from __future__ import annotations import contextlib import dataclasses import dis import functools import inspect import logging import os import sys import textwrap import threading import traceback import types import warnings import weakref from enum import Enum from os.path import dirname, join from typing import Any, Callable, Dict, List, Optional, Set, Tuple, TYPE_CHECKING, Union from unittest.mock import patch import torch import torch.fx import torch.utils._pytree as pytree import torch.utils.checkpoint from torch import _guards from torch._subclasses import fake_tensor from torch.fx.experimental.proxy_tensor import make_fx from torch.fx.graph import _PyTreeCodeGen, _PyTreeInfo from torch.nn.parallel.distributed import DistributedDataParallel from ..fx import GraphModule from .backends.registry import CompilerFn, lookup_backend from .hooks import Hooks if TYPE_CHECKING: from torch._C._dynamo.eval_frame import ( # noqa: F401 reset_code, set_eval_frame, set_guard_error_hook, set_guard_fail_hook, skip_code, unsupported, ) else: for name in dir(torch._C._dynamo.eval_frame): if name.startswith("__"): continue globals()[name] = getattr(torch._C._dynamo.eval_frame, name) from . import config, convert_frame, external_utils, skipfiles, utils from .exc import CondOpArgsMismatchError, ResetRequired, UserError, UserErrorType from .mutation_guard import install_generation_tagging_init from .types import DynamoCallback from .utils import compile_times log = logging.getLogger(__name__) from torch._dispatch.python import enable_python_dispatcher from torch.fx.experimental import proxy_tensor always_optimize_code_objects = utils.ExactWeakKeyDictionary() null_context = contextlib.nullcontext import sympy from torch.fx.experimental.symbolic_shapes import ( ConstraintViolationError, StrictMinMaxConstraint, ) from torch.utils._sympy.value_ranges import ValueRanges # See https://github.com/python/typing/pull/240 class Unset(Enum): token = 0 unset = Unset.token compile_lock = threading.RLock() most_recent_backend: Optional[CompilerFn] = None DONT_WRAP_FILES = { # For tracing into fx modules inspect.getsourcefile(GraphModule), join(dirname(dirname(__file__)), "onnx/_internal/fx/dynamo_graph_extractor.py"), } class OptimizedModule(torch.nn.Module): """ Wraps the original nn.Module object and later patches its forward method to optimized self.forward method. """ def __init__(self, mod: torch.nn.Module, dynamo_ctx): super().__init__() # Installs the params/buffer self._orig_mod = mod self.dynamo_ctx = dynamo_ctx self._initialize() def _initialize(self): # Do this stuff in constructor to lower overhead slightly if isinstance(self._orig_mod.forward, types.MethodType) and skipfiles.check( inspect.getsourcefile(self._orig_mod.forward) ): # This may be a torch.nn.* instance in skipfiles.py which # won't trigger a frame evaluation workaround to add an extra # frame we can capture self.forward = self.dynamo_ctx(external_utils.wrap_inline(self._orig_mod)) else: # Invoke hooks outside of dynamo then pickup the inner frame self.forward = self.dynamo_ctx(self._orig_mod.__call__) if hasattr(self._orig_mod, "_initialize_hook"): self._forward = self.forward self.forward = self._call_lazy_check def __getstate__(self): state = dict(self.__dict__) state.pop("forward", None) state.pop("__call__", None) return state def __setstate__(self, state): self.__dict__ = state self._initialize() def __getattr__(self, name): if name == "_orig_mod": return self._modules["_orig_mod"] return getattr(self._orig_mod, name) def _call_lazy_check(self, *args, **kwargs): if hasattr(self._orig_mod, "_initialize_hook"): # In the case of a lazy module, we want to run # the pre-hooks which initialize it. # Afterwards, lazy module deletes its pre-hooks # to avoid treating it as lazy on subsequent recompile. assert len(kwargs) == 0 self._orig_mod._infer_parameters(self._orig_mod, args) return self._forward(*args, **kwargs) def __dir__(self): orig_mod_attrs = self._orig_mod.__dir__() return orig_mod_attrs + [ attr for attr in super().__dir__() if attr not in orig_mod_attrs ] def remove_from_cache(f): """ Make sure f.__code__ is not cached to force a recompile """ if isinstance(f, types.CodeType): reset_code(f) elif hasattr(f, "__code__"): reset_code(f.__code__) elif hasattr(getattr(f, "forward", None), "__code__"): reset_code(f.forward.__code__) else: from . import reset reset() log.warning("could not determine __code__ for %s", f) def nothing(): pass def innermost_fn(fn): """ In case of nesting of _TorchDynamoContext calls, find the innermost function. TorchDynamo caches on fn.