import contextlib import functools import itertools import logging import os import weakref from dataclasses import dataclass from functools import partial from typing import Any, Callable, Dict, List, Optional, Tuple, Type, TypeVar, Union from weakref import ReferenceType import torch from torch._guards import Source from torch._ops import OpOverload from torch._prims_common import ( elementwise_dtypes, ELEMENTWISE_TYPE_PROMOTION_KIND, is_float_dtype, is_integer_dtype, ) from torch._subclasses.meta_utils import MetaConverter from torch.fx.operator_schemas import normalize_function from torch.multiprocessing.reductions import StorageWeakRef from torch.overrides import TorchFunctionMode from torch.utils._mode_utils import no_dispatch from torch.utils._python_dispatch import TorchDispatchMode from torch.utils._pytree import PyTree, tree_flatten, tree_map, tree_map_only from torch.utils._stats import count, count_label from torch.utils.weak import WeakIdRef log = logging.getLogger(__name__) pytree = torch.utils._pytree T = TypeVar("T") TensorWeakRef = Any aten = torch._ops.ops.aten CONSTANT_NUMEL_LIMIT = 1 RECURSION_COUNT = 0 # Small helper that increments recursion count, and # resets it when the object goes out of scope. Useful # if you don't want to increase indentation which is # what a context manager would do. class IncrementRecursionCount: def __init__(self): global RECURSION_COUNT RECURSION_COUNT += 1 def __del__(self): global RECURSION_COUNT RECURSION_COUNT -= 1 @dataclass class UnsupportedFakeTensorException(RuntimeError): reason: str @dataclass class DynamicOutputShapeException(RuntimeError): func: OpOverload @dataclass class DataDependentOutputException(RuntimeError): func: OpOverload _device_not_kwarg_ops = ( aten._resize_output_.default, aten._nested_tensor_from_tensor_list.default, aten._nested_tensor_from_tensor_list.out, aten.pin_memory.default, aten.is_pinned.default, aten.to.device, aten.to.prim_Device, aten._pin_memory.default, aten._pin_memory.out, aten._resize_output.default, aten._resize_output.out, ) # this op is never actually used _non_kwarg_device_constructors = (aten._list_to_tensor,) def contains_tensor_types(type): tensor_type = torch._C.TensorType.get() return type.isSubtypeOf(tensor_type) or any( contains_tensor_types(e) for e in type.containedTypes() ) _like_tensor_constructors = ( aten.empty_like.default, aten.empty_like.out, aten.full_like.default, aten.full_like.out, aten.ones_like.default, aten.ones_like.out, aten.rand_like.default, aten.rand_like.out, aten.randn_like.default, aten.randn_like.out, aten.randint_like.default, aten.randint_like.out, aten.randint_like.low_dtype, aten.randint_like.low_dtype_out, aten.zeros_like.default, aten.zeros_like.out, aten.new_empty.default, aten.new_empty.out, aten.new_empty_strided.default, aten.new_empty_strided.out, aten.new_full.default, aten.new_full.out, aten.new_zeros.default, aten.new_zeros.out, aten.new_ones.default, aten.new_ones.out, ) @functools.lru_cache(None) def _is_tensor_constructor(func: OpOverload): assert isinstance(func, OpOverload) schema = func._schema if any(contains_tensor_types(arg.type) for arg in schema.arguments): return False # TODO: no real reason to restrict multiple outputs return ( len(schema.returns) == 1 and schema.returns[0].type is torch._C.TensorType.get() ) @functools.lru_cache(None) def get_schema_info(func): return torch._C._SchemaInfo(func._schema) # type: ignore[attr-defined] # many of the decompositions registered to torch/_prims do not at the moment model # aliasing or strides, so as an incremental step, just enable the decompositions in # torch/_decomp/decompositions.py. # decomps are used for aot autograd tracing so we would like to unify on their # implementation and add additional testing to them @functools.lru_cache(None) def torch_decomp_decompositions(func): from torch._decomp import decomposition_table decompositions = torch._decomp.decompositions decomp_attrs = [getattr(decompositions, attr) for attr in dir(decompositions)] return decomposition_table[func] in decomp_attrs def tree_flatten_only(ty: Type[T], pytree: PyTree): flat_vals, _ = tree_flatten(pytree) return [elem for elem in flat_vals if isinstance(elem, ty)] # Similar to `MetaConverter`, this is a class for converting # multiple tensors into fake tensors which share the same view/storage # structure. Like `MetaConverter`, it uses `WeakIdRef` to # hold a weak reference for all memoized tensors. class FakeTensorConverter(object): @property def tensor_memo(self): return self.meta_converter.tensor_memo meta_converter: MetaConverter constant_storage_mapping: Dict[StorageWeakRef, List[ReferenceType]] def __init__(self): self.meta_converter = MetaConverter() # map from to storage to corresponding constant tensors self.constant_storage_mapping = {} def add_constant_storage_mapping(self, fake_tensor): # when you have a constant, aliased tensor: # const_tensor.add_(torch.rand([1])) # all aliases of it must become no longer const assert isinstance(fake_tensor, FakeTensor) and fake_tensor.constant is not None weak_st = StorageWeakRef(fake_tensor.constant._typed_storage()) # we need a map from a weak storage to all of its corresponding # constant tensors. python doesn't have the weak value equivalent # of defaultdict(list), so