into the
# graph that don't involve tensors.
#
# * Some operators return tuples; we need to recursively handle their
# contents
#
# * Some operators have side effects that will affect subsequent AOTAutograd
# tracing but don't otherwise return anything.
#
# * Some operators return symbolic ints/floats/bools which can go in the
# graph and be traced (but only if they're actually symbolic! If they're
# static you don't want to put them in the graph, which means you
# shouldn't call this function.)
#
# The common theme is that you only use this function WHEN YOU ARE TRACING
# SOMETHING INTO THE GRAPH. This is sort of obvious, because you can't call
# this function without a proxy.
def wrap_fx_proxy_cls(
target_cls, tx, proxy, example_value=None, subclass_type=None, **options
):
from ..symbolic_convert import InstructionTranslatorBase
assert isinstance(tx, InstructionTranslatorBase)
if "guards" in options and options["guards"] is not None:
tx.output.guards.update(options["guards"])
assert "example_value" not in proxy.node.meta, f"{proxy.node.meta['example_value']}"
initial_example_value = example_value
def _clone_input(value):
if isinstance(value, torch.Tensor):
# tensor subclasses will not be converted to FakeTensors and need to be cloned
if not (
isinstance(value, FakeTensor)
or (
# Is functional tensor fakeified by this instance of Dynamo
torch._is_functional_tensor(value)
and maybe_get_fake_mode(value) is tx.fake_mode
)
or value.is_nested
):
# NB: ensure strides are preserved
value = clone_input(value)
return value
# See NOTE: [Deferring tensor pack/unpack hooks until runtime]
with torch._dynamo.utils._disable_saved_tensors_hooks_during_tracing():
# with preserve_rng_state():
if example_value is None:
# only allow_non_graph_fake in this instance because we handle the non-fake
# cases properly below.
example_value = get_fake_value(proxy.node, tx, allow_non_graph_fake=True)
# Handle recursive calls here
elif maybe_get_fake_mode(example_value) is tx.fake_mode:
pass
elif isinstance(example_value, torch.Tensor):
if tx.export:
# The legacy behavior for real value cache with subclasses was
# to perform a clone WITHOUT preserving the subclass. It's
# not entirely clear this is what you actually want though.
with torch._C.DisableTorchFunctionSubclass():
proxy.tracer.real_value_cache[proxy.node] = _clone_input(
example_value
)
# NB: If we're ignoring subclass, then the expectation is you will
# take the returned TensorVariable and wrap it into a more
# accurate TensorVariable that is able to track subclass-ness;
# otherwise this is wrong!
kwargs = {
"is_tensor": target_cls
in (TensorVariable, TensorWithTFOverrideVariable),
}
assert "source" in options and options["source"] is not None
kwargs["source"] = options["source"]
example_value = wrap_to_fake_tensor_and_record(
example_value, tx=tx, **kwargs
)
if (
isinstance(example_value, torch.Tensor)
and example_value.device.type != "meta"
and (maybe_get_fake_mode(example_value) is not tx.fake_mode)
):
raise InternalTorchDynamoError(
"`example_value` needs to be a `FakeTensor`"
f"wrapped by this instance of Dynamo. Found: {example_value}"
)
if isinstance(example_value, torch.Tensor):
is_parameter = isinstance(example_value, torch.nn.Parameter)
