# Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the BSD-style license found in the # LICENSE file in the root directory of this source tree. import contextlib import functools from typing import Any, Callable, Dict, List, Optional, Tuple, Union import torch import torch.utils._pytree as pytree from torch.fx import Tracer, GraphModule from torch._subclasses.fake_tensor import FakeTensorMode from torch._dispatch.python import enable_python_dispatcher, enable_pre_dispatch import torch.fx as fx from torch.fx.passes.shape_prop import _extract_tensor_metadata from contextlib import contextmanager, nullcontext import inspect from dataclasses import dataclass import weakref import operator from torch.utils._stats import count import logging from torch.overrides import TorchFunctionMode from torch.utils._python_dispatch import ( TorchDispatchMode, _pop_mode_temporarily, _get_current_dispatch_mode, ) from torch._subclasses import FakeTensor from .symbolic_shapes import ShapeEnv, SymDispatchMode, SymNode from torch.fx import Proxy import torch.fx.traceback as fx_traceback from torch import SymInt, SymFloat, SymBool from torch.utils.weak import WeakTensorKeyDictionary __all__ = ["PythonKeyTracer", "dispatch_trace", "make_fx", "DecompositionInterpreter", "py_sym_types", "get_innermost_proxy_mode"] aten = torch.ops.aten prim = torch.ops.prim log = logging.getLogger(__name__) not_implemented_log = torch._logging.getArtifactLogger(__name__, "not_implemented") CURRENT_DECOMPOSITION_TABLE: Dict[torch._ops.OpOverload, Callable] = {} CONSTANT_NUMEL_LIMIT = 1 # We currently convert all SymInt to proxies before we use them. # This could plausibly be handled at the Dynamo level. pytree._register_pytree_node(torch.Size, lambda x: (list(x), None), lambda xs, _: tuple(xs)) def fake_signature(fn, nargs): """FX gets confused by varargs, de-confuse it""" argnames = ",".join(f"arg{i}" for i in range(nargs)) return eval(f"lambda {argnames}: fn({argnames})", {"fn": fn}) @contextmanager def decompose(decomposition_table): global CURRENT_DECOMPOSITION_TABLE old_decomposition_table = CURRENT_DECOMPOSITION_TABLE CURRENT_DECOMPOSITION_TABLE = decomposition_table try: yield CURRENT_DECOMPOSITION_TABLE finally: CURRENT_DECOMPOSITION_TABLE = old_decomposition_table # ensure we cannot collide with other properties proxy_slot = object() no_default = object() py_sym_types = (SymInt, SymFloat, SymBool) def is_sym_node(node): assert hasattr(node, 'meta'), "All nodes traced with proxy_tensor should have meta" return "val" in node.meta and isinstance(node.meta['val'], py_sym_types) def set_proxy_slot(obj, tracer, proxy): if isinstance(obj, torch.Tensor): # We DO want to clobber proxies whenever we run an inplace operation # on a tensor, and it affects the metadata on the proxy. tracer.tensor_tracker[obj] = proxy else: # NB: Never clobber pre-existing proxy. Although the proxies # are in principle equivalent, when we do graph partitioning # we need there not to be spurious dependencies on tangent inputs. # This works because primals get their SymInts set first, and # THEN later we allocate tangent inputs. Make sure if a SymInt # is derivable from a primal that we use that. assert isinstance(obj, SymNode), type(obj) if obj not in tracer.symnode_tracker: tracer.symnode_tracker[obj] = proxy def has_proxy_slot(obj, tracer): assert isinstance(obj, (torch.Tensor, SymNode)), type(obj) return get_proxy_slot(obj, tracer, False, lambda _: True) # the default argument is what to return if the slot is not set. # the transform argument is handy if you need to extract a subfield from # the successfully looked up result (but NOT the default.) def get_proxy_slot(obj, tracer, default=no_default, transform=lambda x: x): if isinstance(obj, torch.Tensor): tracker = tracer.tensor_tracker else: assert isinstance(obj, SymNode), type(obj) tracker = tracer.symnode_tracker if obj not in tracker: if default is no_default: raise RuntimeError(f"{obj} is not tracked with proxy for {tracer}") return default return transform(tracker[obj]) def snapshot_fake(val): return val.detach() def extract_val(val): if isinstance(val, FakeTensor): return snapshot_fake(val) elif isinstance(val, py_sym_types): return val elif isinstance(val, (list, tuple)): return val.