# mypy: allow-untyped-defs import copy import warnings from collections.abc import Sequence from itertools import chain from typing import Any, Optional import torch import torch.utils._pytree as pytree from torch._export.non_strict_utils import ( _enter_enable_graph_inputs_of_type_nn_module, _exit_enable_graph_inputs_of_type_nn_module, _get_graph_inputs_of_type_nn_module, ) from torch._export.utils import _check_input_constraints_for_graph from torch.export.unflatten import _assign_attr, _AttrKind from torch.fx.graph import _PyTreeCodeGen, _PyTreeInfo from ._remove_effect_tokens_pass import _remove_effect_tokens from ._tree_utils import reorder_kwargs from .exported_program import ( ExportedProgram, ExportGraphSignature, InputKind, OutputKind, ) def _check_inputs_match(args, kwargs, in_spec: pytree.TreeSpec) -> list: reordered_kwargs = reorder_kwargs(kwargs, in_spec) flat_args_with_path, received_spec = pytree.tree_flatten_with_path( (args, reordered_kwargs) ) if received_spec != in_spec: raise ValueError( # noqa: B904 "Trying to flatten user inputs with exported input tree spec: \n" f"{in_spec}\n" "but actually got inputs with tree spec of: \n" f"{received_spec}.\n" "Please check that the inputs have the same number of args " "and kwargs as the ones you used when tracing." ) return flat_args_with_path @torch._dynamo.disable def _check_input_constraints_pre_hook(self, args, kwargs): if not self.validate_inputs: return flat_args_with_path = _check_inputs_match(args, kwargs, self._in_spec) _check_input_constraints_for_graph( [node for node in self.graph.nodes if node.op == "placeholder"], flat_args_with_path, self.range_constraints, ) def _unlift_inputs_as_getattr( gm: torch.fx.GraphModule, lifted_inputs: Sequence[Optional[str]], ) -> tuple[dict[str, torch.fx.Node], dict[str, torch.fx.Node]]: """ Unlift inputs referring to params/buffers/constants as getattr nodes in the graph """ unlifted_name_to_node = {} input_name_to_node = {} placeholder_nodes = [node for node in gm.graph.nodes if node.op == "placeholder"] assert len(lifted_inputs) == len(placeholder_nodes) for input_node, lifted_node in zip(placeholder_nodes, lifted_inputs): if lifted_node is None: input_name_to_node[input_node.name] = input_node else: with gm.graph.inserting_after(input_node): getattr_node = gm.graph.get_attr(lifted_node) input_node.replace_all_uses_with(getattr_node) metadata = input_node.meta gm.graph.erase_node(input_node) getattr_node.meta = metadata unlifted_name_to_node[lifted_node] = getattr_node return unlifted_name_to_node, input_name_to_node def _insert_copy_for_mutations( gm: torch.fx.GraphModule, mutated_outputs: Sequence[Optional[str]], unlifted_name_to_node: dict[str, torch.fx.Node], input_name_to_node: dict[str, torch.fx.Node], ) -> None: """ Find the all the buffers and inputs that were mutated and insert copy_ operators to reflect mutations. """ output_node = None for node in gm.graph.nodes: if node.op == "output": output_node = node break assert output_node is not None outputs = pytree.tree_flatten(output_node.args)[0] assert len(outputs) == len(mutated_outputs) user_output_nodes = [] return_nodes_to_copy = {} for return_node, mutated_node_name in zip(outputs, mutated_outputs): if mutated_node_name is None: user_output_nodes.append(return_node) continue if mutated_node_name in unlifted_name_to_node: mutated_node = unlifted_name_to_node[mutated_node_name] elif mutated_node_name in input_name_to_node: mutated_node = input_name_to_node[mutated_node_name] else: raise RuntimeError( f"Could not find {mutated_node_name} in either buffer or input nodes" ) with gm.graph.inserting_before(output_node): copy_node = gm.graph.call_function( torch.ops.aten.copy_.default, (mutated_node, return_node) ) return_nodes_to_copy[return_node] = copy_node output_args = [ return_nodes_to_copy[node] if node in return_nodes_to_copy else node for node in user_output_nodes ] with gm.graph.inserting_before(output_node): # Only return user outputs new_output = gm.graph.output(tuple(output_args)) output_node.replace_all_uses_with(new_output) gm.graph.erase_node(output_node) new_output.name = output_node.name new_output.meta.update(output_node.meta) def _get_codegen( in_spec: pytree.TreeSpec, out_spec: Optional[pytree.TreeSpec], forward_arg_names: Optional[list[str]] = None, ) -> _PyTreeCodeGen: """ Create the codegen for the graph module based on the in/out specs """ if forward_arg_names: names = forward_arg_names else: if ( in_spec.type == tuple and in_spec.num_children == 2 and in_spec.children_specs[0].type == tuple and in_spec.children_specs[1].type == dict ): # if in_spec contains the args (tuple) and kwargs (dict) names = [f"arg_{i}" for i in range(in_spec.children_specs[0].num_children)] # add kwarg names names.extend(in_spec.children_specs[1].context) else: names = [f"arg_{i}" for i in range(in_spec.num_children)] return _PyTreeCodeGen( _PyTreeInfo( names, in_spec, out_spec, ) ) def _unlift( gm: torch.fx.GraphModule, lifted_inputs: Sequence[Optional[str]], mutated_outputs: Sequence[Optional[str]], in_spec: pytree.TreeSpec, out_spec: Optional[pytree.TreeSpec], state_dict: dict[str, Any], constants: dict[str, Any], forward_arg_names: Optional[list[str]] = None, ): """ Args: lifted_inputs: A list matching the graph module's input nodes. For an input node that is referring to a lifted parameter/buffer, this list will contain the fqn the corresponding attribute. Otherwise, this list will contain None. This is used to unlift the lifted parameters as get_attr nodes. mutated_outputs: A list matching the graph module's output nodes. For an output node that is referring to a mutated buffer or user input, this list will contain the name of the corresponding buffer or user input that needs to be mutated. Otherwise, this list will contain None. This is used to re-insert an inplace copy_ operator to copy the mutated values back to the original node. """ unlifted_name_to_node, input_name_to_node = _unlift_inputs_as_getattr( gm, lifted_inputs ) _insert_copy_for_mutations( gm, mutated_outputs, unlifted_name_to_node, input_name_to_node ) gm.graph._codegen = _get_codegen(in_spec, out_spec, forward_arg_names) gm.graph.lint() gm.recompile() return gm def _register_attrs_to_new_gm( new_gm: torch.fx.GraphModule, graph_signature: ExportGraphSignature, state_dict: dict[str, Any], constants: dict[str, Any], ) -> None: non_persistent_buffers = set(graph_signature.non_persistent_buffers) for name in graph_signature.buffers: if name in non_persistent_buffers: persistent = False value = constants[name] else: persistent = True value = state_dict[name] _assign_attr( value, new_gm, name, attr_kind=_AttrKind.BUFFER, persistent=persistent ) for name in graph_signature.parameters: value = state_dict[name] _assign_attr( value, new_gm, name, attr_kind=_AttrKind.PARAMETER, ) # Technically this doesn't account for the aliased multiple constants but # it is ok because we have a separate pass later in the stack that populates # the final gm. for name in chain( graph_signature.lifted_custom_objs, graph_signature.lifted_tensor_constants ): value = constants[name] _assign_attr( value, new_gm, name, attr_kind=_AttrKind.CONSTANT, ) class _StatefulGraphModuleFactory(type): """ Metaclass that ensures a private constructor for _StatefulGraphModule """ def __call__(cls, *args, **kwargs): raise TypeError( f"{cls.__module__}.{cls.__qualname__} has no public constructor. " ) def _create(cls, root, graph, range_constraints=None): return super().__call__( root, graph, range_constraints=range_constraints, ) class _StatefulGraphModule(torch.fx.GraphModule, metaclass=_StatefulGraphModuleFactory): def __init__(self, root, graph, range_constraints=None): super().__init__(root, graph) # Need to fix up non-persistent buffers. self.range_constraints = range_constraints or [] self.validate_inputs = True def _create_stateful_graph_module( plain_graph_module: torch.fx.GraphModule, range_constraints, ep: ExportedProgram, ) -> _StatefulGraphModule: stateful_gm = _StatefulGraphModule._create( plain_graph_module, plain_graph_module.graph, range_constraints=range_constraints, ) module_types = _get_graph_inputs_of_type_nn_module(ep.example_inputs) stateful_gm.register_forward_pre_hook( lambda *args, **kwargs: _enter_enable_graph_inputs_of_type_nn_module( module_types ) ) stateful_gm.register_forward_pre_hook( _check_input_constraints_pre_hook, with_kwargs=True ) stateful_gm.register_forward_hook( lambda *args, **kwargs: _exit_enable_graph_inputs_of_type_nn_module( module_types ), always_call=True, ) # When we have a constant that has requires_grad=True, we need to detach it # when we unlift as the tensors that require gradients should be registered # via parameters. But this is problematic when we have aliasing two constants # because when we call detach, they will become different tensors. This dict # keeps track of this logic. original_tensor_to_detached_tensor = {} # Fix up lifted tensor constants. # fx.GraphModule() constructor silently turns a constant attribute of plain_graph_module # into a buffer in stateful_gm and creates an inconsistency with graph_signature. # We fix this by de-registering these buffers in lifted_tensor_constants # and call _assign_attr(attr_kind=CONSTANT) to register them as constants. for constant_fqn in ep.graph_signature.lifted_tensor_constants: # Sometimes, the constant can require gradient, this is probably a bug in user code, # e.g. `self.const = torch.randn(2, 2, requires_grad=True)`. # We call detach on the constant_val since they're tensor contants and we don't need to # compute their gradients anyway. # Users should properly register it as parameter if they want it to require gradient. buffer = stateful_gm.get_buffer(constant_fqn) if buffer.requires_grad: warnings.warn( f"A model attribute `{constant_fqn}` requires gradient. " f"but it's not properly registered as a parameter. " f"torch.export will detach it and treat it as a constant tensor " f"but please register it as parameter instead." ) detached_buffer = buffer.detach() original_tensor_to_detached_tensor[buffer] = detached_buffer buffer = detached_buffer *prefix, field = constant_fqn.rsplit(".") submod = torch.fx.graph_module._get_attr_via_attr_list(stateful_gm, prefix) delattr(submod, field) _assign_attr(buffer, stateful_gm, constant_fqn, attr_kind=_AttrKind.CONSTANT) # Constants are not preserved well when we create a new GraphModule unlike param/buffers for const_name, value in ep.constants.items(): if not torch.fx.graph_module._has_attr(stateful_gm, const_name): if isinstance(value, torch.Tensor): if value.requires_grad: warnings.warn( f"A model attribute `{const_name}` requires gradient " f"but it's not properly registered as a parameter. " f"torch.export will detach it and treat it as a constant tensor " f"but please register it as parameter instead." ) if value in original_tensor_to_detached_tensor: value = original_tensor_to_detached_tensor[value] else: detached_value = value.detach() original_tensor_to_detached_tensor[value] = detached_value value = detached_value _assign_attr( value, stateful_gm, const_name, attr_kind=_AttrKind.CONSTANT, ) # Fix up non-persistent buffers. torch.fx does not distinguish between # persistent and non-persistent buffers, so we must restore that distinction # here. for buffer in ep.graph_signature.non_persistent_buffers: _assign_attr( plain_graph_module.get_buffer(buffer), stateful_gm, buffer, attr_kind=_AttrKind.BUFFER, persistent=False, ) return stateful_gm def _unlift_exported_program_lifted_states(ep: ExportedProgram) -> torch.nn.Module: # TODO T206340015 if ep.verifiers[0].dialect != "TRAINING": ep = _remove_effect_tokens(ep) new_gm = torch.fx.GraphModule(ep.graph_module, copy.deepcopy(ep.graph)) _register_attrs_to_new_gm(new_gm, ep.graph_signature, ep.state_dict, ep.constants) forward_arg_names = ( sig.forward_arg_names if (sig := ep.module_call_graph[0].signature) else None ) lifted_inputs: list[Optional[str]] = [ ( in_spec.target if in_spec.kind in ( InputKind.BUFFER, InputKind.CONSTANT_TENSOR, InputKind.PARAMETER, InputKind.CUSTOM_OBJ, ) else None ) for in_spec in ep.graph_signature.input_specs ] mutated_outputs: list[Optional[str]] = [ ( out_spec.target if out_spec.kind in (OutputKind.BUFFER_MUTATION, OutputKind.USER_INPUT_MUTATION) else None ) for out_spec in ep.graph_signature.output_specs ] new_gm = _unlift( new_gm, lifted_inputs, mutated_outputs, ep.call_spec.in_spec, ep.call_spec.out_spec, ep.state_dict, ep.constants, forward_arg_names=forward_arg_names, ) unlift_gm = _create_stateful_graph_module(new_gm, ep.range_constraints, ep) unlift_gm.meta.update(ep.graph_module.meta) return unlift_gm