we are using a WeakValueDictionary as one if weak_st not in self.constant_storage_mapping: self.constant_storage_mapping[weak_st] = [] self.constant_storage_mapping[weak_st].append(weakref.ref(fake_tensor)) def invalidate_constant_aliases(self, tensor): assert not isinstance(tensor, FakeTensor) weak_st = StorageWeakRef(tensor._typed_storage()) if weak_st not in self.constant_storage_mapping: return for weak_tensor_ref in self.constant_storage_mapping[weak_st]: ten = weak_tensor_ref() if ten is not None: ten._fix_weakref() ten.constant = None del self.constant_storage_mapping[weak_st] def _get_memo(self, t): if WeakIdRef(t) in self.tensor_memo: out = self.tensor_memo[WeakIdRef(t)] out._fix_weakref() return out return None def set_tensor_memo(self, t, v): th = WeakIdRef(t) # hold a weak ref to self, otherwise it will be kept alive # by the del_ten closure self_weak_ref = weakref.ref(self) def del_ten(): self_ref = self_weak_ref() if self_ref is None: return # on shutdown, th may not be in memo self_ref.tensor_memo.pop(th, None) weakref.finalize(t, del_ten) self.tensor_memo[th] = v def from_real_tensor( self, fake_mode, t, make_constant=False, shape_env=None, ignore_subclass=False, *, source=None, ): maybe_memo = self._get_memo(t) if maybe_memo is not None: return maybe_memo existing_device = t.device # not yet supported in metatensors if t.is_quantized: raise UnsupportedFakeTensorException("quantized nyi in meta tensors") if type(t) is torch.nn.Parameter: assert not make_constant def mk_fake_tensor(make_meta_t): # NB: don't use in_kernel_invocation_manager. to # ensure FakeTensor can internally do constant computation # as necessary. Invocation manager is "more correct" as # it works for more operators in make_meta_t, but # invariant is that make_meta_t only calls factories # for which it is not strictly necessary to use the # invocation manager (I think!) with no_dispatch(): return FakeTensor( fake_mode, make_meta_t(), existing_device, constant=t if make_constant else None, ) out = self.meta_converter( t, shape_env=shape_env, callback=mk_fake_tensor, ignore_subclass=ignore_subclass, source=source, ) if out is NotImplemented: raise UnsupportedFakeTensorException("meta converter nyi") if make_constant: self.add_constant_storage_mapping(out) # NB: meta_converter set the memo return out # If you specify the device, it MUST be a meta tensor. def from_meta_and_device(self, fake_mode, t, device): assert ( t.device.type == "meta" ), f"tensor's device must be `meta`, got {t.device.type} instead" maybe_memo = self._get_memo(t) if maybe_memo is not None: return maybe_memo out = FakeTensor(fake_mode, t, device) self.set_tensor_memo(t, out) return out # You can have a real tensor that you need to convert into a fake tensor. # If you have a meta tensor already, call from_meta_and_device. # # You're allowed to pass a meta tensor to be turned into a fake # tensor; although an odd thing to do, this can occur if you're doing # cross ref testing and the inner test is already operating on meta tensors. def __call__( self, fake_mode, t, *, make_constant=False, shape_env=None, ignore_subclass=False, source=None, ): return self.from_real_tensor( fake_mode, t, make_constant, shape_env=shape_env, ignore_subclass=ignore_subclass, source=source, ) op_implementations = [] def register_op_impl(run_impl_check: Union[Callable[[OpOverload], bool], OpOverload]): def impl_decorator(op_impl): global op_implementations if isinstance(run_impl_check, OpOverload): op_implementations.append((lambda func: func == run_impl_check, op_impl)) else: op_implementations.append((run_impl_check, op_impl)) return op_impl return impl_decorator @register_op_impl( lambda func: (_is_tensor_constructor(func) or func in _like_tensor_constructors) ) def constructors(fake_mode, func, *args, **kwargs): assert func not in _non_kwarg_device_constructors _, new_kwargs = normalize_function( func, args=args, kwargs=kwargs, normalize_to_only_use_kwargs=True ) if func in _like_tensor_constructors: default_device = new_kwargs["input"].device # TODO: file issue args = (new_kwargs.pop("input"),) else: # cpu is default device if none is specified default_device = torch.device("cpu") args = () out_device = new_kwargs.pop("device", None) out_device = out_device if out_device is not None else default_device new_kwargs["device"] = torch.device("meta") # _like constructors have fake tensor inputs (maybe this causes the non-like # to fail? hmmm) with in_kernel_invocation_manager(fake_mode): r = func(*args, **new_kwargs) return FakeTensor(fake_mode, r, out_device) @register_op_impl(lambda func: func in (aten.to.prim_Device, aten.to.device)) def non_kwarg_to(fake_mode, func, *args, **kwargs): _, new_kwargs = normalize_function( func, args, kwargs, normalize_to_only_use_kwargs=True ) input_device = new_kwargs["device"] out_device = input_device if input_device else new_kwargs["input"].device new_kwargs["device"] = torch.device("meta") inp = new_kwargs.pop("input") with in_kernel_invocation_manager(fake_mode): r = func(inp, **new_kwargs) # TODO: I think this does the wrong thing if r is inp return fake_mode.fake_tensor_converter.from_meta_and_device( fake_mode, r, out_device ) # Dont default to default device handling, # since the device of `the_template` is ignored @register_op_impl(aten.resize_as_.default) def resize_as_(fake_mode, func, *args, **kwargs): with in_kernel_invocation_manager(fake_mode): return func(*args, **kwargs) @register_op_impl(aten._sparse_coo_tensor_with_dims_and_tensors.default) def _sparse_coo_tensor_with_dims_and_tensors(fake_mode, func, *args, **kwargs): # TODO: remove me return constructors(fake_mode, func, *args, **kwargs) # index.Tensor data-dependent in only some conditions @register_op_impl( lambda func: torch.Tag.dynamic_output_shape in func.tags # type: ignore[attr-defined] and func != aten.index.Tensor ) def dyn_shape(fake_mode, func, *args, **kwargs): raise DynamicOutputShapeException(func) @register_op_impl(lambda func: func is torch.ops.aten._local_scalar_dense.default) def local_scalar_dense(fake_mode, func, arg): if fake_mode.shape_env is None: # Without symints/symfloats, cannot handle this raise DataDependentOutputException(func) if is_float_dtype(arg.dtype): return fake_mode.shape_env.create_unbacked_symfloat() elif is_integer_dtype(arg.dtype): return fake_mode.shape_env.create_unbacked_symint() else: raise NotImplementedError(f"local_scalar_dense/item NYI for {arg.dtype}") # NB: this must be ordered after local_scalar_dense @register_op_impl( lambda func: torch.Tag.data_dependent_output in func.tags # type: ignore[attr-defined] ) def data_dep(fake_mode, func, *args, **kwargs): raise DataDependentOutputException(func) # Bool Indices get Expanded as Masks # See: IndexingUtils.h:expandTensors def check_no_bool_index_tensors(func, self, indices): for index in indices: if index is not None and index.dtype in (torch.bool, torch.uint8): raise DynamicOutputShapeException(func) def run_and_return_new_tensor_of_input_device(fake_mode, func, args, kwargs): _, new_kwargs = normalize_function( func, args=args, kwargs=kwargs, normalize_to_only_use_kwargs=True ) out_device = new_kwargs["input"].device with in_kernel_invocation_manager(fake_mode): out = func(*args, **kwargs) return FakeTensor(fake_mode, out, out_device) # Dont default to default device handling, # Since op can take in non-zero sized cpu # index tensors with cuda self @register_op_impl(aten.index.Tensor) def index_tensor(fake_mode, func, *args, **kwargs): # dynamic shape op if indices are bool/uint8 check_no_bool_index_tensors(func, *args, **kwargs) return run_and_return_new_tensor_of_input_device(fake_mode, func, args, kwargs) # takes in multiple-devices, dont default to default device handling @register_op_impl(aten.index_put.default) def index_put(fake_mode, func, *args, **kwargs): return run_and_return_new_tensor_of_input_device(fake_mode, func, args, kwargs) # same with index_put, but return the input @register_op_impl(aten.index_put_.default) def index_put_(fake_mode, func, *args, **kwargs): with in_kernel_invocation_manager(fake_mode): out = func(*args, **kwargs) _, new_kwargs = normalize_function( func, args=args, kwargs=kwargs, normalize_to_only_use_kwargs=True ) return new_kwargs["input"] @register_op_impl(lambda fn: fn in _device_not_kwarg_ops) def nyi(fake_mode, func, *args, **kwargs): assert func not in _device_not_kwarg_ops, f"NYI: {func}" @register_op_impl( lambda func: func in (aten.convolution.default, aten.convolution_backward.default) ) def conv(fake_mode, func, *args, **kwargs): _, kwargs = normalize_function( func, args=args, kwargs=kwargs, normalize_to_only_use_kwargs=True ) device = kwargs["input"].fake_device # need to re-enable mode so the tensors report fake device with fake_mode: # if the input is unsqueezed is done in Convolution.cpp we get segfault k = kwargs["weight"].ndim if k == 3 and not kwargs["input"].is_mkldnn and not kwargs["input"].is_xpu: mem_fmt = None else: if func is aten.convolution.default: conv_backend = torch._C._select_conv_backend(**kwargs) else: conv_backend = torch._C._select_conv_backend( kwargs["input"], kwargs["weight"], bias=None, stride=kwargs["stride"], padding=kwargs["padding"], dilation=kwargs["dilation"], transposed=kwargs["transposed"], output_padding=kwargs["output_padding"], groups=kwargs["groups"], bias_sizes=kwargs["bias_sizes"], ) mem_fmt = torch._C._conv_determine_backend_memory_format( kwargs["input"], kwargs["weight"], conv_backend ) def convert(t, mem_fmt): if t is None: return t if mem_fmt is not None: t = t.to(memory_format=mem_fmt) return FakeTensor(fake_mode, t, device) with in_kernel_invocation_manager(fake_mode): out = func(**kwargs) if func is aten.convolution.default: return convert(out, mem_fmt) else: return ( convert(out[0], mem_fmt), convert(out[1], mem_fmt), convert(out[2], None), ) FAST_OP_IMPLEMENTATIONS = {} # Unlike register_op_impl, these don't do the slow iteration for # run_impl_check, and these run BEFORE decompositions def register_fast_op_impl(func: OpOverload): def impl_decorator(op_impl): FAST_OP_IMPLEMENTATIONS[func] = op_impl return op_impl return impl_decorator # infer_size_impl in ExpandUtils def infer_size(a, b): dimsA = len(a) dimsB = len(b) ndim = max(dimsA, dimsB) expandedSizes = [0] * ndim for i in range(ndim - 1, -1, -1): offset = ndim - 1 - i dimA = dimsA - 1 - offset dimB = dimsB - 1 - offset sizeA = a[dimA] if dimA >= 0 else 1 sizeB = b[dimB] if dimB >= 0 else 1 if not (sizeA == sizeB or sizeA == 1 or sizeB == 1): raise RuntimeError( f"The size of tensor a ({sizeA}) " f"must