__class__([extract_val(x) for x in val]) elif isinstance(val, torch.Tensor): if not val.is_sparse: # NB: Kinda hacky, but we should try to get val as the metadata # everywhere # TODO: This doesn't properly track storages. A more robust # approach would be to maintain a per-trace FakeTensorMode and # from_real_tensor to create fake values (don't forget to # snapshot_fake) fake_tensor_mode = FakeTensorMode(allow_fallback_kernels=True) with fake_tensor_mode: return torch.empty_strided(val.shape, val.stride(), device=val.device, dtype=val.dtype) else: return None # What invariants do we have for the 'val' set on the FX node? It has accurate # metadata... but only for metadata that exists "below" all other subsystems # (most notably autograd, but also vmap, functorch transforms, etc). This means # you can get the dtype, shape, stride, storage, but you CANNOT get requires_grad, # grad_fn, _base (_base actually may be set due to recursive call to # ADInplaceOrView, but you shouldn't rely on it.) def set_meta(proxy, val): proxy.node.meta['val'] = extract_val(val) # Best effort tensor_meta setting; prefer using val! if isinstance(val, FakeTensor): proxy.node.meta['tensor_meta'] = _extract_tensor_metadata(val) elif isinstance(val, torch.Tensor) and not val.is_sparse: proxy.node.meta['tensor_meta'] = _extract_tensor_metadata(val) return proxy def thunkify(f, *args, **kwargs): """ Delays computation of f until it's called again Also caches the result """ return functools.lru_cache(1)(functools.partial(f, *args, **kwargs)) def track_tensor(tensor, proxy, *, constant, tracer): def try_set_proxy_slot(outer_s, proxy_callable, *args): assert callable(proxy_callable) if isinstance(outer_s, SymInt): inner_s = outer_s.node set_proxy_slot(inner_s, tracer, thunkify(proxy_callable, outer_s, *args)) # The basic idea is that we need to associate each tensor/SymInt # with a Proxy. How do we setup this association? We just store # the proxy on the proxy slot of the object, keyed on the tracer # (so that if we have multiple tracers at the same time, they # don't clobber each other.) for i, s in enumerate(tensor.shape): try_set_proxy_slot(s, lambda x, i: set_meta(torch.ops.aten.sym_size(proxy, i), x), i) for i, s in enumerate(tensor.stride()): try_set_proxy_slot(s, lambda x, i: set_meta(torch.ops.aten.sym_stride(proxy, i), x), i) try_set_proxy_slot(tensor.numel(), lambda x: set_meta(torch.ops.aten.sym_numel(proxy), x)) try_set_proxy_slot(tensor.storage_offset(), lambda x: set_meta(torch.ops.aten.sym_storage_offset(proxy), x)) set_proxy_slot(tensor, tracer, _ProxyTensor(proxy, constant)) def track_tensor_tree(inner_res, proxy_res, *, constant, tracer): def wrap_with_proxy(e, proxy, constant): if isinstance(e, torch.Tensor): track_tensor(e, proxy, tracer=tracer, constant=constant) set_meta(proxy, e) elif isinstance(e, py_sym_types): # NB: eagerly set meta here, so that the numbering is in order set_meta(proxy, e) set_proxy_slot(e.node, tracer, lambda: proxy) elif isinstance(e, (tuple, list)): if isinstance(proxy, fx.Proxy): set_meta(proxy, e) # example use case: allreduce_ returns ([tensor], work) for idx, ee in enumerate(e): wrap_with_proxy(ee, proxy[idx], get_constant(idx)) def