# NB: In most (all?) cases, this does not actually do a clone.
# (WARNING: this means that if we mutate metadata on the fake
# tensor, the stored example value will update too!)
example_value = _clone_input(example_value)
set_example_value(proxy.node, example_value)
specialized_props = target_cls.specialize(example_value)
# TODO: not sure about this fake mode test
if (
isinstance(example_value, torch._subclasses.fake_tensor.FakeTensor)
and example_value.fake_mode is tx.fake_mode
):
tensor_type = subclass_type if subclass_type else torch.Tensor
specialized_props["class_type"] = (
torch.nn.Parameter if is_parameter else tensor_type
)
options.update(specialized_props)
return target_cls(proxy, **options)
elif (
hasattr(proxy.node.target, "__name__")
and proxy.node.target.__name__ == "set_state"
and isinstance(proxy.node.target.__self__, torch._C.Generator)
or proxy.node.target == torch.random.set_rng_state
):
return TorchInGraphFunctionVariable(proxy.node.target)
elif (
proxy.node.target == torch._C._DisableFuncTorch
or proxy.node.target == torch.cuda._is_in_bad_fork
):
return UserDefinedObjectVariable(example_value)
elif istype(example_value, torch.Size) and all(
isinstance(x, int) for x in example_value
):
sizes = [ConstantVariable.create(x) for x in example_value]
return SizeVariable(sizes, **options)
elif isinstance(example_value, (tuple, list)):
set_example_value(proxy.node, example_value)
unpacked = []
for i, val in enumerate(example_value):
if val is None:
# nn.MultiheadAttention() can return None, see issue #175
unpacked.append(
ConstantVariable.create(None, **options),
)
else:
proxy_i = proxy.tracer.create_proxy(
kind="call_function",
target=operator.getitem,
args=(proxy, i),
kwargs={},
)
if "source" in options:
source = options["source"]
options_i = options.copy()
options_i["source"] = GetItemSource(
base=source, index=i, index_is_slice=False
)
else:
# use the same options object as parent
options_i = options
# WARNING: this assumes the same target_cls as this tuple/list call
unpacked.append(
wrap_fx_proxy_cls(
target_cls=target_cls,
tx=tx,
proxy=proxy_i,
example_value=val,
**options_i,
)
)
if isinstance(example_value, torch.Size):
# NB: Keep the old proxy around. See SizeVariable for an
# explanation why
return SizeVariable(unpacked, proxy, **options)
elif istype(example_value, tuple):
return TupleVariable(unpacked, **options)
elif istype(example_value, (list, immutable_list)):
return ListVariable(unpacked, mutable_local=MutableLocal(), **options)
else:
assert example_value.__class__.__module__ == "torch.return_types" or hasattr(
example_value, "_fields"
), f"expected {example_value.__class__.__module__} == torch.return_types or named tuple but got {type(example_value)}"
return NamedTupleVariable(unpacked, example_value.__class__, **options)
elif example_value is None or proxy.node.target is torch.manual_seed:
return ConstantVariable.create(None, **options)
elif isinstance(example_value, (torch.SymInt, torch.SymFloat, torch.SymBool)):
set_example_value(proxy.node, example_value)
return SymNodeVariable(proxy, example_value, **options)
elif (
inspect.isclass(proxy.node.target)
and issubclass(proxy.node.target, _StreamBase)
) or proxy.node.target in [
device_interface.current_stream
for _, device_interface in get_registered_device_interfaces()
]:
set_example_value(proxy.node, example_value)
return StreamVariable(proxy, example_value, example_value.device, **options)
elif (
inspect.isclass(proxy.node.target) and issubclass(proxy.node.target, _EventBase)
) or proxy.node.target in [
device_interface.Event
for _, device_interface in get_registered_device_interfaces()
]:
set_example_value(proxy.node, example_value)
return EventVariable(proxy, example_value, **options)
elif proxy.node.target == "query" and proxy.node.op == "call_method":
set_example_value(proxy.node, example_value)
return ConstantVariable(example_value, **options)
elif (
example_value is not None
and isinstance(example_value, _EventBase)
and proxy.node.target == "record_event"
and proxy.node.op == "call_method"
):
set_example_value(proxy.node, example_value)
return EventVariable(proxy, example_value, **options)
elif isinstance(example_value, int) and proxy.node.target in [
torch.sym_int,
getattr,
operator.getitem,
torch._utils._element_size,
torch.seed,
operator.mod,
torch._functorch.vmap._validate_and_get_batch_size,
# some mac builds are missing torch.distributed.get_rank()
getattr(torch.distributed, "get_rank", _missing),
getattr(torch.distributed, "get_world_size", _missing),
# This always wants to be in the graph, even if the constraint
# results in a constant int
torch._constrain_as_size,
]:
set_example_value(proxy.node, example_value)
return ConstantVariable.create(example_value, **options)
elif isinstance(example_value, torch.backends.cuda.SDPAParams):
from .sdpa import SDPAParamsVariable
set_example_value(proxy.node, example_value)
return SDPAParamsVariable(proxy, **options)
elif isinstance(example_value, bool) and proxy.node.target in [
torch.backends.cuda.can_use_flash_attention,
torch.backends.cuda.can_use_efficient_attention,
]:
set_example_value(proxy.node, example_value)
return ConstantVariable.create(example_value, **options)
elif (
isinstance(example_value, (int, float, bool))
and proxy.node.target is call_torchbind
):
set_example_value(proxy.node, example_value)
return ConstantVariable.create(example_value, **options)
else:
unimplemented(
"torch.* op returned non-Tensor "
+ f"{typestr(example_value)} {proxy.node.op} {proxy.node.target}"
)
