import copy import dataclasses import functools import io import json import pathlib import re import sys import types import warnings import weakref import zipfile from collections import OrderedDict from contextlib import contextmanager from typing import Any, Callable, Dict, List, Optional, Tuple, Union from unittest.mock import patch import sympy import torch import torch._dynamo import torch.fx import torch.fx._pytree as fx_pytree import torch.utils._pytree as pytree from torch._decomp import core_aten_decompositions, get_decompositions from torch._dispatch.python import enable_python_dispatcher from torch._dynamo.exc import UserError, UserErrorType from torch._dynamo.source import ConstantSource from torch._export.passes.collect_tracepoints_pass import CollectTracepointsPass from torch._functorch.aot_autograd import aot_export_module, GraphSignature from torch._functorch.eager_transforms import functionalize from torch._guards import detect_fake_mode from torch._ops import OpOverload from torch._subclasses.fake_tensor import FakeTensor, FakeTensorMode from torch.export import _create_constraint, _Dim, Constraint from torch.export.exported_program import ( ExportedProgram, ModuleCallEntry, ModuleCallSignature, _disable_prexisiting_fake_mode, ) from torch.export.graph_signature import ( _sig_to_specs, ArgumentSpec, ConstantArgument, ExportGraphSignature, InputKind, InputSpec, OutputKind, OutputSpec, SymIntArgument, TensorArgument, ) from torch.fx import traceback as fx_traceback from torch.fx._compatibility import compatibility from torch.fx.experimental.proxy_tensor import make_fx, maybe_disable_fake_tensor_mode from torch.fx.experimental.symbolic_shapes import ( ConstraintViolationError, GuardOnDataDependentSymNode, ShapeEnv, StrictMinMaxConstraint, ) from torch.fx.graph import _PyTreeCodeGen, _PyTreeInfo from torch.utils._sympy.value_ranges import ValueRangeError, ValueRanges from .exported_program import ( _create_stateful_graph_module, _process_constraints, CallSpec, ) from .passes.add_runtime_assertions_for_constraints_pass import ( _AddRuntimeAssertionsForInlineConstraintsPass, ) from .passes.lift_constant_tensor_pass import lift_constant_tensor_pass from .passes.remove_runtime_assertions import _RemoveRuntimeAssertionsPass from .passes.replace_sym_size_ops_pass import _replace_sym_size_ops_pass from .passes.replace_view_ops_with_view_copy_ops_pass import ( ReplaceViewOpsWithViewCopyOpsPass, ) from .wrappers import _wrap_submodules def _process_dynamic_shapes( f: Callable, args: Tuple[Any, ...], kwargs: Optional[Dict[str, Any]] = None, dynamic_shapes: Optional[Union[Dict[str, Any], Tuple[Any]]] = None, ) -> Optional[List[Constraint]]: if dynamic_shapes is None or len(dynamic_shapes) == 0: return None kwargs = kwargs if kwargs is not None else {} from collections.abc import Mapping, Sequence def tree_zip(combined_args, dynamic_shapes): if isinstance(combined_args, (tuple, list)): if not isinstance(dynamic_shapes, Sequence): raise UserError( UserErrorType.INVALID_INPUT, f"Expected dynamic_shapes of a {type(combined_args)} to be a Sequence, " f"got {dynamic_shapes} instead", ) if len(combined_args) != len(dynamic_shapes): raise UserError( UserErrorType.INVALID_INPUT, f"Expected {dynamic_shapes} to have {len(combined_args)} items", ) for i, shape in enumerate(dynamic_shapes): yield from tree_zip(combined_args[i], shape) elif isinstance(combined_args, dict): if not isinstance(dynamic_shapes, Mapping): raise UserError( UserErrorType.INVALID_INPUT, f"Expected dynamic_shapes of a {type(combined_args)} to be a Mapping, " f"got {dynamic_shapes} instead", ) if len(combined_args) != len(dynamic_shapes): raise UserError( UserErrorType.INVALID_INPUT, f"Expected {dynamic_shapes} to have {len(combined_args)} items", ) for k, shape in dynamic_shapes.items(): yield from tree_zip(combined_args[k], shape) elif dataclasses.is_dataclass(combined_args): if not type(dynamic_shapes) == type(combined_args): raise UserError( UserErrorType.INVALID_INPUT, f"Expected dynamic_shapes of a {type(combined_args)} to be a {type(combined_args)}, " f"got {dynamic_shapes} instead", ) for f in dataclasses.fields(combined_args): yield from tree_zip(getattr(combined_args, f.name), getattr(dynamic_shapes, f.name)) elif isinstance(combined_args, torch.Tensor): yield (combined_args, dynamic_shapes) else: if dynamic_shapes is not None: raise UserError( UserErrorType.INVALID_INPUT, f"Expected dynamic_shapes of a {type(combined_args)} to be None, " f"got {dynamic_shapes} instead", ) def to_constraint(dim, tensor, i): constraint = dynamic_dim(tensor, i, debug_name=dim.