__code__ object, so its necessary to find the innermost function to pass on the optimize, run, disable etc. """ unaltered_fn = fn while hasattr(unaltered_fn, "_torchdynamo_orig_callable"): unaltered_fn = unaltered_fn._torchdynamo_orig_callable assert callable(unaltered_fn) return unaltered_fn @contextlib.contextmanager def enable_dynamic(enable: bool = True, export: bool = False): if not enable: yield return # dynamic=True used to mean fully dynamic. However, with automatic dynamic, the default flipped to # deriving dynamism. For back compat, and forward compat for when dynamic=True is default, we take # dynamic=True here to mean "fully dynamic from the start". with config.patch(assume_static_by_default=False): yield class _TorchDynamoContext: def __init__( self, callback: DynamoCallback, on_enter=nothing, backend_ctx_ctor=null_context, patch_fn=nothing, first_ctx=False, *, export=False, dynamic=False, ): super().__init__() assert callable(callback) or callback is False or callback is None self.callback: DynamoCallback = callback self.prior: Union[Unset, DynamoCallback] = unset self.on_enter = on_enter self.extra_ctx_ctor = backend_ctx_ctor self.first_ctx = first_ctx self.export = export self.dynamic = dynamic patch_fn() def __enter__(self): if config.raise_on_ctx_manager_usage: raise RuntimeError( "torch._dynamo.optimize(...) is used with a context manager. " "Please refer to https://github.com/pytorch/torchdynamo#usage-example " "to use torch._dynamo.optimize(...) as an annotation/decorator. " ) self.on_enter() self.prior = set_eval_frame(self.callback) self.backend_ctx = self.extra_ctx_ctor() self.backend_ctx.__enter__() self.dynamic_ctx = enable_dynamic(self.dynamic, self.export) self.dynamic_ctx.__enter__() def __exit__(self, exc_type, exc_val, exc_tb): assert self.prior is not unset set_eval_frame(self.prior) self.prior = unset # TODO: This is totally not the right way to chain contexts manually self.dynamic_ctx.__exit__(exc_type, exc_val, exc_tb) self.backend_ctx.__exit__(exc_type, exc_val, exc_tb) def __call__(self, fn): fn = innermost_fn(fn) # Optimize the forward method of torch.nn.Module object if isinstance(fn, torch.nn.Module): mod = fn new_mod = OptimizedModule(mod, self) # Save the function pointer to find the original callable while nesting # of decorators. new_mod._torchdynamo_orig_callable = mod.forward return new_mod assert callable(fn) try: filename = inspect.getsourcefile(fn) except TypeError: filename = None if ( (filename is None or skipfiles.check(filename)) and ( getattr(fn, "__name__", "") not in ["_call_impl", "_wrapped_call_impl"] ) and filename not in DONT_WRAP_FILES ): # call to a builtin without a frame for us to capture fn = external_utils.wrap_inline(fn) callback = self.callback on_enter = self.on_enter backend_ctx_ctor = self.extra_ctx_ctor @functools.wraps(fn) def _fn(*args, **kwargs): if ( not isinstance(self, DisableContext) and torch.fx._symbolic_trace.is_fx_tracing() ): if config.error_on_nested_fx_trace: raise RuntimeError( "Detected that you are using FX to symbolically trace " "a dynamo-optimized function. This is not supported at the moment." ) else: return fn(*args, **kwargs) on_enter() prior = set_eval_frame(callback) backend_ctx = backend_ctx_ctor() backend_ctx.__enter__() dynamic_ctx = enable_dynamic(self.dynamic, self.export) dynamic_ctx.__enter__() try: return fn(*args, **kwargs) finally: set_eval_frame(prior) dynamic_ctx.__exit__(None, None, None) backend_ctx.__exit__(None, None, None) # hooks to properly handle inlining if isinstance(self, DisableContext): _fn._torchdynamo_disable = True # type: ignore[attr-defined] else: _fn._torchdynamo_inline = fn # type: ignore[attr-defined] # Save the function pointer to find the original callable while nesting # of decorators. _fn._torchdynamo_orig_callable = fn # type: ignore[attr-defined] # If the function is called using torch._dynamo.optimize decorator, we # should prevent any type of skipping. if callback not in (None, False): if not hasattr(fn, "__code__"): raise RuntimeError( textwrap.dedent( """ torch._dynamo.optimize is called on a non function object. If this is a callable class, please wrap the relevant code into a function and optimize the wrapper function. >> class CallableClass: >> def __init__(self): >> super().__init__() >> self.relu = torch.nn.ReLU() >> >> def __call__(self, x): >> return self.relu(torch.sin(x)) >> >> def print_hello(self): >> print("Hello world") >> >> mod = CallableClass() If you want to optimize the __call__ function and other code, wrap that up in a function >> def wrapper_fn(x): >> y = mod(x) >> return y.sum() and then optimize the wrapper_fn >> opt_wrapper_fn = torch._dynamo.optimize(wrapper_fn) """ ) ) always_optimize_code_objects[fn.