# Tracks the sources of all fake tensors we wrap in Dynamo.
# Used by shape guard computation.
@dataclasses.dataclass
class TrackedFake:
fake: Union[FakeTensor, SymInt]
source: Source
# Is None when fake is SymInt
symbolic_context: Optional[SymbolicContext]
def __hash__(self) -> int:
return hash((self.fake, self.source.name()))
def __eq__(self, other: object) -> bool:
if isinstance(other, TrackedFake):
return self.fake is other.fake and self.source.name() == other.source.name()
return False
# Performs automatic dynamic dim determination.
# Returns a SymbolicContext
def _automatic_dynamic(
e, tx, source, static_shapes, outer_only=False
) -> SymbolicContext:
# strided NT not supported
if e.is_nested and not isinstance(
e, torch.nested._internal.nested_tensor.NestedTensor
):
unimplemented("torch.compile does not support strided NestedTensor")
name = source.name()
prior_policy = tx.output.tracing_context.tensor_to_context.get(e, None)
shape_env_to_source_to_symbol_cache = (
prior_policy.shape_env_to_source_to_symbol_cache if prior_policy else None
)
# Get base context if the tensor is a view
view_base_context: Optional[SymbolicContext] = None
if e._is_view():
base_source = AttrSource(source, "_base")
view_base_context = _automatic_dynamic(e._base, tx, base_source, static_shapes)
if is_traceable_wrapper_subclass(e) and not outer_only:
# Get symbolic context for outer tensor
outer_context = _automatic_dynamic(
e, tx, source, static_shapes, outer_only=True
)
# Get symbolic contexts for inner tensors
inner_contexts = {} # mapping from attr -> symbolic context
attrs, _ = type(e).__tensor_flatten__(e)
for attr in attrs:
inner_tensor = getattr(e, attr)
inner_source = AttrSource(source, attr)
inner_contexts[attr] = _automatic_dynamic(
inner_tensor, tx, inner_source, static_shapes
)
return SubclassSymbolicContext(
dynamic_sizes=outer_context.dynamic_sizes,
constraint_sizes=outer_context.constraint_sizes,
view_base_context=view_base_context,
tensor_source=outer_context.tensor_source,
shape_env_to_source_to_symbol_cache=outer_context.shape_env_to_source_to_symbol_cache,
inner_contexts=inner_contexts,
)
if static_shapes:
return StatefulSymbolicContext(
dynamic_sizes=[DimDynamic.STATIC] * e.dim(),
constraint_sizes=[None] * e.dim(),
view_base_context=view_base_context,
tensor_source=source,
shape_env_to_source_to_symbol_cache=shape_env_to_source_to_symbol_cache,
)
# We preserve the dynamism of inputs. For example, when users call
# make_fx(torch.cond, tracing_mode="symbolic")(*args), inputs have SymInt sizes.
from torch.fx.experimental.symbolic_shapes import is_nested_int
if any(isinstance(s, SymInt) and not is_nested_int(s) for s in e.size()):
return StatefulSymbolicContext(
dynamic_sizes=[
DimDynamic.DYNAMIC if isinstance(s, SymInt) else DimDynamic.STATIC
for s in e.size()
],
constraint_sizes=[None] * e.dim(),
view_base_context=view_base_context,
tensor_source=source,
shape_env_to_source_to_symbol_cache=shape_env_to_source_to_symbol_cache,
)
# Prep for automatic dynamic
frame_state_entry = None
if name not in tx.output.frame_state:
# If there is no entry for this source, add the tensor to frame state with its current static size.
# E.g., {} -> {"x": [2, 4]}
frame_state_entry = FrameStateSizeEntry(None, None)
frame_state_entry.size = list(e.size())
else:
frame_state_entry = tx.output.frame_state[name]
if frame_state_entry.size is not None:
if e.ndim != len(frame_state_entry.size):