__name__) if dim.min != 2: constraint = constraint >= dim.min if dim.max != sys.maxsize - 1: constraint = constraint <= dim.max return constraint from collections import defaultdict symbols = defaultdict(list) bounds: Dict[str, Tuple[int, int]] = {} def check_same_bounds(dim): if dim.__name__ in symbols: min_, max_ = bounds[dim.__name__] if dim.min != min_ or dim.max != max_: this_ = _Dim.readable(dim.__name__, min_, max_) that_ = _Dim.readable(dim.__name__, dim.min, dim.max) raise UserError( UserErrorType.INVALID_INPUT, f"Found different definitions {this_} and {that_} " f"for the same symbolic dimension {dim}!" ) else: bounds[dim.__name__] = (dim.min, dim.max) def update_symbols(tensor, shape): if isinstance(shape, dict): for i, dim in shape.items(): if isinstance(dim, _Dim): check_same_bounds(dim) symbols[dim.__name__].append(to_constraint(dim, tensor, i)) else: if dim is not None: raise UserError( UserErrorType.INVALID_INPUT, f"Unexpected item #{i} ({dim}) in dynamic_shape {shape} of Tensor, " "try None instead", ) elif isinstance(shape, (tuple, list)): for i, dim in enumerate(shape): if isinstance(dim, _Dim): check_same_bounds(dim) symbols[dim.__name__].append(to_constraint(dim, tensor, i)) else: if dim is not None: raise UserError( UserErrorType.INVALID_INPUT, f"Unexpected item #{i} ({dim}) in dynamic_shape {shape} of Tensor, " "try None instead", ) else: if shape is not None: raise UserError( UserErrorType.INVALID_INPUT, f"Unexpected dynamic_shape {shape} of Tensor, " "try None instead", ) import inspect if isinstance(f, ExportedProgram): f = f.module() signature = inspect.signature(f.forward) if isinstance(f, torch.nn.Module) else inspect.signature(f) combined_args = signature.bind(*args, **kwargs).arguments # This means user didn't specify dynamic shapes with argument names. combined_args = combined_args if isinstance(dynamic_shapes, Mapping) else list(combined_args.values()) # type: ignore[assignment] for tensor, shape in tree_zip(combined_args, dynamic_shapes): update_symbols(tensor, shape) constraints = [] for dynamic_dims in symbols.values(): primary, *others = dynamic_dims if others: for other in others: constraints.append(primary == other) else: constraints.append(primary) return constraints def export__RC__( f: Callable, args: Tuple[Any, ...], kwargs: Optional[Dict[str, Any]] = None, *, dynamic_shapes: Optional[Union[Dict[str, Any], Tuple[Any]]] = None, strict: bool = True, preserve_module_call_signature: Tuple[str, ...] = (), ) -> ExportedProgram: """ API for exporting with dynamic shape specifications instead of constraints. It should be considered "release candidate" (RC), meant to replace `export`. Here, `dynamic_shapes` is expected to be a dict from argument names of `f` to dynamic shape specifications OR a tuple where each element corresponds to the original order of the arguments defined in the function signature ,as follows: - The dynamic shape of a tensor argument can be specified as: - Either a dict from dynamic dimension indices to Dim types. It is not required to include static dimension indices in this dict, but when they are, they should be mapped to None. - Or a tuple of Dim types or None. The Dim types correspond to dynamic dimensions, whereas static dimensions are denoted by None. - Arguments that are dicts or tuples of tensors are recursively specified by using mappings or sequences of contained specifications. See `export` for documentation of `f`, `args`, `kwargs` and return. """ constraints = _process_dynamic_shapes(f, args, kwargs, dynamic_shapes) return _export( f, args, kwargs, constraints=constraints, strict=strict, preserve_module_call_signature=preserve_module_call_signature ) def dynamic_dim(t: torch.Tensor, index: int, debug_name: Optional[str] = None): if not isinstance(t, torch.Tensor): raise UserError( UserErrorType.DYNAMIC_DIM, f"Expected tensor as input to dynamic_dim but got {type(t)}" ) if t.dim() < 1: raise UserError( UserErrorType.DYNAMIC_DIM, "Cannot mark 0-dimension tensors to be dynamic" ) if index >= t.dim(): raise UserError( UserErrorType.DYNAMIC_DIM, f"Expected the dimension passed to dynamic_dim to be in the range [0:{t.dim()-1}]" f" but got {index}, which is out of bounds for the given tensor." ) return _create_constraint( weakref.ref(t), id(t), index, StrictMinMaxConstraint( vr=ValueRanges(lower=2, upper=sympy.oo), warn_only=False ), debug_name=debug_name, ) @dataclasses.dataclass class ExportDynamoConfig: """ Manage Export-specific configurations of Dynamo. """ allow_rnn: bool = True DEFAULT_EXPORT_DYNAMO_CONFIG = ExportDynamoConfig() DECOMP_TABLE = core_aten_decompositions() # TODO(zhxchen17) This