__code__] = True return _fn class OptimizeContext(_TorchDynamoContext): @staticmethod def _different_backend(old, new): return not (old == new or old is None) def __init__( self, callback, backend_ctx_ctor, first_ctx=False, *, export=False, dynamic=False, ): def on_enter(): global most_recent_backend if OptimizeContext._different_backend(most_recent_backend, compiler_fn): if config.raise_on_backend_change: raise ResetRequired() else: warnings.warn( "changing options to `torch.compile()` may require " "calling `torch._dynamo.reset()` to take effect" ) most_recent_backend = compiler_fn install_generation_tagging_init() compiler_fn = innermost_fn(callback) super().__init__( callback=callback, on_enter=on_enter, backend_ctx_ctor=backend_ctx_ctor, patch_fn=TorchPatcher.patch, first_ctx=first_ctx, export=export, dynamic=dynamic, ) class RunOnlyContext(_TorchDynamoContext): def __init__(self): # cudagraph trees relies on generation increment def on_enter(): torch._dynamo.mutation_guard.GenerationTracker.generation += 1 super().__init__(callback=False, on_enter=on_enter) class DisableContext(_TorchDynamoContext): def __init__(self): super().__init__(callback=None) def first_real_inst_idx(code): if sys.version_info < (3, 11): return 0 for inst in dis.get_instructions(code): if inst.opname == "RESUME": return inst.offset // 2 raise RuntimeError("RESUME instruction not found in code") def catch_errors_wrapper(callback, hooks: Hooks): @functools.wraps(callback) def catch_errors(frame, cache_size, frame_state): assert frame_state is not None if ( # TODO: the first condition is not covered by any test frame.f_lasti >= first_real_inst_idx(frame.f_code) or skipfiles.check(frame.f_code.co_filename) or config.disable ): log.debug("skipping %s %s", frame.f_code.co_name, frame.f_code.co_filename) return None if frame.f_code.co_filename == "" and frame.f_code.co_name == "__new__": # nametuple constructor return None if config.optimize_ddp: ddp_module = DistributedDataParallel._get_active_ddp_module() if ddp_module: with compile_lock: from torch._dynamo.backends.distributed import DDPOptimizer ddp_optimizer = DDPOptimizer( bucket_bytes_cap=ddp_module.bucket_bytes_cap, backend_compile_fn=callback._torchdynamo_orig_callable, ) hijacked_callback = convert_frame.convert_frame( ddp_optimizer.compile_fn, hooks=hooks, ) return hijacked_callback(frame, cache_size, hooks, frame_state) with compile_lock: return callback(frame, cache_size, hooks, frame_state) catch_errors._torchdynamo_orig_callable = callback # type: ignore[attr-defined] return catch_errors def _optimize_catch_errors( compile_fn, hooks: Hooks, backend_ctx_ctor=null_context, export=False, dynamic=False ): return OptimizeContext( catch_errors_wrapper(compile_fn, hooks), backend_ctx_ctor=backend_ctx_ctor, first_ctx=True, export=export, dynamic=dynamic, ) def get_compiler_fn(compiler_fn): from .repro.after_dynamo import wrap_backend_debug if hasattr(compiler_fn, "compiler_name"): compiler_str = compiler_fn.compiler_name elif isinstance(compiler_fn, str): compiler_str = compiler_fn else: compiler_str = None compiler_fn = lookup_backend(compiler_fn) return wrap_backend_debug(compiler_fn, compiler_str) class _NullDecorator(contextlib.nullcontext): # type: ignore[type-arg] def __call__(self, fn): assert callable(fn) return fn def check_if_dynamo_supported(): if sys.platform == "win32": raise RuntimeError("Windows not yet supported for torch.compile") if sys.version_info >= (3, 12): raise RuntimeError("Python 3.12+ not yet supported for torch.compile") def is_dynamo_supported(): try: check_if_dynamo_supported() return True except Exception: return False def optimize( backend="inductor", *, nopython=False, guard_export_fn=None, guard_fail_fn=None, disable=False, dynamic=False, ): """ The main entrypoint of TorchDynamo. Do graph capture and call backend() to optimize extracted graphs. Args: backend: One of the two things: - Either, a function/callable taking a torch.fx.GraphModule and example_inputs and returning a python callable that runs the graph faster. One can also provide additional context for the backend, like torch.jit.fuser("fuser2"), by setting the backend_ctx_ctor attribute. See AOTAutogradMemoryEfficientFusionWithContext for the usage. - Or, a string backend name in `torch._dynamo.list_backends()` nopython: If True, graph breaks will be errors and there will be a single whole-program graph. disable: If True, turn this decorator into a no-op dynamic: If True, turn on dynamic shapes support Example Usage:: @torch._dynamo.optimize() def toy_example(a, b): ... """ check_if_dynamo_supported() # Note: The hooks object could be global instead of passed around, *however* that would make # for a confusing API usage and plumbing story wherein we nest multiple .optimize calls. # There is some prior art around this, w/r/t nesting backend calls are enforced to be the same # compiler, however, this feels onerous for callback and hooks, and it feels better to give our users an # easier to understand UX at the cost of a little more plumbing on our end. hooks = Hooks(guard_export_fn=guard_export_fn, guard_fail_fn=guard_fail_fn) torch._C._log_api_usage_once("torch._dynamo.optimize") if