match the size of tensor b ({sizeB}) " f"at non-singleton dimension {i})" ) expandedSizes[i] = sizeB if sizeA == 1 else sizeA return tuple(expandedSizes) def make_fast_binary_impl(slow_ref): def fast_binary_impl(mode, *args, **kwargs): def slow(msg): count_label(f"slow {msg}") with mode: return slow_ref(*args, **kwargs) count_label("attempt fast") # Fast path (based off of TensorIterator fast path). # Unfortunately, there is no way to easily deduplicate # this with either the TensorIterator C++ implementation # (which we don't want to SymIntify, and also the algorithm # here is slightly different from TensorIterator to allow # for broadcasting), nor the PrimTorch implementation # (which does not actually implement a fast path.) operands = args # compute_shape has_scalars = False has_tensors = False final_shape = None for op in operands: shape = op.shape if isinstance(op, torch.Tensor) else () if len(shape) == 0: has_scalars = True else: has_tensors = True if final_shape is None: final_shape = shape # TODO: Minor optimization: track if the shapes # were equal so you can skip the equality check # below if unnecessary final_shape = infer_size(final_shape, shape) assert final_shape is not None # Do some extra safety checks to see if the output # stride is obvious for op in operands: if isinstance(op, torch.Tensor) and op.shape == final_shape: break else: return slow("both tensors nontrivially broadcast") # compute_types cpu = torch.device("cpu") common_device = cpu common_dtype = None output_dtype = None has_different_input_dtypes = False for op in operands: if not isinstance(op, torch.Tensor): # Use elementwise_dtypes for the tricky case has_different_input_dtypes = True continue if common_device == cpu and not op.device.type == "cpu": common_device = op.device # Slightly simplified here as target_dtype cannot vary if common_dtype is None: common_dtype = op.dtype elif common_dtype != op.dtype: has_different_input_dtypes = True if has_different_input_dtypes: # compute promotion # TODO: we don't need the compute type _, common_dtype = elementwise_dtypes( *operands, type_promotion_kind=ELEMENTWISE_TYPE_PROMOTION_KIND.DEFAULT ) # check all tensors on same device # cpu scalars are assumed allow current_cpu_scalars_on_non_cpu = 0 max_cpu_scalars_on_non_cpu = 1 # hard coded atm for op in operands: if not isinstance(op, torch.Tensor): continue if common_device != cpu and op.dim() == 0 and op.device == cpu: if current_cpu_scalars_on_non_cpu >= max_cpu_scalars_on_non_cpu: return slow("error") current_cpu_scalars_on_non_cpu += 1 elif op.device != common_device: return slow("error") # compute_fast_setup_type is_contiguous = True is_channels_last = True # TODO: is_non-overlapping_and_dense (not bound from Python # no inplace, no out, everything defined for op in operands: if not isinstance(op, torch.Tensor): continue is_contiguous = is_contiguous and op.is_contiguous( memory_format=torch.contiguous_format ) is_channels_last = is_channels_last and op.is_contiguous( memory_format=torch.channels_last ) if is_contiguous: # do contiguous count_label("fast is_contiguous") return FakeTensor( mode, torch.empty( final_shape, dtype=common_dtype, device="meta", memory_format=torch.contiguous_format, ), device=common_device, ) if is_channels_last: count_label("fast channels_last") # do channels last return FakeTensor( mode, torch.empty( final_shape, dtype=common_dtype, device="meta", memory_format=torch.channels_last, ), device=common_device, ) return slow("no contiguity match") return fast_binary_impl @functools.lru_cache(None) def get_fast_op_impls(): import torch._refs register_fast_op_impl(torch.ops.aten.add.Tensor)( make_fast_binary_impl(torch._refs.add) ) register_fast_op_impl(torch.ops.aten.sub.Tensor)( make_fast_binary_impl(torch._refs.sub) ) register_fast_op_impl(torch.ops.aten.mul.Tensor)(make_fast_binary_impl(torch._refs.mul)) # type: ignore[has-type] register_fast_op_impl(torch.ops.aten.div.Tensor)( make_fast_binary_impl(torch._refs.div) ) return FAST_OP_IMPLEMENTATIONS @contextlib.contextmanager def in_kernel_invocation_manager(fake_mode): # See: note [Fake Tensor Dispatch Keys] prev_in_kernel = fake_mode.in_kernel_invocation meta_in_tls = torch._C._meta_in_tls_dispatch_include() assert meta_in_tls == prev_in_kernel, f"{meta_in_tls}, {prev_in_kernel}" guard = torch._C._DisableTorchDispatch() # type: ignore[attr-defined] fake_mode.in_kernel_invocation = True torch._C._set_meta_in_tls_dispatch_include(True) try: yield finally: fake_mode.in_kernel_invocation = prev_in_kernel torch._C._set_meta_in_tls_dispatch_include(prev_in_kernel) del guard # Return if the function allows Python numbers to bind to Tensors def should_allow_numbers_as_tensors(func: OpOverload): return torch._C._should_allow_numbers_as_tensors( func.name().split("::")[-1].split(".")