is not needed if we output pre_dispatch graph upfront from export(). @contextmanager def _disable_decomp_table(): global DECOMP_TABLE prev, DECOMP_TABLE = DECOMP_TABLE, {} try: yield finally: DECOMP_TABLE = prev @compatibility(is_backward_compatible=False) def capture_pre_autograd_graph( f: Callable, args: Tuple[Any], kwargs: Optional[Dict[str, Any]] = None, constraints: Optional[List[Constraint]] = None, ) -> torch.nn.Module: """ A helper function that is intended to trace a module before any pre-autograd decomposition is run. The produced module will be "non-functional" and composed of aten operators. Later this API will be deleted in favor of more general torch.export API. Args: f: A callable to be traced args: example positional inputs. kwargs: optional example keyword inputs. constraints: A optional list of constraints on the dynamic arguments specifying their possible range of their shapes Returns: An nn.Module containing the traced method. """ decomp_table = { torch.ops.aten.dropout.default: torch.ops.aten.dropout.default.decompose, torch.ops.aten.batch_norm.default: torch.ops.aten.batch_norm.default.decompose, torch.ops.aten._batch_norm_impl_index.default: torch.ops.aten._batch_norm_impl_index.default.decompose, torch.ops.aten.native_batch_norm.default: torch.ops.aten.native_batch_norm.default.decompose, } if kwargs is None: kwargs = {} with torch._dynamo.config.patch(dataclasses.asdict(DEFAULT_EXPORT_DYNAMO_CONFIG)): m = torch._dynamo.export( f, constraints=constraints, assume_static_by_default=True, tracing_mode="symbolic", decomposition_table=decomp_table, pre_dispatch=True, aten_graph=True, )( *args, **kwargs, )[0] def _train(self, mode: bool = True): raise NotImplementedError("Calling train() is not supported yet.") def _eval(self, mode: bool = True): raise NotImplementedError("Calling eval() is not supported yet.") _, _, _, fake_mode = _convert_input_to_fake(m, args, kwargs) m.meta["inline_constraints"] = { k: v for k, v in fake_mode.shape_env.runtime_var_to_range.items() if re.match(r"^[if]\d+$", str(k)) } flat_args, _ = pytree.tree_flatten((args, kwargs or {})) range_constraints, equality_constraints = _process_constraints(m, 0, flat_args) unlifted_m = _create_stateful_graph_module( m, range_constraints=range_constraints, equality_constraints=equality_constraints, ) unlifted_m.train = types.MethodType(_train, m) # type: ignore[method-assign] unlifted_m.eval = types.MethodType(_eval, m) # type: ignore[method-assign] return unlifted_m def _convert_input_to_fake(gm, args, kwargs): if len(args) == 0 and len(kwargs) == 0 and len(dict(gm.named_parameters())) == 0 and len(dict(gm.named_buffers())) == 0: return [], {}, {}, None fake_inps: List[torch.Tensor] = [] fake_mode = None for node in gm.graph.nodes: if node.op == "placeholder" and "val" in node.meta: fake_val = node.meta["val"] if fake_val is not None and isinstance(fake_val, torch.Tensor): fake_inps.append(fake_val) if detected_fake_mode := detect_fake_mode(fake_inps): fake_mode = detected_fake_mode assert fake_mode is not None, "Cannot find fake_mode attatched to the graph's placeholders." count = 0 def convert_to_fake(x): nonlocal count val = fake_inps[count] count += 1 return val fake_args = pytree.tree_map_only(torch.Tensor, convert_to_fake, args) # TODO properly use the cached fake tensor fake_kwargs = pytree.tree_map_only(torch.Tensor, fake_mode.from_tensor, kwargs) fake_params_buffers = pytree.tree_map_only(torch.Tensor, functools.partial(fake_mode.from_tensor, static_shapes=True), {**dict(gm.named_parameters(remove_duplicate=False)), **dict(gm.named_buffers(remove_duplicate=False))}) return fake_args, fake_kwargs, fake_params_buffers, fake_mode def _replace_param_buffer_names(param_buffer_table, sig): for spec in sig.input_specs: spec.target = param_buffer_table.get(spec.target, spec.target) for spec in sig.output_specs: spec.target = param_buffer_table.get(spec.target, spec.target) def _normalize_nn_module_stack(gm_torch_level, root_cls): # Append a root module to every nn_module_stack. root = "L['self']" root_key = re.sub(r'[^a-zA-Z0-9]', '_', root) for gm in gm_torch_level.modules(): if not isinstance(gm, torch.fx.GraphModule): continue for node in gm.graph.nodes: if node.op in ["placeholder", "output"]: continue add_root = True if nn_module_stack := node.meta.get("nn_module_stack", {}): path, ty = next(iter(nn_module_stack.values())) assert issubclass(ty, torch.nn.Module) # TODO Figure out why sometimes we have root sometimes we don't. if path == root and ty is root_cls: add_root = False if add_root: def normalize_path(path): try: parts = [] class Path: def __getattr__(self, name): parts.append(name) return self def __getitem__(self, idx): parts.append(str(idx)) return self eval(path, {"L": {"self": Path()}}) return ".".join(parts) except Exception: # TODO(zhxchen17) Remove this. return path nn_module_stack = {root_key: (root, root_cls), **nn_module_stack} node.meta["nn_module_stack"] = { key: (normalize_path(path), ty) for key, (path, ty) in nn_module_stack.items() } def _export_to_torch_ir( f: Callable, args: Tuple[Any, ...], kwargs: Optional[Dict[str, Any]] = None, constraints: Optional[List[Constraint]] = None, *, preserve_module_call_signature: Tuple[str, ...] = (), disable_constraint_solver: bool = False, ) -> torch.fx.GraphModule: """ Traces either an nn.Module's forward function or just a callable with PyTorch operations inside and produce a torch.fx.GraphModule in torch IR. """ constraints = constraints or [] kwargs = kwargs or {} if not isinstance(args, tuple): raise UserError(UserErrorType.INVALID_INPUT, f"Expecting `args` to be a tuple of example positional inputs, got {type(args)}") # We convert to nn.Module because __call__ of ExportedProgram # is untracable right now. if isinstance(f, ExportedProgram): f = f.module() with torch._dynamo.config.patch(dataclasses.asdict(DEFAULT_EXPORT_DYNAMO_CONFIG)): try: module_call_specs: Dict[str, Dict[str, pytree.TreeSpec]] = {} with _wrap_submodules(f, preserve_module_call_signature, module_call_specs): gm_torch_level, _ = torch._dynamo.export( f, constraints=constraints, assume_static_by_default=True, tracing_mode="symbolic", disable_constraint_solver=disable_constraint_solver, )( *args, **kwargs, ) except (ConstraintViolationError, ValueRangeError) as e: raise UserError(UserErrorType.CONSTRAINT_VIOLATION, str(e)) # noqa: TRY200 except GuardOnDataDependentSymNode as e: raise UserError( # noqa: TRY200 UserErrorType.ANTI_PATTERN, f"Consider annotating your code using torch._constrain_as_*(). {str(e)}", case_name="constrain_as_size_example", ) gm_torch_level.meta["module_call_specs"] = module_call_specs return gm_torch_level def export( f: Callable, args: Tuple[Any, ...], kwargs: Optional[Dict[str, Any]] = None, constraints: Optional[List[Constraint]] = None, *, strict: bool = True, preserve_module_call_signature: Tuple[str, ...] = (), ) -> ExportedProgram: if constraints is not None: warnings.warn( "Using `constraints` to specify dynamic shapes for export is DEPRECATED " "and will not be supported in the future. " "Please use `dynamic_shapes` instead (see docs on `torch.export.export`).", DeprecationWarning, stacklevel=2, ) return _export( f, args, kwargs, constraints, strict=strict, preserve_module_call_signature=preserve_module_call_signature, ) def _unlift_user_inputs_to_buffers( gm_torch_level: torch.fx.GraphModule, aot_export_args ) -> List[str]: flat_args = pytree.tree_leaves(aot_export_args) user_input_names = [] with gm_torch_level.graph.inserting_before(): for i, (arg, node) in enumerate(zip(flat_args, gm_torch_level.graph.nodes)): assert node.op == "placeholder" user_input_names.append(node.name) if isinstance(arg, torch.Tensor): assert not hasattr(gm_torch_level, node.name) gm_torch_level.register_buffer(node.name, arg) get_attr = gm_torch_level.graph.get_attr(node.name) node.replace_all_uses_with(get_attr) get_attr.meta = copy.copy(node.meta) for node in list(gm_torch_level.graph.nodes): if node.op == "placeholder": assert len(node.users) == 0 gm_torch_level.graph.erase_node(node) gm_torch_level.recompile() return user_input_names def _lift_buffers_to_user_inputs( gm: torch.fx.GraphModule, graph_signature: GraphSignature, user_input_names: List[str] ) -> Dict[str, str]: assert len(graph_signature.user_inputs) == 0 assert graph_signature.backward_signature is None names = set(user_input_names) placeholders = [node for node in gm.graph.nodes if node.op == "placeholder"] # user inputs are always added in the end start = len(graph_signature.parameters) end = start + len(graph_signature.buffers) buffer_nodes = placeholders[start:end] last_placeholder_node = placeholders[-1] if len(placeholders) > 0 else None old_nodes: Dict[str, torch.fx.Node] = {} for node in buffer_nodes: buffer_name = graph_signature.inputs_to_buffers[node.name] if buffer_name not in names: continue old_nodes[buffer_name] = node replaces = {} new_node_names: Dict[str, str] = {} with gm.graph.inserting_after(last_placeholder_node): for name in reversed(user_input_names): new_node = gm.graph.placeholder(name) new_node.target = new_node.name new_node_names[name] = new_node.name if name in old_nodes: old_node = old_nodes[name] new_node.meta = copy.copy(old_node.meta) old_node.replace_all_uses_with(new_node) replaces[old_node.name] = new_node.name new_node_names = dict(reversed(new_node_names.items())) for old_node in old_nodes.values(): gm.graph.erase_node(old_node) gm.recompile() graph_signature.buffers = [b for b in graph_signature.buffers if b not in names] graph_signature.inputs_to_buffers = { i: b for i, b in graph_signature.inputs_to_buffers.items() if b not in names } user_inputs_to_mutate = { o: b for o, b in graph_signature.buffers_to_mutate.items() if b in names } graph_signature.buffers_to_mutate = { o: b for o, b in graph_signature.buffers_to_mutate.items() if b not in names } graph_signature.user_inputs.extend(new_node_names.values()) # type: ignore[arg-type] graph_signature.user_outputs = [ replaces[o] if o in replaces else o for o in graph_signature.user_outputs ] return user_inputs_to_mutate # type: ignore[return-value] def _export_non_strict( mod, fake_args, fake_kwargs, fake_params_buffers, *, transform=lambda x: x # TODO(zhxchen17) Revisit if this is needed later. ): # This _reparametrize_module makes sure inputs and module.params/buffers have the same fake_mode, # otherwise aot_export_module will error out because it sees a mix of fake_modes. # And we want aot_export_module to use the fake_tensor mode in dynamo to keep the pipeline easy to reason about. with torch.nn.utils.stateless._reparametrize_module(mod, fake_params_buffers): gm, graph_signature = transform(aot_export_module)( mod, (*fake_args, *fake_kwargs.values()), trace_joint=False ) # NOTE: aot_export adds symint metadata for placeholders with int values; # since these become specialized, we replace such metadata with the original values flat_args = pytree.tree_leaves((fake_args, fake_kwargs)) index = 0 total_param_buffers = len(graph_signature.parameters) + len(graph_signature.buffers) for node in gm.graph.nodes: if node.op == "placeholder": if index >= total_param_buffers: user_arg = flat_args[index - total_param_buffers] if not isinstance(user_arg, torch.Tensor): node.meta["val"] = user_arg index += 1 is_joint = graph_signature.backward_signature is not None def make_argument_spec(node) -> ArgumentSpec: assert "val" in node.meta, f"{node} has no 'val' metadata field" val = node.meta["val"] if isinstance(val, FakeTensor): return TensorArgument(name=node.name) elif isinstance(val, torch.SymInt): return SymIntArgument(name=node.name) else: return ConstantArgument(value=val) input_specs, output_specs = _sig_to_specs( user_inputs=set(graph_signature.user_inputs), inputs_to_parameters=graph_signature.inputs_to_parameters, # type: ignore[arg-type] inputs_to_buffers=graph_signature.inputs_to_buffers, # type: ignore[arg-type] user_outputs=set(graph_signature.user_outputs), # type: ignore[arg-type] buffer_mutations=graph_signature.buffers_to_mutate, # type: ignore[arg-type] user_input_mutations=gm.meta.get("user_inputs_to_mutate", {}), # type: ignore[arg-type] grad_params=graph_signature.backward_signature.gradients_to_parameters if is_joint else {}, # type: ignore[arg-type, union-attr] grad_user_inputs=graph_signature.backward_signature.gradients_to_user_inputs if is_joint else {}, # type: ignore[arg-type, union-attr] loss_output=graph_signature.backward_signature.loss_output if is_joint else None, # type: ignore[arg-type, union-attr] inputs=[make_argument_spec(node) for node in gm.graph.nodes if node.op == "placeholder"], outputs=[make_argument_spec(node) for node in pytree.tree_leaves(next(iter(reversed(gm.graph.nodes))).args)], ) export_graph_signature = ExportGraphSignature(input_specs=input_specs, output_specs=output_specs) tensor_constants = lift_constant_tensor_pass(gm, export_graph_signature) @dataclasses.dataclass class _ExportedProgramNonStrict: gm: torch.fx.GraphModule sig: ExportGraphSignature tensor_constants: Dict[str, torch.Tensor] return _ExportedProgramNonStrict( gm, export_graph_signature, tensor_constants, ) def _get_params_buffers(mod: torch.nn.Module) -> Dict[str, torch.Tensor]: params_buffers: Dict[str, torch.Tensor] = {} for name, param in mod.named_parameters(remove_duplicate=False): params_buffers[name] = param for name, buffer in mod.named_buffers(remove_duplicate=False): params_buffers[name] = buffer return params_buffers @_disable_prexisiting_fake_mode def _export( f: Callable, args: Tuple[Any, ...], kwargs: Optional[Dict[str, Any]] = None, constraints: Optional[List[Constraint]] = None, *, strict: bool = True, preserve_module_call_signature: Tuple[str, ...] = (), ) -> ExportedProgram: """ Traces either an nn.Module's forward function or just a callable with PyTorch operations inside and produce a ExportedProgram. Args: m: the `nn.Module` or callable to trace. args: example positional inputs. kwargs: optional example keyword inputs. constraints: A optional list of constraints on the dynamic arguments specifying their possible range of their shapes preserve_module_call_signature: A list of submodule paths for which the original calling conventions are preserved as metadata. Returns: An ExportedProgram containing the traced method. """ constraints = constraints or [] kwargs = kwargs or {} if not strict: assert isinstance(f, torch.nn.Module) assert len(preserve_module_call_signature) == 0 assert len(constraints) == 0, "dynamic shape NYI" assert len(kwargs) == 0, "keyword arguments NYI" out_spec = None def _tuplify_outputs(aot_export): def _aot_export_non_strict(mod, args, **kwargs): class Wrapper(torch.nn.Module): def __init__(self, mod): super().__init__() self._export_root = mod def forward(self, *args, **kwargs): nonlocal out_spec flat_outs, out_spec = pytree.tree_flatten(self._export_root(*args, **kwargs)) return tuple(flat_outs) gm, sig = aot_export(Wrapper(mod), args, **kwargs) def strip_root(x): return x[len('_export_root.'):] if x.startswith('_export_root.') else x sig.parameters = pytree.tree_map(strip_root, sig.parameters) sig.buffers = pytree.tree_map(strip_root, sig.buffers) sig.inputs_to_buffers = pytree.tree_map(strip_root, sig.inputs_to_buffers) sig.inputs_to_parameters = pytree.tree_map(strip_root, sig.inputs_to_parameters) sig.buffers_to_mutate = pytree.tree_map(strip_root, sig.buffers_to_mutate) return gm, sig return _aot_export_non_strict ep_non_strict = _export_non_strict(f, args, {}, f.state_dict(), transform=_tuplify_outputs) assert out_spec is not None return ExportedProgram( ep_non_strict.gm, ep_non_strict.gm.graph, ep_non_strict.sig, _get_params_buffers(f), {}, [], [ModuleCallEntry("", ModuleCallSignature([], [], pytree.tree_flatten((args, {}))[1], out_spec))], (args, kwargs), tensor_constants=ep_non_strict.tensor_constants, ) gm_torch_level = _export_to_torch_ir( f, args, kwargs, constraints, preserve_module_call_signature=preserve_module_call_signature, ) params_buffers = _get_params_buffers(gm_torch_level) # We detect the fake_mode by looking at gm_torch_level's placeholders, this is the fake_mode created in dynamo. fake_args, fake_kwargs, fake_params_buffers, dynamo_fake_mode = _convert_input_to_fake(gm_torch_level, args, kwargs) # First, we want to pass through the graph to try populating # val field for getattr if there is anything missing. # THis can happen when quantization adds extra params and forgets # to update "val" for node in gm_torch_level.graph.nodes: if node.op == "get_attr" and "val" not in node.meta: attr = getattr(gm_torch_level, node.target) # Checks if it is not a HigherOrderOp branch or a module if not isinstance(attr, torch.nn.Module): assert dynamo_fake_mode is not None, ( "Cannot find dynamo_fake_mode. This could be due to the exported graph module have no placeholders." ) node.meta["val"] = dynamo_fake_mode.from_tensor(attr, static_shapes=True) # When aot_export lifts the params, we lose the nn_module_stack # and source_fn from the param nodes as they are treated as fresh inputs # Therefore, we manually extract them before calling into aot_export params_buffers_to_node_meta = {} for node in gm_torch_level.graph.nodes: target = node.target meta = node.meta if node.op == "call_module": submodule = getattr(gm_torch_level, target) if isinstance(submodule, torch.nn.Module): for name, _ in submodule.named_parameters(recurse=True, remove_duplicate=False): params_buffers_to_node_meta[target + "." + name] = meta for name, _ in submodule.named_buffers(recurse=True, remove_duplicate=False): params_buffers_to_node_meta[target + "." + name] = meta if node.op == "get_attr": submodule = getattr(gm_torch_level, target) if not isinstance(submodule, torch.fx.GraphModule): params_buffers_to_node_meta[target] = meta # If the call_function uses param as input, we also need to update params' meta # with this call_function node's meta. # This is basically the same flow as torch.fx.traceback.preserve_meta() if node.op == "call_function" and not isinstance(node.target, torch._ops.HigherOrderOperator): for arg in node._input_nodes: if arg.op == "get_attr": for entry in torch.fx.proxy._COPY_META_FIELDS: if entry in meta: params_buffers_to_node_meta[arg.target][entry] = meta[entry] # Fix the graph output signature to be tuple if scalar out_spec = orig_out_spec = gm_torch_level._out_spec assert out_spec is not None # aot_export expect the return type to always be a tuple. if out_spec.type not in (list, tuple): out_spec = pytree.TreeSpec(tuple, None, [out_spec]) orig_args = gm_torch_level.graph._codegen.pytree_info.orig_args # type: ignore[attr-defined] gm_torch_level.graph._codegen = _PyTreeCodeGen( _PyTreeInfo( orig_args, gm_torch_level._in_spec, out_spec, ) ) gm_torch_level.recompile() param_buffer_table: Dict[str, str] = {} if isinstance(f, torch.nn.Module): param_lookup: Dict[int, List[str]] = {} buffer_lookup: Dict[int, List[str]] = {} for name, param in f.named_parameters(remove_duplicate=False): param_lookup.setdefault(id(param), []).append(name) for name, buffer in f.named_buffers(remove_duplicate=False): buffer_lookup.setdefault(id(buffer), []).append(name) for dynamo_name, dynamo_param in gm_torch_level.named_parameters(remove_duplicate=False): assert dynamo_name not in param_buffer_table if id(dynamo_param) in param_lookup: param_buffer_table[dynamo_name] = param_lookup[id(dynamo_param)].pop() for dynamo_name, dynamo_buffer in gm_torch_level.named_buffers(remove_duplicate=False): assert dynamo_name not in param_buffer_table if id(dynamo_buffer) in buffer_lookup: param_buffer_table[dynamo_name] = buffer_lookup[id(dynamo_buffer)].pop() if isinstance(f, torch.nn.Module): _normalize_nn_module_stack(gm_torch_level, type(f)) def _process_user_inputs(aot_export): def _aot_export_strict(gm_torch_level: torch.fx.GraphModule, args, **kwargs): user_input_names = _unlift_user_inputs_to_buffers(gm_torch_level, args) gm, graph_signature = aot_export(gm_torch_level, (), **kwargs) user_inputs_to_mutate = _lift_buffers_to_user_inputs(gm, graph_signature, user_input_names) # TODO unfortunately preserving graph-level metadata is not # working well with aot_export. So we manually copy it. # (The node-level meta is addressed above.) gm.meta.update(gm_torch_level.meta) assert "user_inputs_to_mutate" not in gm.meta gm.meta["user_inputs_to_mutate"] = user_inputs_to_mutate return gm, graph_signature return _aot_export_strict # Note: aot_export_module doesn't accept kwargs, we'd like to reorder the kwargs as an OrderedDict # to follow the order in orig_args and correctly call module ep_non_strict = _export_non_strict( gm_torch_level, fake_args, _reorder_kwargs_by_names(orig_args, fake_args, fake_kwargs), fake_params_buffers, transform=_process_user_inputs ) gm = ep_non_strict.gm export_graph_signature = ep_non_strict.sig tensor_constants = ep_non_strict.tensor_constants # After aot_export, set the param/buffer metadata back into placeholders # Technically, users can still construct this data from param names # without relying on this metadata for node in gm.graph.nodes: if node.op == "placeholder": if node.target in export_graph_signature.inputs_to_parameters: param_name = export_graph_signature.inputs_to_parameters[node.target] if param_name in params_buffers_to_node_meta: for k, v in params_buffers_to_node_meta[param_name].items(): node.meta[k] = v if node.target in export_graph_signature.inputs_to_buffers: buffer_name = export_graph_signature.inputs_to_buffers[node.target] if buffer_name in params_buffers_to_node_meta: for k, v in params_buffers_to_node_meta[buffer_name].items(): node.meta[k] = v # The unbacked symint symbols are updated in aot_export # so we serialize them here instead of inside dynamo # dynamo_fake_mode can be None if there's no placeholder in gm_torch_level if dynamo_fake_mode: gm.meta["inline_constraints"] = { k: v for k, v in dynamo_fake_mode.shape_env.runtime_var_to_range.items() if re.match(r"^[if]\d+$", str(k)) } num_lifted = next( (i for i, s in enumerate(export_graph_signature.input_specs) if s.kind == InputKind.USER_INPUT), 0 ) flat_args, orig_in_spec = pytree.tree_flatten((args, kwargs)) range_constraints, equality_constraints = _process_constraints( gm, num_lifted, flat_args, ) if isinstance(f, torch.nn.Module): _replace_param_buffer_names(param_buffer_table, export_graph_signature) params_buffers = {param_buffer_table.get(name, name): tensor for name, tensor in params_buffers.items()} module_call_signatures = { fqn: ModuleCallSignature(inputs=[], outputs=[], **specs) for fqn, specs in gm_torch_level.meta["module_call_specs"].items() } if len(preserve_module_call_signature) > 0: res = CollectTracepointsPass(module_call_signatures, export_graph_signature)(gm) assert res is not None gm = res.graph_module assert orig_out_spec is not None exported_program = ExportedProgram( gm, gm.graph, export_graph_signature, # TODO(zhxchen17) Return empty state_dict for functions. params_buffers, range_constraints, equality_constraints, [ModuleCallEntry("", ModuleCallSignature(inputs=[], outputs=[], in_spec=orig_in_spec, out_spec=orig_out_spec))] + [ModuleCallEntry(fqn, sig) for