disable or os.environ.get("TORCHDYNAMO_DISABLE", "") == "1": return _NullDecorator() backend = get_compiler_fn(backend) # Find if backend has any extra context manager backend_ctx_ctor = getattr(backend, "backend_ctx_ctor", null_context) if nopython: return optimize_assert( backend, dynamic=dynamic, hooks=hooks, ) return _optimize_catch_errors( convert_frame.convert_frame(backend, hooks=hooks), hooks, backend_ctx_ctor, dynamic=dynamic, ) # TODO(voz): Consider making "explain" output alongside a run / part of a run @patch("torch._dynamo.symbolic_convert.explain", True) def explain(f, *args, **kwargs): # TODO(voz): Do we want a decorator for this? from . import reset reset() graphs: List[torch.fx.GraphModule] = [] break_reasons: List[Any] = [] op_count: int = 0 ops_per_graph: List[torch.fx.Node] = [] out_guards: List[_guards.Guard] = [] def dynamo_graph_accumulating_compiler(gm: torch.fx.GraphModule, example_inputs): from .backends.debugging import _explain_graph_detail nonlocal graphs nonlocal op_count nonlocal ops_per_graph nonlocal break_reasons gm, graphs, op_count, ops_per_graph, break_reasons = _explain_graph_detail( gm, graphs, op_count, ops_per_graph, break_reasons ) return gm.forward def guard_export_print(guards): nonlocal out_guards out_guards.extend(guards) with patch(f"{__name__}.most_recent_backend", None): opt_f = optimize( dynamo_graph_accumulating_compiler, nopython=False, guard_export_fn=guard_export_print, )(f) # TODO(voz): We may have instances of `f` that mutate inputs, we should track sideffects and reject. opt_f(*args, **kwargs) graph_count = len(graphs) # For the explanation summary, dedupe reasons by the innermost stack frame and dedupe by it. deduped_reasons = {} for reason in break_reasons: innermost_frame = reason.user_stack[-1] # __repr__ uniquely identifies a FrameSummary so we can use it for deduping deduped_reasons[repr(innermost_frame)] = reason formatted_list = "" for idx, break_reason in enumerate(deduped_reasons.values()): formatted_stack = "".join(traceback.format_list(break_reason.user_stack)) msg = f"{idx + 1}. Reason: {break_reason.reason}\n User Stack: {formatted_stack}\n" formatted_list += msg graph_break_count = graph_count - 1 compile_time = compile_times(repr="str") # TODO(voz): Do we want a decorator for this? reset() from .backends.debugging import ExplainOutput return ExplainOutput( graphs, graph_count, graph_break_count, break_reasons, op_count, ops_per_graph, out_guards, compile_time, ) @dataclasses.dataclass class ConstraintTarget: """ This represents input tensor dimensions. Don't create this class directly; instead, use :func:`torch._export.dynamic_dim`. """ w_tensor: weakref.ReferenceType[torch.Tensor] # TODO: We don't need t_id; we can get it off of w_tensor t_id: int dim: int @dataclasses.dataclass class Constraint(ConstraintTarget): """ This represents constraints on input tensor dimensions, e.g., requiring them to be fully polymorphic or within some range. Don't create this class directly; instead, use :func:`torch._export.dynamic_dim`. """ # NOTE(avik): In the future, this could be Union[StrictMinMaxConstraint, ] constraint_range: StrictMinMaxConstraint # Represent that `constraint_range` is shared with another ConstraintTarget, which # typically arises because of a specified equality with another dynamic dimension. shared: Optional[ConstraintTarget] = None def _clone_with_range(self, lower=2, upper=sympy.oo): constraint_range = StrictMinMaxConstraint( vr=self.constraint_range.vr & ValueRanges(lower=lower, upper=upper), warn_only=False, ) return Constraint( self.w_tensor, self.t_id, self.dim, constraint_range, self.shared ) def __ge__(self, lower): return self._clone_with_range(lower=lower) def __gt__(self, lower): return self._clone_with_range(lower=lower + 1) def __le__(self, upper): return self._clone_with_range(upper=upper) def __lt__(self, upper): return self._clone_with_range(upper=upper - 1) def __bool__(self): # NOTE(avik): We do not support compound expressions like a <= x <= b. # This is because Python implicitly desugars them into bool(a <= x) and bool(x <= b), # and moreover, enforces that any overload of __bool__ must return True or False. # FWIW, sympy also raises TypeError in this case. raise TypeError( "Cannot determine truth value of Constraint. " "If you are trying to combine Constraints with logical connectives, " "you can specify them separately instead." ) @property def serializable_spec(self): # We need a serialization compatible format of the constraint so that it # can be savedin the graph module w/o breaking the module serialization. # The saved constraints will be used directly for the post-exporting pass # that converts constraints to runtime assertion. The saved constraints # will not be saved in the serialized