[0] ) class FakeTensorConfig: debug = os.environ.get("TORCH_FAKE_TENSOR_DEBUG", False) class FakeTensor(torch.Tensor): """ Meta tensors give you the ability to run PyTorch code without having to actually do computation through tensors allocated on a `meta` device. Because the device is `meta`, meta tensors do not model device propagation. FakeTensor extends MetaTensors to also carry an additional `fake_device` which tracks devices that would have been used. """ fake_device: torch.device fake_mode: "FakeTensorMode" constant: Optional[torch.Tensor] @property def device(self): if self.fake_mode.in_kernel_invocation: return torch.device("meta") else: return self.fake_device # Note: [Fake Tensor Dispatch Keys] # In order to model the behavior of device-specific autocast # and autograd logic, we update the dispatch keys of FakeTensors # to reflect their fake device. This includes the BackendComponent # (DispatchKey::Meta -> DispatchKey::CUDA), and also the BackendComponent # related Autocast and Autograd keys. __torch__dispatch__ sits below # Autocast and Autograd, and is only invoked when we are at the # kernel for the BackendComponent. Then, we add Meta to the # thread-local dispatch include set to hit the meta kernel # instead of the kernel of the BackendComponent for the fake device. # The `device_for_backend_keys` does that below @staticmethod def __new__(cls, fake_mode, elem, device, constant=None): self = torch.Tensor._make_subclass( cls, elem, elem.requires_grad, dispatch_device=True, device_for_backend_keys=device, ) assert elem.device.type == "meta", elem.device.type device = device if isinstance(device, torch.device) else torch.device(device) # NB: it is fine, if a little confusing, for device to be meta # (we are faking a meta tensor in that case). However, it often # indicates some sort of confusion (e.g., you accidentally passed # in a meta tensor when you should have passed in the real tensor). # So by default we disallow meta, and if you are working in a situation # where it is helpful (e.g., crossref testing) you can turn it back # on if not fake_mode.allow_meta: assert device.type != "meta" # normalize cuda device. if device.type == "cuda" and device.index is None: device = torch.device(f"cuda:{torch.cuda.current_device()}") self.fake_device = device # type: ignore[attr-defined] self.fake_mode = fake_mode # type: ignore[attr-defined] self.constant = constant # type: ignore[attr-defined] if FakeTensorConfig.debug: import traceback self._debug_trace = traceback.extract_stack() # type: ignore[attr-defined] return self # In some circumstances, a conventional torch.Tensor constructor # will get rewritten to call into FakeTensor. We must provide an # __init__ method that can accept the Python interpreters initialization # in such a situation; we must also be able to handle direct fake # tensor construction via FakeTensor(). # # In particular, the __init__ call will look funny in the following case: # # with FakeTensorMode(): # x = torch.Tensor([1, 2, 3]) # # this desugars into: # # with FakeTensorMode(): # x = torch.Tensor.__new__([1, 2, 3]) # # NB: x is a fake tensor, because of the mode! # x.__init__([1, 2, 3]) # not the normal fake tensor args! # def __init__(self, *args, **kwargs): super().__init__() @staticmethod def from_tensor(t, fake_mode): return fake_mode.from_tensor(t) # TODO: resolve error in default __repr__ def __repr__(self): with in_kernel_invocation_manager(self.fake_mode): self_repr = super().__repr__() return f"FakeTensor({self_repr}, {self.fake_device})" @classmethod @count def __torch_dispatch__(cls, func, types, args=(), kwargs=None): # need to handle here to avoid infinite recursion # see [in_kernel_invocation] if func == torch.ops.prim.device.default: assert len(args) == 1 and isinstance(args[0], FakeTensor) if args[0].fake_mode.in_kernel_invocation: return torch.device("meta") else: return args[0].fake_device # Because fake mode can return NotImplemented (if it sees a subclass # it doesn't know how to deal with), this test here is important # because the next dispatch after a fake mode will attempt to use # subclasses of tensors to dispatch, and any FakeTensor arguments # will be considered eligible. if any(not issubclass(t, FakeTensor) and t is not torch.Tensor for t in types): return NotImplemented fake_mode = None for arg in itertools.chain(tree_flatten(args)[0], tree_flatten(kwargs)[0]): if isinstance(arg, FakeTensor): if fake_mode is None: fake_mode = arg.fake_mode else: assert fake_mode is arg.fake_mode, "Mixing modes NYI" assert fake_mode is not None with fake_mode: # type: ignore[attr-defined] return func(*args, **kwargs) @staticmethod def _find_common_device(func, args, kwargs) -> Tuple[torch.device, bool]: # Returns: (common_device, has_scalar_only_inputs) # cpu - zero-dim tensors can be called in cuda kernels, # so overwrite the common_device if it the only existing # device comes from a cpu zero-dim tensor common_device = None has_scalar_only_inputs = False is_cpu_zero_dim = None def cpu_zero_dim(t): return t.device.type == "cpu" and t.dim() == 0 def merge_devices(t): nonlocal common_device nonlocal is_cpu_zero_dim if not isinstance(t, FakeTensor): return if common_device is None: common_device = t.device is_cpu_zero_dim = cpu_zero_dim(t) return t_is_cpu_zero_dim = cpu_zero_dim(t) if t.device == common_device: if is_cpu_zero_dim: is_cpu_zero_dim = t_is_cpu_zero_dim return # mismatching devices ! # if current tensor is cpu 0 dim, defer to existing device if t_is_cpu_zero_dim: return # current device is from cpu 0 dim tensor, overwrite if is_cpu_zero_dim: common_device = t.device is_cpu_zero_dim = t_is_cpu_zero_dim return # mismatching devices of non-zero dim tensors, throw # This might be valid behavior and need to