fqn, sig in module_call_signatures.items()], (args, kwargs), tensor_constants=tensor_constants, ) if len(range_constraints) > 0 or len(equality_constraints) > 0: exported_program = exported_program._transform( _AddRuntimeAssertionsForInlineConstraintsPass(range_constraints, equality_constraints) ) return exported_program def _reorder_kwargs_by_names(arg_names: List[str], args: Tuple[Any], kwargs: Dict[str, Any]): assert len(arg_names) == len(args) + len(kwargs), ( f"Total number of arg names is expected to be {len(arg_names)} " f"but got {len(args)} positional args, {len(kwargs)} kwargs." ) return {kw_name: kwargs[kw_name] for kw_name in arg_names[len(args):]} def save( ep: ExportedProgram, f: Union[str, pathlib.Path, io.BytesIO], *, extra_files: Optional[Dict[str, Any]] = None, opset_version: Optional[Dict[str, int]] = None, ) -> None: from .serde.serialize import serialize, SerializedArtifact from .serde.schema import SCHEMA_VERSION artifact: SerializedArtifact = serialize(ep, opset_version) if isinstance(f, (str, pathlib.Path)): f = str(f) with zipfile.ZipFile(f, 'w') as zipf: # Save every field the SerializedArtifact to a file for field in dataclasses.fields(artifact): field_name = field.name serialized_field = getattr(artifact, field_name) zipf.writestr(f"serialized_{field_name}.json", serialized_field) zipf.writestr('version', str(SCHEMA_VERSION)) # Add extra files if provided if extra_files: for extra_file_name, content in extra_files.items(): encoded_content = content.encode('utf-8') zipf.writestr(f"extra_files/{extra_file_name}", encoded_content) def load( f: Union[str, pathlib.Path, io.BytesIO], *, extra_files: Optional[Dict[str, Any]] = None, expected_opset_version: Optional[Dict[str, int]] = None, ) -> ExportedProgram: if isinstance(f, (str, pathlib.Path)): f = str(f) with zipfile.ZipFile(f, 'r') as zipf: # Check the version version = int(zipf.read('version')) from .serde.schema import SCHEMA_VERSION if version != SCHEMA_VERSION: raise RuntimeError( f"Serialized version {version} does not match our current " f"schema version {SCHEMA_VERSION}." ) from .serde.serialize import deserialize, SerializedArtifact # Load serialized_ep and serialized_state_dict from the zip file artifact: SerializedArtifact = SerializedArtifact( **{ field.name: zipf.read(f"serialized_{field.name}.json") for field in dataclasses.fields(SerializedArtifact) } ) # Deserialize ExportedProgram ep = deserialize(artifact) # Populate extra_files map if extra_files is not None: for filename in extra_files.keys(): extra_files[filename] = zipf.read(f"extra_files/{filename}").decode('utf-8') return ep def aot_compile( f: Callable, args: Tuple[Any], kwargs: Optional[Dict[str, Any]] = None, *, constraints: Optional[List[Constraint]] = None, dynamic_shapes: Optional[Dict[str, Any]] = None, options: Optional[Dict[str, Any]] = None, remove_runtime_assertions: bool = False, disable_constraint_solver: bool = False, ) -> str: """ Note: this function is not stable yet Traces either an nn.Module's forward function or just a callable with PyTorch operations inside, generates executable cpp code from the program, and returns the path to the generated shared library Args: f: the `nn.Module` or callable to trace. args: example positional inputs. kwargs: optional example keyword inputs. constraints: A optional list of constraints on the dynamic arguments specifying their possible range of their shapes dynamic_shapes: An experimental new feature designed to subsume ``constraints``. A dict mapping argument names of ``f`` to their dynamic shape specifications, as follows. Dynamic shape specifications can be a dict from dynamic dimensions to ``Dim`` types, or a tuple/list of ``Optional[Dim]`` corresponding to each input dimension. options: A dictionary of options to control inductor disable_constraint_solver: Whether the dim constraint solver must be disabled. Returns: Path to the generated shared library """ if constraints is not None: warnings.warn( "The constraints field is deprecated. " "Please use dynamic_shapes instead." ) from torch._inductor.decomposition import select_decomp_table if constraints is None: constraints = _process_dynamic_shapes(f, args, kwargs, dynamic_shapes) # We want to export to Torch IR here to utilize the pre_grad passes in # inductor, which run on Torch IR. gm = _export_to_torch_ir( f, args, kwargs, constraints, disable_constraint_solver=disable_constraint_solver ) flat_example_inputs = pytree.arg_tree_leaves(*args, **(kwargs or {})) with torch.no_grad(): so_path = torch._inductor.aot_compile(gm, flat_example_inputs, options) # type: ignore[arg-type] return so_path