_unlift.py

# Copyright (c) Meta Platforms, Inc. and 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 copy
from typing import Any

import torch
import torch._export
from torch.fx.graph import _PyTreeCodeGen, _PyTreeInfo
from torch.utils import _pytree as pytree


Val = Any


def _unlift(gm, inp_pos_to_param_buffer_name, in_spec, out_spec, state_dict):
    count = 0
    # Step 1: make lifted params as get_attr
    for node in gm.graph.nodes:
        if node.op == "placeholder":
            if count in inp_pos_to_param_buffer_name:
                with gm.graph.inserting_after(node):
                    getattr_node = gm.graph.get_attr(
                        inp_pos_to_param_buffer_name[count]
                    )
                    node.replace_all_uses_with(getattr_node)
                    metadata = node.meta
                    gm.graph.erase_node(node)
                    getattr_node.meta = metadata
            count += 1

    # Step 2: Fix the input/output of the graph now that we deleted
    # some args.
    gm.graph.lint()
    names = [f"arg_{i}" for i in range(len(in_spec.children_specs))]
    gm.graph._codegen = _PyTreeCodeGen(
        _PyTreeInfo(
            names,
            in_spec,
            out_spec,
        )
    )
    gm.recompile()

    # Step 3: Find state references in HigherOrderOps and recursively
    # fix them.
    for node in gm.graph.nodes:
        if node.op == "call_function" and node.target == torch.ops.cond:
            pred, true_graph, false_graph, operands = node.args
            true_gm = getattr(gm, true_graph.name)
            false_gm = getattr(gm, false_graph.name)
            inp_pos_to_param_buffer_name_for_submod = {}
            real_operands = []
            for ix, operand in enumerate(operands):
                if operand.target in inp_pos_to_param_buffer_name.values():
                    inp_pos_to_param_buffer_name_for_submod[ix] = operand.target
                    true_gm.register_buffer(operand.target, state_dict[operand.target])
                    false_gm.register_buffer(operand.target, state_dict[operand.target])
                else:
                    real_operands.append(operand)
            node.args = (pred, true_graph, false_graph, real_operands)

            _, in_spec = pytree.tree_flatten(real_operands)

            _unlift(
                true_gm,
                inp_pos_to_param_buffer_name_for_submod,
                in_spec,
                None,
                state_dict,
            )
            _unlift(
                false_gm,
                inp_pos_to_param_buffer_name_for_submod,
                in_spec,
                None,
                state_dict,
            )
        if node.op == "call_function" and node.target.__name__ == "map_impl":
            body_graph, num_mapped, *operands = node.args
            body_gm = getattr(gm, body_graph.name)
            inp_pos_to_buffer_name_for_submod = {}
            real_operands = []
            for ix, operand in enumerate(operands):
                if operand.target in inp_pos_to_param_buffer_name.values():
                    inp_pos_to_buffer_name_for_submod[ix] = operand.target
                    body_gm.register_buffer(operand.target, state_dict[operand.target])
                else:
                    real_operands.append(operand)
            node.args = (body_graph, num_mapped, *real_operands)

            _, in_spec = pytree.tree_flatten(real_operands)

            _unlift(
                body_gm, inp_pos_to_buffer_name_for_submod, in_spec, None, state_dict
            )
    gm.graph.lint()
    gm.graph.eliminate_dead_code()
    gm.recompile()
    return gm


def unlift_exported_program_lifted_states(
    ep: torch.export.exported_program.ExportedProgram,
):
    new_gm = copy.deepcopy(ep.graph_module)

    # TODO Fix the period in params/buffers names later
    # maybe a pass to replace graph signature with fixed names
    param_buffer_name_to_corrected_name = {}

    for name, stuff in ep.state_dict.items():
        if name in ep.graph_signature.buffers:
            if "." in name:
                new_gm.register_buffer(name.replace(".", "_"), stuff)
                param_buffer_name_to_corrected_name[name] = name.replace(".", "_")
            else:
                new_gm.register_buffer(name, stuff)
        elif name in ep.graph_signature.parameters:
            if "." in name:
                new_gm.register_parameter(name.replace(".", "_"), stuff)
                param_buffer_name_to_corrected_name[name] = name.replace(".", "_")
            else:
                new_gm.register_parameter(name, stuff)
        else:
            raise AssertionError("encountered not registered param/buffer")

    count = 0
    inp_pos_to_param_buffer_name = {}
    for node in new_gm.graph.nodes:
        if node.op == "placeholder":
            if node.name in ep.graph_signature.inputs_to_buffers:
                buffer_name = ep.graph_signature.inputs_to_buffers[node.name]
                if buffer_name in param_buffer_name_to_corrected_name:
                    inp_pos_to_param_buffer_name[count] = (
                        param_buffer_name_to_corrected_name[buffer_name]
                    )
                else:
                    inp_pos_to_param_buffer_name[count] = buffer_name
            if node.name in ep.graph_signature.inputs_to_parameters:
                param_name = ep.graph_signature.inputs_to_parameters[node.name]
                if param_name in param_buffer_name_to_corrected_name:
                    inp_pos_to_param_buffer_name[count] = (
                        param_buffer_name_to_corrected_name[param_name]
                    )
                else:
                    inp_pos_to_param_buffer_name[count] = param_name
            count += 1
    new_gm = _unlift(
        new_gm,
        inp_pos_to_param_buffer_name,
        ep.call_spec.in_spec,
        ep.call_spec.out_spec,
        ep.state_dict,
    )
    return new_gm