get_constant(idx): if constant is None: return None else: return constant[idx] wrap_with_proxy(inner_res, proxy_res, constant) return inner_res def maybe_disable_fake_tensor_mode(): # TODO: figure out if this API generally makes sense and bake it into the # library mb_fake_mode = _get_current_dispatch_mode() if isinstance(mb_fake_mode, FakeTensorMode): return _pop_mode_temporarily() else: return nullcontext() @dataclass class _ProxyTensor: proxy: Proxy constant: Optional[torch.Tensor] def fetch_sym_proxy(tracer): def inner(e): n = e.node if n.constant is not None: return n.constant else: # NB: we REQUIRE all symints to be tracked return get_proxy_slot(n, tracer)() return inner def fetch_tensor_proxy(tracer): return lambda t: get_proxy_slot(t, tracer, t) HANDLED_TYPES = (torch.Tensor, torch.nn.Parameter, FakeTensor) def proxy_call(proxy_mode, func, pre_dispatch, args, kwargs): unrecognized_types = [] def can_handle_tensor(x): r = type(x) in HANDLED_TYPES or has_proxy_slot(x, proxy_mode.tracer) if proxy_mode._allow_fake_constant: r = r or type(x) in (torch._subclasses.FakeTensor,) if not r: unrecognized_types.append(type(x)) return r # If there are any tensor subclasses, we need to handle those tensor subclasses first # TODO: we could use types to test this if not pytree.tree_all_only(torch.Tensor, can_handle_tensor, (args, kwargs)): not_implemented_log.debug("ProxyTensorMode tensors without proxy had unrecognized subclasses: %s", unrecognized_types) return NotImplemented if func in CURRENT_DECOMPOSITION_TABLE: with proxy_mode: r = CURRENT_DECOMPOSITION_TABLE[func](*args, **kwargs) if r is not NotImplemented: return r # For pre-autograd tracing, we do not want to run CompositeImplicit decomps. if not pre_dispatch: with proxy_mode: r = func.decompose(*args, **kwargs) if r is not NotImplemented: return r tracer = proxy_mode.tracer f_args, f_kwargs = pytree.tree_map_only(torch.Tensor, fetch_tensor_proxy(tracer), (args, kwargs)) # If there are SymInts, we also should not consider this constant. # However, fake tensor handling of SymInts is sufficiently broken that # I couldn't write a test for this case all_constant = ( pytree.tree_all_only(_ProxyTensor, lambda t: t.constant is not None, (f_args, f_kwargs)) # TODO: maybe constant SymInts should also be allowed? Not sure if # this can happen and pytree.tree_all_only((SymInt, SymFloat, SymBool), lambda _: False, (args, kwargs)) ) if torch.Tag.data_dependent_output in func.tags: # Check if all of the Tensor inputs are constants if all_constant: const_args, const_kwargs = pytree.tree_map_only( _ProxyTensor, lambda t: t.constant, (f_args, f_kwargs) ) with maybe_disable_fake_tensor_mode(): return func(*const_args, **const_kwargs) # If any of the Tensor inputs are "real" (not FakeTensor), we may # incorrectly burn in constants by allowing this access. Raise # an error in this case if pytree.tree_all_only(torch.Tensor, lambda t: not isinstance(t, FakeTensor), (args, kwargs)): raise RuntimeError( f"It appears that you're trying to get value out of a tracing tensor with {func} - erroring out! " "It's likely that this is caused by data-dependent control flow or similar. " "It may be possible to trace this with dynamic shapes; try setting tracing_mode='symbolic' " "in your make_fx call." ) proxy_args, proxy_kwargs = pytree.tree_map_only( (SymInt, SymFloat, SymBool), fetch_sym_proxy(proxy_mode.tracer), pytree.tree_map_only(_ProxyTensor, lambda e: e.proxy, (f_args, f_kwargs)) ) # When we trace through a torch.tensor invocation, you never actually # see a torch.ops.aten.tensor call. Instead, the way this function is # implemented internally is that we allocate a plain tensor (this is # *guaranteed* to