recompute.py

# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import contextlib
import weakref

import paddle
from paddle import framework
from paddle.autograd import PyLayer
from paddle.autograd.py_layer import LegacyPyLayer
from paddle.distributed.fleet.meta_parallel.parallel_layers.random import (
    get_rng_state_tracker,
)
from paddle.framework import core, in_dygraph_mode

from ..utils.log_util import logger

__all__ = []


def detach_variable(inputs):
    out = []
    for inp in inputs:
        if not isinstance(inp, (core.eager.Tensor, core.VarBase)):
            out.append(inp)
            continue

        x = inp.detach()
        x.stop_gradient = inp.stop_gradient
        out.append(x)
    return tuple(out)


def check_recompute_necessary(inputs):
    if not any(
        not input_.stop_gradient
        for input_ in inputs
        if isinstance(input_, (core.eager.Tensor, paddle.Tensor))
    ):
        logger.warning(
            "[Recompute]: None of the inputs to current recompute block need grad, "
            "therefore there is NO need to recompute this block in backward !"
        )


@contextlib.contextmanager
def swith_rng_state_tracker(rng_state, tracker):
    orig_rng_state = paddle.get_rng_state()
    orig_rng_tracker = get_rng_state_tracker().get_states_tracker()
    paddle.set_rng_state(rng_state)
    get_rng_state_tracker().set_states_tracker(tracker)
    try:
        yield
    finally:
        paddle.set_rng_state(orig_rng_state)
        get_rng_state_tracker().set_states_tracker(orig_rng_tracker)


class LegacyRecomputeFunction(LegacyPyLayer):
    @staticmethod
    def forward(ctx, run_function, preserve_rng_state, *args):
        # store for recomputing
        ctx.run_function = run_function
        ctx.preserve_rng_state = preserve_rng_state

        # NOTE the number of outputs of backward() should be equal to the number of tensors in forward()'s input
        # the order of tensors in backward()'s output should be the same as tensors in forward()'s input
        # None tensor inputs will be filtered in backward inputs.

        # save input for backward
        ctx.inputs = []
        ctx.tensor_indices = []
        tensor_inputs = []
        for i, arg in enumerate(args):
            if paddle.is_tensor(arg):
                tensor_inputs.append(arg)
                ctx.tensor_indices.append(i)
                ctx.inputs.append(None)
            else:
                ctx.inputs.append(arg)
        ctx.save_for_backward(*tensor_inputs)

        # NOTE recompute with restore RNG only support one senario where one process for one cuda gpu.
        # one process with multiple gpu and mix-gpu-cpu senarios are not support
        if ctx.preserve_rng_state:
            ctx.fw_rng_state = paddle.get_rng_state()
            ctx.fwd_rng_state_tracker = (
                get_rng_state_tracker().get_states_tracker()
            )

        # TODO support AMP
        tracer = framework._dygraph_tracer()
        ctx.is_fw_autocast = (
            False if tracer._amp_level == core.AmpLevel.O0 else True
        )
        if tracer._amp_level == core.AmpLevel.O2:
            ctx.amp_level = 'O2'
        elif tracer._amp_level in (core.AmpLevel.O1, core.AmpLevel.O0):
            ctx.amp_level = 'O1'
        else:
            raise ValueError(
                "unsupported amp level: {}".format(tracer._amp_level)
            )

        if tracer._amp_dtype == 'float16':
            ctx.amp_dtype = 'float16'
        elif tracer._amp_dtype in ('bfloat16', 'float32'):
            ctx.amp_dtype = 'bfloat16'
        else:
            raise ValueError(
                "unsupported amp dtype: {}".format(tracer._amp_dtype)
            )

        ctx.amp_white_list, ctx.amp_black_list = tracer._get_amp_op_list()

        with paddle.no_grad():
            outputs = run_function(*args)
        return outputs

    @staticmethod
    def backward(ctx, *args):
        with paddle.fluid.dygraph.guard():
            # TODO need to check the recompute calling is vaild or not

            # Restore inputs
            inputs = list(ctx.inputs)
            tensor_indices = ctx.tensor_indices
            tensors = ctx.saved_tensor()
            for i, idx in enumerate(tensor_indices):
                inputs[idx] = tensors[i]