# If there is already an entry, and the dim mismatches, replace the frame state entry with None.
# E.g. {"x": [2, 3, 4]} -> {"x": None}
log.debug(
"automatic dynamic %s dim %s != %s",
name,
e.ndim,
frame_state_entry.size,
)
frame_state_entry.size = None
else:
# If there is already an entry, and the dim matches, for every size in the frame state which
# disagrees with the current static size, replace it with None. E.g., {"x": [2, 3]} -> {"x": [2, None]}
for i, dim in enumerate(frame_state_entry.size):
if dim is not None and e.size()[i] != dim:
log.debug(
"automatic dynamic %s size(%s) %s != %s",
name,
i,
e.size(i),
dim,
)
frame_state_entry.size[i] = None
# TODO: index export_constraints ahead of time so we don't have to
# do a linear scan every time here
t_id = id(e)
dim2constraint = {}
def update_dim2constraint(dim, constraint_range, debug_name):
if dim in dim2constraint:
from torch.fx.experimental.symbolic_shapes import StrictMinMaxConstraint
old_constraint_range, old_debug_name = dim2constraint[dim]
new_constraint_range = StrictMinMaxConstraint(
vr=constraint_range.vr & old_constraint_range.vr,
warn_only=False,
)
# It is possible for (non-None) old_debug_name and debug_name to be different
# but this will only happen the corresponding Dims can be derived equal.
new_debug_name = old_debug_name or debug_name
dim2constraint[dim] = new_constraint_range, new_debug_name
else:
dim2constraint[dim] = constraint_range, debug_name
if tx.output.export_constraints:
for constraint in tx.output.export_constraints:
if constraint.t_id == t_id:
update_dim2constraint(
constraint.dim, constraint.constraint_range, constraint.debug_name
)
if constraint.shared is not None and constraint.shared.t_id == t_id:
# We process constraint ranges for each shared dimension separately
# so that we can directly check range constraint violations on them
# without looking up which other shared dimensions have this info.
# In other words, for this t_id, we will have processed all of its
# constraint ranges, no matter where / how they were specified, by
# by the end of this loop.
update_dim2constraint(
constraint.shared.dim,
constraint.constraint_range,
constraint.debug_name,
)
dynamic_dims = []
constraint_dims = []
for i in range(e.dim()):
# NB: mark dynamic has precedence over static
marked_unbacked = i in getattr(e, "_dynamo_unbacked_indices", set())
marked_dynamic = i in getattr(e, "_dynamo_dynamic_indices", set())
marked_weak_dynamic = i in getattr(e, "_dynamo_weak_dynamic_indices", set())
marked_static = i in getattr(e, "_dynamo_static_indices", set())
# NB: both static and dynamic have precedence over
automatic_dynamic = config.automatic_dynamic_shapes and (
frame_state_entry.size is None or frame_state_entry.size[i] is None
)
# Reflect the user directive in the frame_state
# For dynamic, apply None always
if frame_state_entry.size and marked_dynamic:
log.debug("automatic dynamic %s marked dynamic", name)
frame_state_entry.size[i] = None
# We will process constraints first, as they will imply that we
# have a dynamic dimension
# Precedence: export constraints > eager constraints
constraint = dim2constraint.get(i)
if constraint is None:
if marked_dynamic and not config.allow_ignore_mark_dynamic:
if hasattr(e, "_dynamo_dynamic_range"):
dim_range = [
dr for dr in e._dynamo_dynamic_range if dr.dim == i
].pop()
if dim_range.min is None and dim_range.max is None:
constraint_dim = RelaxedUnspecConstraint(warn_only=False)
else:
from torch.fx.experimental.symbolic_shapes import (
StrictMinMaxConstraint,
)
constraint_dim = StrictMinMaxConstraint(
vr=ValueRanges(lower=dim_range.min, upper=dim_range.max),
warn_only=False,
)
else:
constraint_dim = RelaxedUnspecConstraint(warn_only=False)
elif not marked_static and automatic_dynamic:
constraint_dim = RelaxedUnspecConstraint(warn_only=True)
else:
constraint_dim = None
else:
constraint_dim, debug_name = constraint
if debug_name is not None:
dim_name = f"{name}.size()[{i}]"
tx.output.shape_env.source_name_to_debug_name[dim_name] = debug_name
constraint_dims.append(constraint_dim)