be a plain tensor, we disable all modes when doing # so), and then call at::lift_fresh on it (to give modes a chance to do # their stuff). Furthermore, the tensor argument to lift_fresh is guaranteed # to be freshly allocated, so we want lift_fresh to be a no-op (directly # returning the input argument). # # Here is the basic problem: when we trace this sequence of executions # into an FX graph, what happens to this call sequence? Traditionally, # tensor constants get interned as buffers on the FX GraphModule. But # this is dangerous. Consider: # # x = torch.tensor(1) # x.add_(2) # # Naively, this traces into: # # t = self._tensor_constant0 # initialized to torch.tensor(1) # x = torch.ops.aten.lift_fresh(t) # x.add_(2) # # If lift_fresh returns t directly, the subsequent add_ call will # modify the tensor constant. Really, the problem is we've violated # the invariant the the argument to lift is fresh. So what we should # preserve the invariant by replacing lift_fresh with lift_fresh_copy: # # t = self._tensor_constant0 # initialized to torch.tensor(1) # x = torch.ops.aten.lift_fresh_copy(t) # x.add_(2) # # This is what the overload modification does. if func is torch.ops.aten.lift_fresh.default: func = torch.ops.aten.lift_fresh_copy.default proxy_out = proxy_mode.tracer.create_proxy('call_function', func, proxy_args, proxy_kwargs, name=proxy_mode.tracer.graph._target_to_str(func.overloadpacket.__name__)) # This makes DCE marginally less likely to DCE inplace operations. # It is not strictly necessary # Kind of a hacky way to test if an op is in-place or not if func.overloadpacket.__name__[-1] == "_" and func.overloadpacket.__name__[0] != "_": if isinstance(args[0], List): # e.g., c10d::allreduce_ returns a list of tensors as the first element # in the output. for i, a in enumerate(args[0]): a.proxy = proxy_out[0][i] else: args[0].proxy = proxy_out out = func(*args, **kwargs) # In some circumstances, we will be tracing in a situation where a tensor # is *statically* known to be a constant (currently, this only happens if # you run torch.tensor; deterministic factory functions like torch.arange # don't get this treatment). When the tensor in question is small, it's # helpful to due constant propagation in case we call item() (in which # case we can return the constant value that is known, rather than give # an error.) The logic here tests if constant propagation is possible # (because all of the inputs are constant). If so, we disable fake tensor # mode (if it is on) and do true compute on the constant. # # It's worth highlighting that we're making a policy decision here. # There is a potential that the tensor is actually quite large, and we # don't actually want to run the compute. The tensor being quite large # is one of the reasons why factory functions don't get this treatment # (since they can be quite large; if a parameter is initialized to a # constant value it will be!) Similarly, there is also a potential # to run an operator that blows up the size of a small tensor; we don't # protect against this case, but we could force, e.g., only single # element constant computation by testing the numel of the result before # propagating const-ness. Similarly, we don't require the constant to # live on CPU, but we could. any_constant = pytree.tree_any_only(_ProxyTensor, lambda t: t.constant is not None, (f_args, f_kwargs)) constant = None # 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 if func is torch.ops.aten.lift_fresh_copy.default and out.numel() <= CONSTANT_NUMEL_LIMIT: with maybe_disable_fake_tensor_mode(): constant = args[0].clone() elif ( torch.Tag.nondeterministic_seeded not in func.tags and all_constant and any_constant and pytree.tree_all_only(torch.Tensor, lambda t: t.numel() <= CONSTANT_NUMEL_LIMIT, out) ): # NB: do NOT include factories as constants with maybe_disable_fake_tensor_mode(): const_args, const_kwargs = pytree.tree_map_only( _ProxyTensor, lambda t: t.constant, (f_args, f_kwargs) ) constant = func(*const_args, **const_kwargs) else: constant = None track_tensor_tree(out, proxy_out, constant=constant, tracer=tracer) return out class PythonKeyTracer(Tracer): def __init__(self): super().