module. # TODO: A better way is needed. Currently we use 't_id' to map the constraint, # which is not reliable return { "t_id": self.t_id, "dim": self.dim, "min": self.constraint_range.vr.lower, "max": self.constraint_range.vr.upper, "shared": ( None if self.shared is None else { "t_id": self.shared.t_id, "dim": self.shared.dim, } ), } def __eq__(self, other): constraint_range = StrictMinMaxConstraint( vr=self.constraint_range.vr & other.constraint_range.vr, warn_only=False, ) return Constraint( self.w_tensor, self.t_id, self.dim, constraint_range, shared=ConstraintTarget(other.w_tensor, other.t_id, other.dim), ) class FlattenInputOutputSignature(torch.fx.interpreter.Transformer): def __init__( self, m: torch.fx.GraphModule, flat_args: Tuple[Any], matched_input_elements_positions: List[int], matched_output_elements_positions: List[int], example_fake_inputs: List[torch.Tensor], fake_mode: Optional[fake_tensor.FakeTensorMode] = None, ): super().__init__(m) matched_input_elements_to_fake = { val: example_fake_inputs[ix] for ix, val in enumerate(matched_input_elements_positions) } self.new_args = [] for i in range(0, len(flat_args)): arg = super(FlattenInputOutputSignature, self).placeholder( f"arg{i}", (), {} ) if i in matched_input_elements_to_fake: arg.node.meta["val"] = matched_input_elements_to_fake[i] else: # Fill node.mata["val"] with faketensor from the input, # if it's not found in matched_input_elements_positions if fake_mode is not None and isinstance(flat_args[i], torch.Tensor): arg.node.meta["val"] = fake_mode.from_tensor(flat_args[i]) self.new_args.append(arg) self.old_args_gen = (self.new_args[i] for i in matched_input_elements_positions) self.matched_output_elements_positions = matched_output_elements_positions def placeholder(self, target, args, kwargs): arg = next(self.old_args_gen) if "val" in self.current_node.meta: arg.node.meta["val"] = self.current_node.meta["val"] if "tensor_dict" in self.current_node.meta: arg.node.meta["tensor_dict"] = self.current_node.meta["tensor_dict"] return arg def output(self, target, args, kwargs): dynamo_result_flat = args[0] lookup = [*dynamo_result_flat, *self.new_args] new_result_flat = [lookup[i] for i in self.matched_output_elements_positions] return super().output(target, (new_result_flat,), {}) def run_node(self, n): self.current_node = n r = super().run_node(n) if "val" in self.current_node.meta: r.node.meta["val"] = self.current_node.meta["val"] return r def export( f: Callable[..., Any], *args, aten_graph: bool = False, pre_dispatch: bool = False, decomposition_table: Optional[ Dict[torch._ops.OpOverload, Callable[..., Any]] ] = None, tracing_mode: str = "symbolic", constraints: Optional[List[Constraint]] = None, assume_static_by_default: bool = False, fake_mode: fake_tensor.FakeTensorMode = None, **kwargs, ) -> Tuple[torch.fx.GraphModule, Set[_guards.Guard]]: """ Export an input function f to a format that can be executed outside of PyTorch using the FX graph. Args: f (callable): A PyTorch function to be exported. *args: Variable length argument list to be passed to the function f. aten_graph (bool): If True, exports a graph with ATen operators. If False, exports a graph with Python operators. Default is False. pre_dispatch (bool): If True, exports a graph with ATen operators, but before any logic in the PyTorch dispatcher has run. This can be useful if you want to apply further tranformations on a graph before running it through autograd, autocast, or any other functionalities that are integrated into the dispatcher. This flag is only valid if aten_graph=True is set. Default is False. decomposition_table (dict): A dictionary that maps operators to their decomposition functions. Required if aten_graph or tracing_mode is specified. Default is None. tracing_mode (str): If "symbolic", turn on dynamic shapes support. Default is "symbolic". fake_mode (fake_tensor.FakeTensorMode): Use this fake_mode instead of creating an internal one. Useful during symbolic tracing, when user input is already fakefied. Implies free fake tensors are allowed on `make_fx`. **kwargs: Arbitrary keyword arguments to be passed to the function f. Returns: A tuple of (graph, guards) Graph: An FX graph representing the execution of the input PyTorch function with the provided arguments and options. Guards: The guards we accumulated during tracing f above Raises: AssertionError: If decomposition_table is specified without setting aten_graph=True, or if graph breaks during tracing in export. AssertionError: If Dynamo input and output is not consistent with traced input/output. Note - this headerdoc was authored by ChatGPT, with slight modifications by the author. """ check_if_dynamo_supported() torch._C._log_api_usage_once("torch._dynamo.export") if decomposition_table is not None: assert ( aten_graph ), "Specifying a