be explicitly modeled, e.g. reshape_as raise RuntimeError( f"Unhandled FakeTensor Device Propagation for {func}, found two different devices {common_device}, {t.device}" ) tree_map(merge_devices, args) tree_map(merge_devices, kwargs) # some functions that allow Python numbers to bind to Tensors # if we have failed to find a device, and we're running one of these operators, # we must have scalar only inputs if should_allow_numbers_as_tensors(func) and common_device is None: # ops with scalar only inputs always have result on cpu has_scalar_only_inputs = True common_device = torch.device("cpu") assert common_device is not None, f"Could not find common device for {func}" return common_device, has_scalar_only_inputs __torch_function__ = torch._C._disabled_torch_function_impl # We keep one instantiation of `fake_tensor_converter` active # for the duration of `with FakeTensorMode()`. # This allows accurate storage aliasing across invocation of # different operators. While this will keep all freshly allocated # tensors alive during `FakeTensorMode`, there will no be no # new allocations of Tensors which have non-meta storage so # memory should not significantly incraese. class FakeTensorMode(TorchDispatchMode): def __init__( self, *, allow_fallback_kernels=True, allow_non_fake_inputs=False, shape_env=None, ): self.allow_fallback_kernels = allow_fallback_kernels self.fake_tensor_converter = FakeTensorConverter() import torch._functorch.config self.allow_meta = torch._functorch.config.fake_tensor_allow_meta # A flag that controls, whether we want to invoke ops on mix of # real weights/global variables and fake inputs self.allow_non_fake_inputs = allow_non_fake_inputs # [in_kernel_invocation] # when FakeTensor is invoked in user code, .device should return # the fake_device of the tensor so that code such as as `if x.is_cuda` # or torch.zeros([10, 10], device=x.device) continues to execute as if # the FakeTensor were real. However, within kernel execution, we return # the `Meta` device because all computation within the kernels should # behave as if the Tensors are on meta devices. Kernels should allocate # new tensors on meta devices, and checks like `is_meta` should return true. # within python refs, we always return the real device by defining # the device property self.in_kernel_invocation = False self.shape_env = shape_env @count def __torch_dispatch__(self, func, types, args=(), kwargs=None): try: return self.dispatch(func, types, args, kwargs) except TypeError: log.exception("fake tensor raised TypeError") raise def dispatch(self, func, types, args=(), kwargs=None): kwargs = kwargs if kwargs else {} if func == torch.ops.prim.device.default: assert len(args) == 1 and isinstance(args[0], FakeTensor) if args[0].fake_mode.in_kernel_invocation: return torch.device("meta") else: return args[0].fake_device if log.getEffectiveLevel() <= logging.DEBUG: log.debug( f"{' ' * RECURSION_COUNT}FakeTensorMode.__torch_dispatch__: {func}" ) incr = IncrementRecursionCount() # Some attribute queries that can be serviced directly # See Note [is_coalesced is dispatched] if func in { torch.ops.aten.is_coalesced.default, torch.ops.aten.dense_dim.default, torch.ops.aten.sparse_dim.default, }: # NB: no_dispatch is ok here too, this func is very simple with in_kernel_invocation_manager(self): return func(*args, **kwargs) flat_arg_fake_tensors = tree_flatten_only(FakeTensor, (args, kwargs)) flat_symints = tree_flatten_only(torch.SymInt, (args, kwargs)) has_symbolic_sizes = ( any([i._has_symbolic_sizes_strides for i in flat_arg_fake_tensors]) or len(flat_symints) > 0 ) converter = self.fake_tensor_converter # To constant propagate through these functions: # 1, If this is a lift, the input tensor is guaranteed to be a # constant, so we keep a copy of the original argument along so # we can query it if we're asked to item() it at some later point # 2, Some functions that allow Python numbers to bind to Tensors, e.g, torch.div if func in self.lift_fns or ( should_allow_numbers_as_tensors(func) and not has_symbolic_sizes and not flat_arg_fake_tensors ): out = func(*args, **kwargs) if self.may_turn_const(out): # NB: not in_kernel_invocation_manager because we're doing real # compute here with no_dispatch(): out = out.clone() return converter(self, out, make_constant=True) # See [subclass inputs] below # NB: If you're seeing a mysterious infinite loop involving fake # tensor, it might be related to this line. Though I'm not sure # how you'll know to read this comment, as this line won't show up # in the stack trace. if self.check_for_subclass(args, kwargs): return NotImplemented # if we are in the dispatch mode, we will enter this function even if the inputs # are not FakeTensors. For now, throw if any non-Fake Tensor inputs # and just support constructors. # this is generated from torch.tensor(), which does not use the # dispatcher, to allow wrapper subclasses to wrap the new tensor if func in self.lift_fns: assert ( len(kwargs) == 0 and len(args) == 1 and type(args[0]) is torch.Tensor ), f"{args} {kwargs}" return converter(self, args[0]) args, kwargs = self.validate_and_convert_non_fake_tensors( func, converter, args, kwargs ) # The current constant handling only support tracing systems # (aot autograd, torchdynamo) where each