__init__(autowrap_modules=()) self.tensor_tracker = WeakTensorKeyDictionary() self.symnode_tracker = weakref.WeakKeyDictionary() # type: ignore[var-annotated] # In general, we don't want to make modules leaves. In principle, users of # this tracer might want to override this in order to turn a couple specific # modules into leaves in the traced graph. def call_module( self, m: torch.nn.Module, forward: Callable[..., Any], args: Tuple[Any, ...], kwargs: Dict[str, Any] ) -> Any: return forward(*args, **kwargs) # We don't want to turn getattr calls into proxies. So we just return the actual value. def getattr(self, attr, attr_val, parameter_proxy_cache): return attr_val def create_arg(self, a: Any): if isinstance(a, torch.nn.Parameter): for n, p in self.root.named_parameters(): if a is p: return self.create_node('get_attr', n, (), {}) qualname: Optional[str] = None if not qualname: i = 0 while True: qualname = f'_param_constant{i}' if not hasattr(self.root, qualname): break i += 1 setattr(self.root, qualname, a) return self.create_node('get_attr', qualname, (), {}) elif isinstance(a, (SymInt, SymFloat, SymBool)): assert a.node.constant is not None return a.node.constant return super().create_arg(a) def unwrap_proxy(self, e): if isinstance(e, torch.Tensor): return get_proxy_slot(e, self, e, lambda e: e.proxy) elif isinstance(e, (torch.SymInt, torch.SymFloat, torch.SymBool)): return get_proxy_slot(e.node, self, e, lambda e: e()) else: return e @torch._disable_dynamo def dispatch_trace( root: Union[torch.nn.Module, Callable], tracer: Tracer, concrete_args: Optional[Tuple[Any, ...]] = None, ) -> GraphModule: graph = tracer.trace(root, concrete_args) name = root.__class__.__name__ if isinstance(root, torch.nn.Module) else root.__name__ return GraphModule(tracer.root, graph, name) def wrap_key(f, tensors, tracer, pre_dispatch: bool): flat_tensors, tensors_spec = pytree.tree_flatten(tensors) dk = torch._C.DispatchKey.PreDispatch if pre_dispatch else None @functools.wraps(f) def wrapped(*proxies): flat_proxies, proxies_spec = pytree.tree_flatten(proxies) assert len(flat_proxies) == len(flat_tensors) with _pop_mode_temporarily(dk) as m: assert isinstance(m, ProxyTorchDispatchMode) track_tensor_tree(flat_tensors, flat_proxies, constant=None, tracer=tracer) out = f(*tensors) out = pytree.tree_map_only( torch.Tensor, lambda t: get_proxy_slot(t, tracer, t, lambda x: x.proxy), out ) out = pytree.tree_map_only( (SymInt, SymFloat, SymBool), lambda t: get_proxy_slot(t.node, tracer)(), out ) return out return wrapped ORIGINAL_ATEN = None @contextmanager def set_original_aten_op(func): global ORIGINAL_ATEN if ORIGINAL_ATEN is None and fx_traceback.has_preserved_node_meta(): ORIGINAL_ATEN = func fx_traceback.current_meta['original_aten'] = func try: yield finally: ORIGINAL_ATEN = None fx_traceback.current_meta['original_aten'] = None else: yield # This mode is **only** used for pre_dispatch tracing. # In particular, we need to make sure that autograd/autocast API's # that do not desugar into dispatcher operators stay in the graph. class PreDispatchTorchFunctionMode(TorchFunctionMode): def __init__(self, tracer): self.tracer = tracer def __torch_function__(self, func, types, args=(), kwargs=None): kwargs = kwargs or {} pre_dispatch_ops = [ torch._C._set_grad_enabled, torch.amp._enter_autocast, torch.amp._exit_autocast, ] if func in pre_dispatch_ops: return self.tracer.create_node("call_function", func, args, {}) # Don't actually run the function! We just want to trace the calls # into a graph. We don't actualy want to change global autograd state. return func(*args, **kwargs) class ProxyTorchDispatchMode(TorchDispatchMode): def __init__(self, tracer, tracing_mode, pre_dispatch=False, _allow_fake_constant=False): dk = torch._C.DispatchKey.PreDispatch if pre_dispatch else None super().__init__(dk) self.tracer = tracer self.tracing_mode = tracing_mode self.enable_tracing = True self.pre_dispatch = pre_dispatch self._allow_fake_constant = _allow_fake_constant self.is_inside_mode = False self.sym_mode = ProxySymDispatchMode(tracer) self.trace_state = {} self._managers = [] @count def __torch_dispatch__(self, func, types, args=(), kwargs=None): with self.sym_mode.enable(False), set_original_aten_op(func): return self.inner_torch_dispatch(func, types, args, kwargs) def __enter__(self): # sym mode first, then us... m = self.sym_mode.enable(True) self._managers.append(m) m.__enter__() return super().__enter__() def __exit__(self, exc_type, exc_value, traceback): m = self._managers.pop() # ...exit us first, then sym mode b = super().__exit__(exc_type, exc_value, traceback) if not b: return m.__exit__(exc_type, exc_value, traceback) else: return m.__exit__(None, None, None) def inner_torch_dispatch(self, func, types, args=(), kwargs=None): if not self.enable_tracing: return func(*args, **kwargs) if func in [prim.device.default]: return func(*args, **kwargs) return proxy_call(self, func, self.pre_dispatch, args, kwargs) class ProxySymDispatchMode(SymDispatchMode): def __init__(self, tracer): super().__init__() self.tracer = tracer # When false, we don't trace operations. If you do this, you MUST # call track_tensor/track_tensor_tree on all results of the operation # to ensure we can adequately track the results self.enable_tracing = True @contextmanager def enable(self, b): old = self.enable_tracing self.enable_tracing = b try: yield finally: self.enable_tracing = old def _compute_proxy(self, func, args, out: Union[SymInt, SymFloat, SymBool]): n_args = tuple( get_proxy_slot(a.node, self.tracer)().node if isinstance(a, py_sym_types) else a for a in args ) # func doesn't have a __torch_function__ that Proxy can interpose, so # we gotta do it manually n_out = self.tracer.create_node("call_function", func, n_args, {}) p_out = fx.Proxy(n_out, self.tracer) set_meta(p_out, out) return p_out def __sym_dispatch__(self, func, types, args, kwargs): if not self.enable_tracing: return func(*args, **kwargs) # Peephole optimize multiply by one # NB: be careful not to trigger guards here! if func == operator.mul: if isinstance(args[1], int) and args[1] == 1: return args[0] elif isinstance(args[0], int) and args[0] == 1: return args[1] # For speed, we assume there are no nested data structures # (otherwise we could use tree_map) # We also assume there are no keyword arguments. assert not kwargs out = func(*args, **kwargs) # If func returned a constant, we don't need to trace; we have # determined that the result is constant (no matter if the inputs # were symbolic) and it is no longer necessary to trace the # computation. This could occur if func triggered some guards. if isinstance(out, py_sym_types): # Delays tracing out the proxies on this op until we actually need it p_out_thunk = thunkify(self._compute_proxy, func=func, args=args, out=out) set_proxy_slot(out.node, self.tracer, p_out_thunk) return out # TODO: I'm not sure what the point of this class is; you can just # make_fx through a regular Interpreter class DecompositionInterpreter(torch.fx.Interpreter): def __init__(self, module: torch.fx.GraphModule, new_graph: torch.fx.Graph, decomposition_table=None, **kwargs): super().