            # paddle.enable_grad()
            tracer = framework._dygraph_tracer()
            tracer._has_grad = True

            # NOTE support AMP
            # need restore auto_cast state as well as w/b list
            if ctx.preserve_rng_state:
                with swith_rng_state_tracker(
                    ctx.fw_rng_state, ctx.fwd_rng_state_tracker
                ):
                    with paddle.amp.auto_cast(
                        enable=ctx.is_fw_autocast,
                        custom_white_list=ctx.amp_white_list,
                        custom_black_list=ctx.amp_black_list,
                        level=ctx.amp_level,
                        dtype=ctx.amp_dtype,
                    ):
                        detached_inputs = detach_variable(tuple(inputs))
                        outputs = ctx.run_function(*detached_inputs)
            else:
                with paddle.amp.auto_cast(
                    enable=ctx.is_fw_autocast,
                    custom_white_list=ctx.amp_white_list,
                    custom_black_list=ctx.amp_black_list,
                    level=ctx.amp_level,
                    dtype=ctx.amp_dtype,
                ):
                    detached_inputs = detach_variable(tuple(inputs))
                    outputs = ctx.run_function(*detached_inputs)

            if isinstance(outputs, core.VarBase):
                outputs = (outputs,)
            assert len(outputs) == len(args)

            # run backward() with only tensor that requires grad
            forward_outputs_with_grad = []
            # NOTE In Transformer-like network, if user put the attention mask into the recompute segment output,
            # pylayer will force the stop_gradient of attention mask to be False, which will make the number of
            # tensor that need grad does not match.
            # the following backward_inputs_with_grad is used to avoid this case.
            backward_inputs_with_grad = []
            for i in range(len(outputs)):
                if (
                    isinstance(outputs[i], core.VarBase)
                    and not outputs[i].stop_gradient
                ):
                    forward_outputs_with_grad.append(outputs[i])
                    backward_inputs_with_grad.append(args[i])

            if len(forward_outputs_with_grad) == 0:
                raise RuntimeError(
                    "none of output has requires_grad=True, this recompute() is not necessary"
                )

            # actually backward
            with paddle.amp.auto_cast(enable=False):
                paddle.autograd.backward(
                    forward_outputs_with_grad, backward_inputs_with_grad
                )

            grads = list(
                inp._grad_ivar()
                for inp in detached_inputs
                if isinstance(inp, core.VarBase)
            )
            return grads


class RecomputeFunction(PyLayer):
    @staticmethod
    def forward(ctx, run_function, preserve_rng_state, *args, **kwargs):
        # store for recomputing
        ctx.run_function = run_function
        ctx.preserve_rng_state = preserve_rng_state
        ctx.kwargs = kwargs

        # NOTE the number of outputs of backward() should be equal to the number of tensors in forward()'s input
        # the order of tensors in backward()'s output should be the same as tensors in forward()'s input
        # None tensor inputs will be filtered in backward inputs.

        # save input for backward
        ctx.inputs = []
        ctx.tensor_indices = []
        tensor_inputs = []
        for i, arg in enumerate(args):
            if paddle.is_tensor(arg):
                tensor_inputs.append(arg)
                ctx.tensor_indices.append(i)
                ctx.inputs.append(None)
            else:
                ctx.inputs.append(arg)
        ctx.save_for_backward(*tensor_inputs)

        # NOTE recompute with restore RNG only support one senario where one process for one cuda gpu.
        # one process with multiple gpu and mix-gpu-cpu senarios are not support
        if ctx.preserve_rng_state:
            ctx.fw_rng_state = paddle.get_rng_state()
            ctx.fwd_rng_state_tracker = (
                get_rng_state_tracker().get_states_tracker()
            )

        # TODO support AMP
        tracer = framework._dygraph_tracer()
        ctx.is_fw_autocast = (
            False if tracer._amp_level == core.AmpLevel.O0 else True
        )
        if tracer._amp_level == core.AmpLevel.O2:
            ctx.amp_level = 'O2'
        elif tracer._amp_level in (core.AmpLevel.O1, core.AmpLevel.O0):
            ctx.amp_level = 'O1'
        else:
            raise ValueError(
                "unsupported amp level: {}".format(tracer._amp_level)
            )

        if tracer._amp_dtype == 'float16':
            ctx.amp_dtype = 'float16'
        elif tracer._amp_dtype in ('bfloat16', 'float32'):
            ctx.amp_dtype = 'bfloat16'
        else:
            raise ValueError(
                "unsupported amp dtype: {}".format(tracer._amp_dtype)
            )

        ctx.amp_white_list, ctx.amp_black_list = tracer._get_amp_op_list()

        with paddle.no_grad():
            outputs = run_function(*args, **kwargs)
        return outputs

    @staticmethod
    def backward(ctx, *args):
        with paddle.fluid.dygraph.guard():
            # TODO need to check the recompute calling is vaild or not