operation is run consecutively. # Because each operation is run in order, we can trace out and support # sequences like: x = torch.tensor(0.); y = x.add_(1) # Whenver a constant is written to but with inputs that cannot be evaluated # statically, such as random_(), we invalidate all constants that alias the input # We will rely on functionalization for use of fake tensors constants as persistent # objects on an FX Graph. # We dispatch size/stride/numel on the FakeTensor not its constant, so bail on inplace_view all_constant = all(e.constant is not None for e in flat_arg_fake_tensors) if ( torch.Tag.nondeterministic_seeded not in func.tags # type: ignore[attr-defined] and torch.Tag.inplace_view not in func.tags # type: ignore[attr-defined] and all_constant and len(flat_arg_fake_tensors) != 0 and not has_symbolic_sizes ): const_args, const_kwargs = pytree.tree_map_only( FakeTensor, lambda t: t.constant, (args, kwargs) ) # NB: not in_kernel_invocation_manager(self) as we want to do REAL # compute with no_dispatch(): out = func(*const_args, **const_kwargs) all_constant = pytree.tree_all_only( torch.Tensor, lambda t: self.may_turn_const(t), out ) if all_constant: return pytree.tree_map_only( torch.Tensor, lambda t: converter(self, t, make_constant=True), out, ) # we weren't able to turn outputs to constants, # so invalidate all constants that might be aliases of the outputs for ten in tree_flatten_only(torch.Tensor, out): converter.invalidate_constant_aliases(ten) # we are falling through to running non constant tensors, any input constant that # is written to must be invalidated self.invalidate_written_to_constants(func, flat_arg_fake_tensors, args, kwargs) # Try for fastpath if has_symbolic_sizes: fast_impl = get_fast_op_impls().get(func) if fast_impl is not None: return fast_impl(self, *args, **kwargs) # If there's a Python meta, prefer that over the decomposition from torch._decomp import meta_table as meta_table if func not in meta_table and not self.cpp_meta_supports_symint(func): from torch._decomp import decomposition_table # Prefer Python decompositions over C++ ones if func in decomposition_table and ( has_symbolic_sizes or ( # TODO: Remove these exclusions, so that we can remove # this leg entirely torch_decomp_decompositions(func) and all(not e.is_sparse for e in flat_arg_fake_tensors) ) ): with self: return decomposition_table[func](*args, **kwargs) with self: # Decomposes CompositeImplicitAutograd ops r = func.decompose(*args, **kwargs) if r is not NotImplemented: return r # prims already wrap FakeTensor inputs to FakeTensor outputs # and do device logic, we dont need do anything but run them # and ensure that Meta kernels are dispatched to (see) # Fake Tensor Dispatch Keys # TODO - we should be use the prim aten impl if "prims::" in func._schema.name and hasattr(func, "prim_meta_impl"): with self: return func.prim_meta_impl(*args, **kwargs) # special handling for funcs registered through `register_op_impl`, # e.g., manipulating args on constructor calls to construct meta tensors # and then afterwards wrapping them to a FakeTensor for run_impl_check, op_impl in op_implementations: if run_impl_check(func): op_impl_out = op_impl(self, func, *args, **kwargs) if op_impl_out != NotImplemented: return op_impl_out # run kernel registered to meta for func, which include # python meta registrations, prims, decomps, and c++ meta fns (structured kernels) try: with in_kernel_invocation_manager(self): r = func(*args, **kwargs) except NotImplementedError as not_implemented_error: # no meta kernel registered, fallback to kernel for the device if not self.allow_fallback_kernels: raise not_implemented_error return run_fallback_kernel(self, func, args, kwargs, not_implemented_error) return self.wrap_meta_outputs_with_default_device_logic(r, func, args, kwargs) # [subclass inputs] # Suppose we enable fake tensor mode. This means that fake tensor # mode will run first. But what if we do an operation that # involves a tensor subclass that will desugar into normal tensor # operations? Without returning NotImplemented, fake tensor mode will run first, # decide that a conversion was made (since there was a non fake # tensor argument), and report an error that converting non # fake tensor is not supported. What we actually wanted to happen # was to give the subclass a chance to figure out what it wants to # before erroring out. Returning NotImplemented here allows this. def check_for_subclass(self, args, kwargs): def check(x): return ( not isinstance(x, FakeTensor) and type(x) is not torch.Tensor and type(x) is not torch.nn.Parameter ) return any([check(x) for x in tree_flatten_only(torch.Tensor, (args, kwargs))]) def validate_and_convert_non_fake_tensors(self, func, converter, args, kwargs): """ Checks if the list of tensors are fake tensors. If not, try to convert them to fake tensors. """ def validate(x): if not isinstance(x, FakeTensor): if torch.Tag.inplace_view in func.tags: # type: ignore[attr-defined] raise Exception( f"Can't call metadata mutating ops on non-Fake Tensor inputs. Found in {func}(*{args}, **{kwargs})" ) if not self.allow_non_fake_inputs: raise Exception( f"Please convert all Tensors to FakeTensors first or instantiate FakeTensorMode " f"with 'allow_non_fake_inputs'. Found in {func}(*{args}, **{kwargs}) " ) return converter(self, x) return x return tree_map_only( torch.Tensor, validate, (args, kwargs), ) def wrap_meta_outputs_with_default_device_logic(self, r, func, args, kwargs): wrap = self.gen_wrap_fn(func, args, kwargs) # if device is specified, use that if kwargs.get("device", None): return tree_map(partial(wrap, device=kwargs["device"]), r) return tree_map(partial(wrap), r) def gen_wrap_fn(self, func, args, kwargs): converter = self.fake_tensor_converter # Lazily initialized, in case there are no tensor returns common_device = None has_scalar_only_inputs = False def wrap(e, device=None): nonlocal common_device nonlocal has_scalar_only_inputs if ( isinstance(e, torch.Tensor) and not isinstance(e, FakeTensor) and converter is not None ): if common_device is None: ( common_device, has_scalar_only_inputs, ) = FakeTensor._find_common_device(func, args, kwargs) if has_scalar_only_inputs: # Under FakeTensorMode, op accepts scalar only inputs, such as aten.add/sub/mul/div, # returns a real scalar tensor on CPU. See TensorMeta() in _prims/__init__.py for details. # We thus directly convert real tensor to fake tensor. return converter(self, e) else: return converter.from_meta_and_device( self, e, device or common_device ) else: return e return wrap def cpp_meta_supports_symint(self, func): if torch.Tag.view_copy in func.tags: # type: ignore[attr-defined] return True return func in [ aten.empty_strided.default, aten.as_strided_scatter.default, aten.as_strided.default, aten.as_strided_.default, aten.zeros.default, aten.detach.default, aten.view_as_real.default, aten.view_as_complex.default, aten.set_.source_Storage_storage_offset, aten._sparse_coo_tensor_with_dims_and_tensors.default, ] @property def lift_fns(self): return (aten.lift_fresh.default, aten.lift_fresh_copy.default) def may_turn_const(self, t): return ( t.numel() <= CONSTANT_NUMEL_LIMIT and not t.is_sparse and not isinstance(t, FakeTensor) and not t.device.type == "meta" ) def invalidate_written_to_constants( self, func, flat_arg_fake_tensors, args, kwargs ): any_constant = any(e.constant is not None for e in flat_arg_fake_tensors) if any_constant and get_schema_info(func).is_mutable(): schema_info = get_schema_info(func) _, new_kwargs = normalize_function( func, args=args, kwargs=kwargs, normalize_to_only_use_kwargs=True ) for k, v in new_kwargs.items(): k = k if (k != "input" or schema_info.has_argument(k)) else "self" if ( isinstance(v, FakeTensor) and schema_info.is_mutable(k) and v.constant is not None ): self.fake_tensor_converter.invalidate_constant_aliases(v.constant) def from_tensor( self, tensor, static_shapes=False, ignore_subclass=False, source: Optional[Source] = None, ): if static_shapes: return self.fake_tensor_converter( self, tensor, ignore_subclass=ignore_subclass, source=source ) return self.fake_tensor_converter( self, tensor, shape_env=self.shape_env, ignore_subclass=ignore_subclass, source=source, ) # NB: returns fake tensors def run_fallback_kernel(fake_mode, func, args, kwargs, orig_not_implemented_exception): # these should all be supported, just to be safe # avoid fallback for operators which inplace modify metadata # because the input fake tensors would be umodified if torch.Tag.inplace_view in func.tags: # type: ignore[attr-defined] raise orig_not_implemented_exception inp_impls = {} # Don't use in_kernel_invocation_manager(fake_mode) as we want to do # REAL compute (not with meta device) with no_dispatch(): def to_real_tensor(e): if isinstance(e, FakeTensor): out = torch.zeros_like(e, device=e.fake_device) if e.is_sparse: out._coalesced_(e.is_coalesced()) inp_impls[id(out)] = e return out return e args = tree_map(to_real_tensor, args) kwargs = tree_map(to_real_tensor, kwargs) r = func(*args, **kwargs) tensor_impls = set() storages = set() for e in tree_flatten((args, kwargs))[0]: if isinstance(e, torch.Tensor): if not e.is_sparse: storages.add(e._typed_storage()._cdata) # TODO: also check metadata change on inputs # proper aliasing/metadata relationship between outputs and inputs will # not be set up, bc of conversion to device, unless we can reuse an # input impl for e in tree_flatten(r)[0]: if id(e) not in inp_impls and ( isinstance(e, torch.Tensor) and not e.is_sparse and e._typed_storage()._cdata in storages ): raise orig_not_implemented_exception def map_out(e): if isinstance(e, torch.Tensor): if id(e) in inp_impls: return inp_impls[id(e)] else: return fake_mode.fake_tensor_converter(fake_mode, e) else: return e return tree_map(map_out, r) # Just for use to allow copying a module to fake tensors, # does not apply elsewhere class FakeCopyMode(TorchFunctionMode): def __init__(self, fake_mode): self.fake_mode = fake_mode def __torch_function__(self, func, types, args=(), kwargs=None): kwargs = kwargs if kwargs else {} # clone will get called in Parameter deepcopy if func == torch._C._TensorBase.clone: return func( self.fake_mode.from_tensor(args[0], static_shapes=True), **kwargs ) elif func == torch.Tensor.__deepcopy__: assert len(args) == 2 and len(kwargs) == 0 tensor, memo = args if id(tensor) in memo: return memo[id(tensor)] out = self.fake_mode.from_tensor(tensor, static_shapes=True) memo[id(tensor)] = out return out else: with torch._C.DisableTorchFunctionSubclass(): return func(*args, **kwargs)