decomposition_table table or tracing mode is illegal without setting aten_graph=True" if pre_dispatch: assert aten_graph, "pre_dispatch=True can only be used when aten_graph=True" f = innermost_fn(f) call_to_inspect = f.forward if isinstance(f, torch.nn.Module) else f original_signature = inspect.signature(call_to_inspect) graph = None out_guards = None graph_captured_input = None graph_captured_result: Optional[Tuple[torch.Tensor, ...]] = None _allow_fake_constant: bool = ( fake_mode is not None ) # Allow fake constants during symbolic tracing def produce_matching(source_args, candidate_args): matched_elements_positions = [] dict_of_source_args = dict() for i in range(0, len(source_args)): element_id = id(source_args[i]) dict_of_source_args[element_id] = i for i in range(0, len(candidate_args)): arg = candidate_args[i] # 1-element tensor arg can be unspec int/float if isinstance(arg, torch.Tensor) and torch.numel(arg) == 1: if id(arg) in dict_of_source_args: matched_elements_positions.append(dict_of_source_args[id(arg)]) elif id(arg.item()) in dict_of_source_args: matched_elements_positions.append( dict_of_source_args[id(arg.item())] ) else: raise AssertionError( "Dynamo input/output is not consistent with traced input/output" ) else: assert ( id(arg) in dict_of_source_args ), "Dynamo input and output is a strict subset of traced input/output" matched_elements_positions.append(dict_of_source_args[id(arg)]) return matched_elements_positions def guard_export_print(guards: Set[_guards.Guard]): nonlocal out_guards assert out_guards is None, "whole graph export entails exactly one guard export" out_guards = guards example_inputs = [] def dynamo_normalization_capturing_compiler( gm: torch.fx.GraphModule, inner_example_inputs ): nonlocal graph assert ( graph is None ), "Tried to emit a second graph during export. Tracing through 'f' must produce a single graph." graph = gm nonlocal fake_mode, example_inputs fake_mode = fake_mode or _guards.detect_fake_mode(inner_example_inputs) example_inputs = inner_example_inputs def result_capturing_wrapper(*graph_inputs): nonlocal graph_captured_result nonlocal graph_captured_input graph_captured_input = graph_inputs assert graph is not None graph_captured_result = graph(*graph_inputs) return graph_captured_result return result_capturing_wrapper flat_args, in_spec = pytree.tree_flatten((args, kwargs)) remove_from_cache(f) constraint_violation_error = None if tracing_mode != "symbolic": assume_static_by_default = True with patch(f"{__name__}.most_recent_backend", None), config.patch( summarize_dim_constraints=True, specialize_int=True, assume_static_by_default=assume_static_by_default, automatic_dynamic_shapes=False, ), torch._guards.export_fake_mode(fake_mode): opt_f = optimize_assert( dynamo_normalization_capturing_compiler, hooks=Hooks( guard_export_fn=guard_export_print, guard_fail_fn=None, ), export=True, export_constraints=constraints, )(f) # TODO(voz): We may have instances of `f` that mutate inputs, we should track sideffects and reject. try: result_traced = opt_f(*args, **kwargs) except ConstraintViolationError as e: constraint_violation_error = e remove_from_cache(f) if ( (shape_env := getattr(fake_mode, "shape_env", None)) is not None and (dim_constraints := shape_env.dim_constraints) is not None and not skipfiles.check(inspect.getsourcefile(call_to_inspect)) ): dim_constraints.solve() msg = dim_constraints.prettify_results(original_signature) forced_specializations = dim_constraints.forced_specializations() if forced_specializations: msg = ( "Some dynamic dimensions need to be specialized because " "the constraints inferred for them are too complex to specify.\n" f"{forced_specializations}\n{msg}" ) if constraint_violation_error: constraint_violation_error.args = ( constraint_violation_error.args[0] + msg, ) else: if forced_specializations: constraint_violation_error = ConstraintViolationError(msg) else: log.info( "Summary of dimension constraints:%s", msg, ) # Error if we have any constraints on static values for k in shape_env.var_to_range.keys(): if isinstance(k, sympy.Integer): constraint_violation_error = ConstraintViolationError( f"{''.join(traceback.format_list(shape_env.var_to_stack[k]))}\n" "It appears that you're trying to set a constraint on a " f"value which we evaluated to have a static value of {k}. " "Scroll up to see where this constraint was set." ) if constraint_violation_error: raise constraint_violation_error assert ( graph is not None ), "Failed to produce a graph during tracing. Tracing through 'f' must produce a single graph." assert out_guards is not None, "Failed to produce guards during tracing" assert fake_mode is not None matched_input_elements_positions = produce_matching(flat_args, graph_captured_input) # NB: This is mostly hitting the cache; Dynamo already converted these