# Now, figure out if the dim is dynamic/duck/static
if marked_unbacked:
dynamic = DimDynamic.SIZE_LIKE_UNBACKED
elif (
constraint_dim is not None
or marked_dynamic
or marked_weak_dynamic
or is_nested_int(e.shape[i])
):
# NB: We could assert static_shapes is False here, but it
# seems better to allow the user to override symbolic_context in this
# case
dynamic = DimDynamic.DYNAMIC
elif static_shapes or config.assume_static_by_default or marked_static:
dynamic = DimDynamic.STATIC
else:
dynamic = DimDynamic.DUCK
dynamic_dims.append(dynamic)
tx.output.frame_state[name] = frame_state_entry
return StatefulSymbolicContext(
dynamic_sizes=dynamic_dims,
constraint_sizes=constraint_dims,
view_base_context=view_base_context,
tensor_source=source,
shape_env_to_source_to_symbol_cache=shape_env_to_source_to_symbol_cache,
)
# See note [Tensor Fakification and Symbol Caching]
def wrap_to_fake_tensor_and_record(
e, tx, *, source: Optional[Source], is_tensor: bool, parent_context=None
):
if (
type(e) in (torch.Tensor, torch.nn.Parameter, FakeTensor)
or isinstance(e, torch.Tensor)
or is_traceable_wrapper_subclass(e)
):
assert source is not None
static_shapes, reason = tensor_always_has_static_shape(
e, is_tensor, guard_source=source.guard_source()
)
if not parent_context:
symbolic_context = _automatic_dynamic(e, tx, source, static_shapes)
else:
# Parent contexts are passed in when we are recursively creating
# fake tensors for subclasses. A better design would be not to create a
# parent/child relationship, but to recursively call _automatic_dynamic
# as we recursively call wrap_to_fake_tensor_and_record. This runs
# into bugs around how meta_utils knows and works to create fake tensors
# with tensor subclasses. Ideally, dynamo would drive both the recursive
# wrap_to_fake_tensor_and_record and _automatic_dynamic policy creation.
assert isinstance(source, AttrSource)
inner_context_name = source.member
symbolic_context = parent_context.inner_contexts[inner_context_name]
log.debug(
"wrap_to_fake %s %s %s %s",
source.name(),
tuple(e.shape),
symbolic_context,
type(e),
)
fake_e = wrap_fake_exception(
lambda: tx.fake_mode.from_tensor(
e,
source=source,
symbolic_context=symbolic_context,
)
)
if (
source is not None
and isinstance(fake_e, FakeTensor)
and (sym_val := fake_e.item_memo) is not None
):
tx.output.tracked_fakes.append(
TrackedFake(sym_val, CallMethodItemSource(source), symbolic_context)
)
if is_traceable_wrapper_subclass(fake_e):
attrs, _ = fake_e.__tensor_flatten__()
for attr in attrs:
fake_inner = getattr(fake_e, attr)
inner = getattr(e, attr)
inner_source = AttrSource(source, attr)
wrap_to_fake_tensor_and_record(
inner,
tx,
source=inner_source,
is_tensor=isinstance(fake_inner, torch.Tensor),
parent_context=symbolic_context,
)
tx.output.tracing_context.tensor_to_context[e] = symbolic_context
if is_sparse_any(fake_e):
# TODO: for TensorGuards, this eventually may need more
# fields for the size/stride of any other constituents
values = fake_e._values() if fake_e.is_sparse else fake_e.values()
tx.output.input_source_to_sizes_strides[source] = {
"size": fake_e.size(),
# TODO: revise this, but for now this stride instead of ()
# avoids SegFault with PYTORCH_TEST_WITH_DYNAMO=1
"stride": (1,) * fake_e.ndim,
"values_size": values.size(),
"values_stride": values.stride(),
}
else:
tx.output.input_source_to_sizes_strides[source] = {
"size": fake_e.size(),
"stride": fake_e.stride(),
}
if (
is_tensor
and not (static_shapes and source.is_nn_module())
and not is_constant_source(source)
):
tx.output.tracked_fakes.append(
TrackedFake(fake_e, source, symbolic_context)
)
tx.output.tracked_fakes_id_to_source[id(e)].append(source)
return fake_e
else:
return e
class SourcelessBuilder:
"""
Like builder, but stateless and does not require a source. Useful for simple type->VT objects, or objects
that are being created/evaporated during inlining (ex: consider a locally made list of tensors we then iterate over
.), such a list should not show up as an artifact from inputs, nor in reconstruction, nor in the graph. However,
there may be reasons to represent it as a ListVariable internally.