__init__(module, **kwargs) self.new_graph = new_graph self.tracer = torch.fx.proxy.GraphAppendingTracer(self.new_graph) # Blegh self.tracer.tensor_tracker = WeakTensorKeyDictionary() # type: ignore[attr-defined] self.tracer.symnode_tracker = weakref.WeakKeyDictionary() # type: ignore[attr-defined] self.decomposition_table = decomposition_table if self.decomposition_table is None: self.decomposition_table = {} self.mode = ProxyTorchDispatchMode(self.tracer, tracing_mode="real") def placeholder(self, target, args, kwargs): out = super().placeholder(target, args, kwargs) proxy = torch.fx.Proxy(self.new_graph.placeholder(target), self.tracer) track_tensor_tree(out, proxy, constant=None, tracer=self.tracer) # TODO handle case where the first character of target is '*' return out def get_attr(self, target, args, kwargs): out = super().get_attr(target, args, kwargs) proxy = torch.fx.Proxy(self.new_graph.get_attr(target), self.tracer) track_tensor_tree(out, proxy, constant=None, tracer=self.tracer) return out # call_function, call_method, call_module get traced automatically by the outer mode. def output(self, target, args, kwargs): out = super().output(target, args, kwargs) def unwrap(e): return get_proxy_slot(e, self.tracer, e, lambda x: x.proxy.node) self.new_graph.output(pytree.tree_map(unwrap, out)) return out def run(self, *args, **kwargs): # Should enter the mode at least once for being able to restore it later # See: https://github.com/pytorch/pytorch/pull/82549#discussion_r934782025 with decompose(self.decomposition_table), self.mode: return super().run(*args, **kwargs) def wrapper_and_args_for_make_fx(func, args, kwargs): # make_fx doesn't support kwargs, so we need to do this flattening # and then unflatten the args before calling func flat_args, spec = pytree.tree_flatten((args, kwargs)) def wrapped(flat_args): fn_args, fn_kwargs = pytree.tree_unflatten(flat_args, spec) return func(*fn_args, **fn_kwargs) return wrapped, flat_args @contextmanager def disable_autocast_cache(): old_value = torch.is_autocast_cache_enabled() torch.set_autocast_cache_enabled(False) try: yield finally: torch.set_autocast_cache_enabled(old_value) def make_fx(f, decomposition_table=None, tracing_mode="real", _allow_non_fake_inputs=False, *, pre_dispatch=False, _allow_fake_constant=False): assert tracing_mode in ["real", "fake", "symbolic"] if decomposition_table is None: decomposition_table = {} @functools.wraps(f) def wrapped(*args): phs = pytree.tree_map(lambda _: fx.PH, args) # type: ignore[attr-defined] fx_tracer = PythonKeyTracer() fake_tensor_mode: Any = nullcontext() if tracing_mode == "real": fake_tensor_mode = nullcontext() elif tracing_mode == "fake": import torch._dynamo fake_tensor_mode = torch._dynamo.utils.detect_fake_mode(args) if fake_tensor_mode is None: fake_tensor_mode = FakeTensorMode( allow_fallback_kernels=True, allow_non_fake_inputs=_allow_non_fake_inputs) elif tracing_mode == "symbolic": import torch._dynamo fake_tensor_mode = torch._dynamo.utils.detect_fake_mode(args) if fake_tensor_mode is None: shape_env = ShapeEnv() fake_tensor_mode = FakeTensorMode( allow_fallback_kernels=False, allow_non_fake_inputs=_allow_non_fake_inputs, shape_env=shape_env) else: shape_env = fake_tensor_mode.shape_env assert shape_env is not None, "shape_env should be set if tracing with 'symbolic'" else: raise AssertionError(f"Unexpected tracing type: {tracing_mode}") python_dispatcher_mode: Any = nullcontext() pre_dispatch_mode: Any = nullcontext() # pre-autograd tracing uses per-dispatch-key