            # Restore inputs
            inputs = list(ctx.inputs)
            tensor_indices = ctx.tensor_indices
            tensors = ctx.saved_tensor()
            for i, idx in enumerate(tensor_indices):
                inputs[idx] = tensors[i]

            # paddle.enable_grad()
            tracer = framework._dygraph_tracer()
            tracer._has_grad = True

            # NOTE support AMP
            # need restore auto_cast state as well as w/b list
            if ctx.preserve_rng_state:
                with swith_rng_state_tracker(
                    ctx.fw_rng_state, ctx.fwd_rng_state_tracker
                ):
                    with paddle.amp.auto_cast(
                        enable=ctx.is_fw_autocast,
                        custom_white_list=ctx.amp_white_list,
                        custom_black_list=ctx.amp_black_list,
                        level=ctx.amp_level,
                        dtype=ctx.amp_dtype,
                    ):
                        detached_inputs = detach_variable(tuple(inputs))
                        outputs = ctx.run_function(
                            *detached_inputs, **ctx.kwargs
                        )
            else:
                with paddle.amp.auto_cast(
                    enable=ctx.is_fw_autocast,
                    custom_white_list=ctx.amp_white_list,
                    custom_black_list=ctx.amp_black_list,
                    level=ctx.amp_level,
                    dtype=ctx.amp_dtype,
                ):
                    detached_inputs = detach_variable(tuple(inputs))
                    outputs = ctx.run_function(*detached_inputs, **ctx.kwargs)

            if isinstance(outputs, (core.VarBase, core.eager.Tensor)):
                outputs = (outputs,)
            assert len(outputs) == len(args)

            # run backward() with only tensor that requires grad
            forward_outputs_with_grad = []
            # NOTE In Transformer-like network, if user put the attention mask into the recompute segment output,
            # pylayer will force the stop_gradient of attention mask to be False, which will make the number of
            # tensor that need grad does not match.
            # the following backward_inputs_with_grad is used to avoid this case.
            backward_inputs_with_grad = []
            for i in range(len(outputs)):
                if (
                    isinstance(outputs[i], (core.VarBase, core.eager.Tensor))
                    and not outputs[i].stop_gradient
                ):
                    forward_outputs_with_grad.append(outputs[i])
                    backward_inputs_with_grad.append(args[i])

            if len(forward_outputs_with_grad) == 0:
                raise RuntimeError(
                    "none of output has requires_grad=True, this recompute() is not necessary"
                )

            # actually backward
            with paddle.amp.auto_cast(enable=False):
                paddle.autograd.backward(
                    forward_outputs_with_grad, backward_inputs_with_grad
                )

            if in_dygraph_mode():
                grads = tuple(
                    inp._grad_ivar()
                    for inp in detached_inputs
                    if isinstance(inp, (core.VarBase, core.eager.Tensor))
                )
            else:
                grads = list(
                    inp._grad_ivar()
                    for inp in detached_inputs
                    if isinstance(inp, (core.VarBase, core.eager.Tensor))
                )
            return grads


def _recompute_without_reentrant(
    function, preserve_rng_state=True, *args, **kwargs
):
    """
    recompute without reentrant, that means use hook to implement the recompute function rather than re-entrant autograd.
    """

    if preserve_rng_state:
        cur_device = paddle.get_device()
        if 'gpu:' not in cur_device:
            raise RuntimeError(
                "Recompute with RNG perserve is not support current device: {}.".format(
                    cur_device
                )
            )
        fw_cuda_rng_state = paddle.get_cuda_rng_state()
        fwd_cuda_rng_state_tracker = (
            get_rng_state_tracker().get_states_tracker()
        )
    tracer = framework._dygraph_tracer()
    is_fw_autocast = False if tracer._amp_level == core.AmpLevel.O0 else True
    if tracer._amp_level == core.AmpLevel.O2:
        amp_level = 'O2'
    elif tracer._amp_level in (core.AmpLevel.O1, core.AmpLevel.O0):
        amp_level = 'O1'

    if tracer._amp_dtype == 'float16':
        amp_dtype = 'float16'
    elif tracer._amp_dtype in ('bfloat16', 'float32'):
        amp_dtype = 'bfloat16'

    amp_white_list, amp_black_list = tracer._get_amp_op_list()

    class Intermediate_Holder:
        pass

    storage = weakref.WeakKeyDictionary()
    holder_list = []

    def pack(x):
        res = Intermediate_Holder()
        holder_list.append(weakref.ref(res))
        return res

