weights_to_outputs_pass.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.

# pyre-unsafe

from torch.export import ExportedProgram
from torch.export.exported_program import OutputKind, OutputSpec, TensorArgument


def weights_to_outputs_pass(
    exported_program: ExportedProgram,
) -> ExportedProgram:
    """
    This pass is for training graphs with gradients returned. It flags the weights as having a gradient attached,
    and appends them to the outputs in order to make the weights easier to handle in memory planning and the emitter.

    Args:
        exported_program: The ExportedProgram to update.

    Returns:
        The modified ExportedProgram.
    """
    if (
        len([node for node in exported_program.graph.nodes if node.op == "placeholder"])
        == 0
    ):
        return exported_program

    gs = exported_program.graph_signature
    gm = exported_program.graph_module

    # Check for/ get gradients
    grad_targets = [
        spec.target
        for spec in gs.output_specs
        if spec.kind == OutputKind.GRADIENT_TO_PARAMETER
    ]

    # If no gradients, return
    if len(grad_targets) == 0:
        return exported_program

    inputs_to_params = gs.inputs_to_parameters

    # Get output node
    output_node = None
    for node in gm.graph.nodes:
        if node.op == "output":
            output_node = node
            break
    assert output_node is not None

    # Get input nodes that are weights with an associated gradient
    placeholder_nodes = [
        node
        for node in gm.graph.nodes
        if node.op == "placeholder"
        and node.target in inputs_to_params.keys()
        and inputs_to_params[node.target] in grad_targets
    ]

    # Flag these placeholder nodes as having a gradient attached so that memory planning will operate on them.
    for node in placeholder_nodes:
        node.meta["weight_has_gradient"] = True

    # add to output node
    new_output_nodes = []
    new_output_nodes.extend(output_node.args[0])
    new_output_nodes.extend(placeholder_nodes)
    # Remove old outputs
    new_output = gm.graph.output(tuple(new_output_nodes))
    output_node.replace_all_uses_with(new_output)
    gm.graph.erase_node(output_node)

    # add to output signature
    for node in placeholder_nodes:
        gs.output_specs.append(
            OutputSpec(
                OutputKind.TOKEN,  # This is a hack. We are returning the raw weights here to make it easier for memory
                # planning and the emitter. There is no outputkind.Parameter so I am using TOKEN which is currently unused in Edge.
                TensorArgument(node.target),
                None,
            )
        )

    # Cleanup the graph.
    exported_program.graph.eliminate_dead_code()
    exported_program.graph_module.recompile()

    return exported_program