modes, # which requires the python dispatcher if tracing_mode == "symbolic" or pre_dispatch: python_dispatcher_mode = enable_python_dispatcher() if pre_dispatch: pre_dispatch_mode = enable_pre_dispatch() proxy_function_mode: Any = nullcontext() if pre_dispatch: proxy_function_mode = PreDispatchTorchFunctionMode(fx_tracer) proxy_mode = ProxyTorchDispatchMode(fx_tracer, tracing_mode, pre_dispatch=pre_dispatch, _allow_fake_constant=_allow_fake_constant) arg_count = 0 def wrap_fake(x): nonlocal arg_count if isinstance(x, torch.Tensor): # TODO: it would be nice to line these up with the names # FX will choose for the placeholders, but we don't # actually know what the names will be at this point yet # NB: the Source here is actually meaningless from torch._dynamo.source import ConstantSource source = ConstantSource(f"input{arg_count}") arg_count += 1 return fake_tensor_mode.from_tensor(x, source=source) # type: ignore[attr-defined] return x sym_mode = proxy_mode.sym_mode wrap_fn_map = { "real": lambda x: x, "fake": wrap_fake, "symbolic": wrap_fake, } args = pytree.tree_map(wrap_fn_map[tracing_mode], args) if not hasattr(inspect.unwrap(f), '__code__') or inspect.unwrap(f).__code__.co_flags & inspect.CO_VARARGS: # FX doesn't support varargs, so we gotta fake up a wrapper # TODO: Would be nice to fix this at the source... func = fake_signature(f, len(phs)) else: func = f # We disable the autocast cache as the autocast cache causes type conversions on parameters to # check a cache, which introduces untracked tensors into the graph # # We also disable tracing by any other tensor proxy-based tracers except the current. The # purpose of `make_fx` is to produce graphmodules as a side effect; its internal execution is # thus irrelevant to any external functional trace. with decompose(decomposition_table), fake_tensor_mode, python_dispatcher_mode, pre_dispatch_mode, proxy_function_mode, \ sym_mode, proxy_mode, disable_autocast_cache(), disable_proxy_modes_tracing(enable_current=True): t = dispatch_trace(wrap_key(func, args, fx_tracer, pre_dispatch), tracer=fx_tracer, concrete_args=tuple(phs)) # TODO: kind of a bad way to do it, should maybe figure out a better way if tracing_mode == "symbolic": t.shape_env = shape_env # type: ignore[assignment] return t return wrapped def get_torch_dispatch_modes(): return torch.utils._python_dispatch._get_current_dispatch_mode_stack() def get_innermost_proxy_mode(): for m in reversed(torch.utils._python_dispatch._get_current_dispatch_mode_stack()): if isinstance(m, ProxyTorchDispatchMode): return m return None @contextlib.contextmanager def disable_proxy_modes_tracing(enable_current=False): modes = get_torch_dispatch_modes() proxy_tensor_modes = [m for m in modes if isinstance(m, ProxyTorchDispatchMode)] if enable_current: proxy_tensor_modes = proxy_tensor_modes[:-1] olds = [(m.enable_tracing, m.sym_mode.enable_tracing) for m in proxy_tensor_modes] for proxy_mode in proxy_tensor_modes: proxy_mode.enable_tracing = False proxy_mode.sym_mode.enable_tracing = False try: yield finally: for proxy_mode, (old, old_sym) in zip(proxy_tensor_modes, olds): proxy_mode.enable_tracing = old proxy_mode.sym_mode.enable_tracing = old_sym def get_isolated_graphmodule(func, args, kwargs, tracing_mode="real"): """A helper function used to get the GraphModule for the given func. It's expected to be used in the ProxyTensor tracing context. It detaches the args and kwargs from the current tracer so that the trace of the current graph module can be created without any side-effects. """ wrapped, all_args = wrapper_and_args_for_make_fx(func, args, kwargs) with disable_proxy_modes_tracing(): gm = make_fx(wrapped, tracing_mode=tracing_mode)(all_args) return gm