example_fake_inputs = [fake_mode.from_tensor(t) for t in example_inputs] flat_results_traced, out_spec_traced = pytree.tree_flatten(result_traced) assert graph_captured_result is not None flat_both = list(graph_captured_result) + flat_args matched_output_elements_positions = produce_matching(flat_both, flat_results_traced) if aten_graph: # Running graph with interpreter is needed for propagating the stack_trace def graph_with_interpreter(*args): with torch.fx.traceback.preserve_node_meta(): return torch.fx.Interpreter(graph).run(*args) with enable_python_dispatcher(), fake_mode: try: graph = make_fx( graph_with_interpreter, decomposition_table=decomposition_table, tracing_mode="real", _allow_non_fake_inputs=True, pre_dispatch=pre_dispatch, _allow_fake_constant=_allow_fake_constant, )(*example_fake_inputs) except CondOpArgsMismatchError as e: # Wrap the internal error to the user-facing error raise UserError(UserErrorType.DYNAMIC_CONTROL_FLOW, str(e)) new_graph = FlattenInputOutputSignature( graph, flat_args, matched_input_elements_positions, matched_output_elements_positions, example_fake_inputs, fake_mode, ).transform() # Store constraints and inputs as metadata for user passes, e.g. turn constraints to runtime check new_graph.meta["input_shape_constraints"] = ( [constraint.serializable_spec for constraint in constraints] if constraints else [] ) def signature_to_fullargspec(sig: inspect.Signature): # Get a list of Parameter objects from the Signature object params = list(sig.parameters.values()) # Separate positional arguments, keyword-only arguments and varargs/varkw args = [ p.name for p in params if p.kind == inspect.Parameter.POSITIONAL_OR_KEYWORD ] kwonlyargs = [ p.name for p in params if p.kind == inspect.Parameter.KEYWORD_ONLY ] varargs = next( (p.name for p in params if p.kind == inspect.Parameter.VAR_POSITIONAL), None ) varkw = next( (p.name for p in params if p.kind == inspect.Parameter.VAR_KEYWORD), None ) # Get default values for positional arguments and keyword-only arguments defaults = tuple( p.default for p in params if p.kind == inspect.Parameter.POSITIONAL_OR_KEYWORD and p.default is not inspect.Parameter.empty ) kwonlydefaults = { p.name: p.default for p in params if p.kind == inspect.Parameter.KEYWORD_ONLY and p.default is not inspect.Parameter.empty } # Get annotations for parameters and return value annotations = {} if sig.return_annotation: annotations = {"return": sig.return_annotation} for parameter in params: annotations[parameter.name] = parameter.annotation # Return a FullArgSpec object with the extracted attributes return inspect.FullArgSpec( args, varargs, varkw, defaults, kwonlyargs, kwonlydefaults, annotations ) # Make dynamo graph to have same input/output spec as user code def argument_names(f: Callable[..., Any], *args, **kwargs) -> List[str]: fullargspec = signature_to_fullargspec(original_signature) # 1. Map `args` 1-to-1 to positional arguments in original signature. input_strs = fullargspec.args[: len(args)] if len(args) > len(fullargspec.args): # 2. If there are more arguments left in `args`, they map to varargs in original # signature. Assign names as {varargs}_0, {varargs}_1, ... assert fullargspec.varargs is not None, "More arguments than expected" input_strs += [ f"{fullargspec.varargs}_{i}" for i in range(0, len(args) - len(input_strs)) ] elif len(args) < len(fullargspec.args): # 3. If there are fewer arguments in `args` than `fullargspec.args`, # it implies these are arguments either with default values, or provided in # `kwargs`. The former can be safely ignored. Because Dynamo.export does not # export them as part of the function signature. The latter will be handled # in the next step. for unprovided_arg in fullargspec.args[ len(args) : -len(fullargspec.defaults or []) ]: assert unprovided_arg in kwargs, f"Missing argument {unprovided_arg}" # 4. Keyword arguments provided in `kwargs`. input_strs += list(kwargs.keys()) # 5. Keyword-only arguments with default values if not provided are not exported # as part of the function signature. for kwonly_arg in fullargspec.kwonlyargs: kwonlydefaults = fullargspec.kwonlydefaults or {} assert ( kwonly_arg in kwargs or kwonly_arg in kwonlydefaults ), f"Missing keyword only argument {kwonly_arg}" return input_strs new_graph.graph._codegen = _PyTreeCodeGen( _PyTreeInfo( argument_names(f, *args, **kwargs), in_spec, out_spec_traced, ) ) new_graph.recompile() return (new_graph, out_guards) def assume_constant_result(fn): fn._dynamo_marked_constant = True return fn def optimize_assert( backend, *, hooks=Hooks(None, None), export=False, export_constraints=None, dynamic=False, ): """ The same as `torch._dynamo.optimize(backend, nopython=True)` """ backend = get_compiler_fn(backend) # Find if backend has any extra context manager backend_ctx_ctor = getattr(backend, "backend_ctx_ctor", null_context) return _optimize_catch_errors( convert_frame.convert_frame_assert( backend, export=export, export_constraints=export_constraints ), hooks, backend_ctx_ctor, export=export, dynamic=dynamic, ) def run(fn=None): """Don't do any dynamic compiles, just use prior optimizations""" if fn is not None: fn = innermost_fn(fn) assert callable(fn) return RunOnlyContext()(fn) return RunOnlyContext() def disable(fn=None, recursive=True): """ Decorator and context manager to disable TorchDynamo If recursive=True, Dynamo is completely skipped on the decorated function frame as well as the recursively invoked functions. If recursive=False, Dynamo skips frames associated with the function code, but still process recursively invoked frames. """ if recursive: if fn is not None: fn = innermost_fn(fn) assert callable(fn) return DisableContext()(fn) return DisableContext() else: return skip(fn) def skip(fn=None): """ Skip frames associated with the function code, but still process recursively invoked frames """ if fn is None: return skip fn = innermost_fn(fn) assert callable(fn) skip_code(fn.__code__) fn._torchdynamo_disable = True return fn class TorchPatcher: @staticmethod @functools.lru_cache(None) def patch(): # Disable TorchDynamo on some torch.* compilers generated frames torch.jit.trace = disable(torch.jit.trace) torch.jit.trace_module = disable(torch.jit.trace_module) torch.jit._get_trace_graph = disable(torch.jit._get_trace_graph) # symbolic_trace creates new frames. We disable Dynamo on such frames torch.fx._symbolic_trace.Tracer.trace = disable( torch.fx._symbolic_trace.Tracer.trace ) torch.onnx.export_to_pretty_string = disable(torch.onnx.export_to_pretty_string) torch.distributions.Distribution.set_default_validate_args(False) proxy_tensor.dispatch_trace = disable(proxy_tensor.dispatch_trace) optimizers = [ opt for opt in torch.optim.__dict__.values() if inspect.isclass(opt) and issubclass(opt, torch.optim.Optimizer) ] # disable dynamo for the wrapper that helps give dynamo hints about entering DDP if hasattr(DistributedDataParallel, "_inside_ddp_forward"): DistributedDataParallel._inside_ddp_forward = disable( DistributedDataParallel._inside_ddp_forward, recursive=False ) # Note: this excludes the optimizers that are unsupported in excluded_opts below from ..optim import ( adadelta, adagrad, adam, adamax, adamw, asgd, nadam, rmsprop, rprop, sgd, ) for opt_mod in ( adadelta, adagrad, adam, adamax, adamw, asgd, nadam, rmsprop, rprop, sgd, ): opt_name = opt_mod.__name__.split(".")[-1] multi_tensor_fn_name = f"_multi_tensor_{opt_name}" fused_fn_name = f"_fused_{opt_name}" if hasattr(opt_mod, multi_tensor_fn_name): setattr( opt_mod, multi_tensor_fn_name, disable(getattr(opt_mod, multi_tensor_fn_name)), ) if hasattr(opt_mod, fused_fn_name): setattr( opt_mod, fused_fn_name, disable(getattr(opt_mod, fused_fn_name)) ) # Note: we don't support sparsity, data-dependent control, or tracing through backwards excluded_opts = {torch.optim.SparseAdam, torch.optim.RAdam, torch.optim.LBFGS} for opt in optimizers: if opt in excluded_opts: opt.step = disable(opt.step) opt.zero_grad = disable(opt.zero_grad) opt.state_dict = disable(opt.state_dict) opt.load_state_dict = disable(opt.load_state_dict) opt.add_param_group = disable(opt.add_param_group) if hasattr(opt, "_init_group"): opt._init_group = disable(opt._init_group) # disable any currently set hooks # Note: we only want to disable the profiling hook # which is the *last* hook applied, we want to keep the no_grad hook hooked = getattr(opt.step, "hooked", False) if hooked: unwrapped_step = getattr(opt.step, "__wrapped__", None) if unwrapped_step: opt.step = unwrapped_step # disable future hooking opt.step.hooked = True # TorchDynamo does not step inside utils.checkpoint function. The flow # looks likes this # 1) TorchDynamo tries to wrap utils.checkpoint in a HigherOrderOp by # speculatively checking if the forward function is safe to trace. # 2) If yes, then Dynamo-generated Fx graph has the wrapped higher # order op. As a result, TorchDynamo does not look inside utils.checkpoint. # 3) If not, then TorchDynamo falls back to eager by performing a graph # break. And here, the following disable wrapper ensures that # TorchDynamo does not trigger again on the frames created by # utils.checkpoint innards. torch.utils.checkpoint.checkpoint = disable(torch.utils.checkpoint.checkpoint) torch._dynamo.variables.lists._register_dynamo_list_to_tree_spec() torch._dynamo.variables.lists._register_dynamo_tuple_to_tree_spec() torch._dynamo.variables.dicts._register_dynamo_dict_to_tree_spec() @staticmethod def suppress_torch_distributed_warnings(fn): def inner_fn(*args, **kwargs): warnings.filterwarnings( "ignore", category=UserWarning, module="torch.distributed" ) return fn(*args, **kwargs) return inner_fn