NOTE - Objects produced here are born UNGUARDED due to the nature of sources!
NOTE - This class is very new! It will have some rough edges, but it was created to stem the bleeding of giant
if/else type->VariableTracker trees that were cropping up all over dynamo.
"""
def __init__(self):
raise AssertionError("Use SourcelessBuilder.create()")
@staticmethod
def create(tx, value) -> VariableTracker:
value_type = type(value)
fast_handler = SourcelessBuilder._type_handlers.get(value_type)
if fast_handler:
return fast_handler(tx, value)
if isinstance(value, VariableTracker):
# This is always valid to call, and useful for recursive calls.
return value
elif isinstance(value, dataclasses._HAS_DEFAULT_FACTORY_CLASS):
return UserDefinedObjectVariable(value)
elif ConstantVariable.is_literal(value):
return ConstantVariable.create(value)
elif callable(value) and trace_rules.lookup_callable(value) is not None:
if is_callable_allowed(value):
tx.output.has_user_defined_allowed_in_graph = True
return trace_rules.lookup_callable(value)(value)
elif is_function_or_wrapper(value):
return trace_rules.lookup(value)(value)
elif isinstance(value, enum.Enum):
return EnumVariable(value)
elif isinstance(value, (type, abc.ABCMeta)):
return UserDefinedClassVariable(value)
elif isinstance(value, types.MethodWrapperType):
return MethodWrapperVariable(value)
elif isinstance(value, torch.fx.graph_module.GraphModule):
return SourcelessGraphModuleVariable(value)
elif isinstance(
value, (torch.utils._pytree.TreeSpec, torch.utils._pytree.LeafSpec)
):
return UserDefinedObjectVariable(value)
elif PlacementVariable.is_placement(value):
return PlacementVariable(value)
elif DeviceMeshVariable.is_device_mesh(value):
return DeviceMeshVariable(value)
elif isinstance(value, re.Pattern):
return RegexPatternVariable(value)
unimplemented(
f"Unexpected type in sourceless builder {value_type.__module__}.{value_type.__qualname__}"
)
@staticmethod
def wrap_constant_literal(value):
assert ConstantVariable.is_literal(value)
return ConstantVariable.create(value=value)
@staticmethod
def make_type_handlers():
create = SourcelessBuilder.create
handlers = {}
for t in common_constant_types:
handlers[t] = lambda tx, value: ConstantVariable(value)
handlers[set] = lambda tx, value: SetVariable(
[create(tx, x) for x in value], mutable_local=MutableLocal()
)
handlers[dict] = lambda tx, value: ConstDictVariable(
{create(tx, k): create(tx, v) for k, v in value.items()},
type(value),
mutable_local=MutableLocal(),
)
handlers[list] = lambda tx, value: ListVariable(
[create(tx, x) for x in value], mutable_local=MutableLocal()
)
handlers[tuple] = lambda tx, value: TupleVariable(
[create(tx, x) for x in value]
)
handlers[torch.Size] = lambda tx, value: SizeVariable(
[create(tx, x) for x in value]
)
handlers[collections.OrderedDict] = handlers[dict]
handlers[immutable_dict] = handlers[dict]
handlers[immutable_list] = handlers[list]
handlers[types.ModuleType] = lambda tx, value: PythonModuleVariable(value)
handlers[
torch.distributions.constraints._Real
] = lambda tx, value: UserDefinedObjectVariable(
value, mutable_local=MutableLocal()
)
handlers[
torch.distributions.constraints._Interval
] = lambda tx, value: UserDefinedObjectVariable(
value, mutable_local=MutableLocal()
)
handlers[
torch.distributions.constraints.Constraint
] = lambda tx, value: UserDefinedObjectVariable(
value, mutable_local=MutableLocal()
)
def passthrough(tx, value):
return value
for cls in VariableTrackerMeta.all_subclasses:
handlers[cls] = passthrough
return handlers
SourcelessBuilder._type_handlers = SourcelessBuilder.make_type_handlers()