    def unpack(x):
        unpack_counter = 0
        if len(storage) == 0:

            def inner_pack(inner_x):
                nonlocal unpack_counter
                unpack_counter += 1

                if holder_list[unpack_counter - 1]() is None:
                    return

                tmp_tensor = core.eager.Tensor(
                    inner_x.dtype,
                    inner_x.shape,
                    inner_x.name + "cpy",
                    core.VarDesc.VarType.LOD_TENSOR,
                    inner_x.persistable,
                )
                inner_x._share_buffer_to(tmp_tensor)
                storage[holder_list[unpack_counter - 1]()] = tmp_tensor
                return

            def inner_unpack(inner_x):
                raise Exception("An unexcepted backward called on a tensor!")

            if preserve_rng_state:
                with swith_rng_state_tracker(
                    fw_cuda_rng_state, fwd_cuda_rng_state_tracker
                ):
                    with paddle.set_grad_enabled(True):
                        with paddle.amp.auto_cast(
                            enable=is_fw_autocast,
                            custom_white_list=amp_white_list,
                            custom_black_list=amp_black_list,
                            level=amp_level,
                            dtype=amp_dtype,
                        ):
                            with paddle.autograd.saved_tensors_hooks(
                                inner_pack, inner_unpack
                            ):
                                unused_outputs = function(*args, **kwargs)
            else:
                with paddle.set_grad_enabled(True), paddle.amp.auto_cast(
                    enable=is_fw_autocast,
                    custom_white_list=amp_white_list,
                    custom_black_list=amp_black_list,
                    level=amp_level,
                    dtype=amp_dtype,
                ), paddle.autograd.saved_tensors_hooks(
                    inner_pack, inner_unpack
                ):
                    unused_outputs = function(*args, **kwargs)

        if x not in storage:
            raise Exception(
                "Not supported to retrieve a tensor saved by autograd multiple times that is no need to recompute."
            )

        return storage[x]

    with paddle.autograd.saved_tensors_hooks(pack, unpack):
        outputs = function(*args, **kwargs)

    return outputs


def recompute(function, *args, **kwargs):
    """
    recompute intermediate activations to save then memory.

    Parameters:
        function(paddle.nn.Layer): layer of sequence of layers that describes part of forward pass of the model
              whose intermediate activations will be released to save memory in forward stage and will be recomputed
              in backward stage for gradient calculation.
        *args(Tensor): inputs to the function.
        **kwargs(Dict): Kwargs should only contain two kinds of key-value params, the one is part of function's key-value params,
                        and the other contains 'preserve_rng_state' and 'use_reentrant'. the key-value pair of preserve_rng_state,
                        which is used to indicate whether to save the forward rng. If it is True, then the last forward rng value
                        will be restored when the forward recalculation of backpropagation is performed, its default value is True.
                        the key-value pair of use_reentrant is used to indicate which implementation of recompute you will be used.
                        'use_reentrant=True' means to use the PyLayer implementation of recompute, 'use_reentrant=False' means to
                        use the Hook implementation of recompute, its default value is True.
    Returns:
        Output of function on args.

    Examples:
        .. code-block:: python

            import paddle
            from paddle.distributed.fleet.utils import recompute
            import random
            # required: gpu
            def get_fc_block(block_idx, input_size, is_last=False):
                block_name = "block_" + str(block_idx)
                block = paddle.nn.Sequential(
                    (block_name + "_fc_0", paddle.nn.Linear(input_size, input_size, bias_attr=False)),
                    (block_name + "_dropout", paddle.nn.Dropout(p=0.5)),
                    (block_name + "_relu_1", paddle.nn.ReLU()),
                    (block_name + "_fc_1", paddle.nn.Linear(input_size, input_size, bias_attr=False)),
                    (block_name + "_relu_2", paddle.nn.ReLU()),
                )
                if is_last:
                    block.add_sublayer(
                        block_name + "_fc_2",
                        paddle.nn.Linear(
                            input_size, 1, bias_attr=False
                        )
                    )
                else:
                    block.add_sublayer(
                        block_name + "_fc_2",
                        paddle.nn.Linear(input_size, input_size, bias_attr=False)
                    )
                return block
            class Naive_fc_net(paddle.nn.Layer):
                def __init__(self, input_size=10,
                            recompute_blocks=[1, 3],
                            recompute_kwargs={}):
                    super(Naive_fc_net, self).__init__()
                    self.recompute_blocks = recompute_blocks
                    self.recompute_kwargs = recompute_kwargs
                    self.runfunc0 = get_fc_block(0, input_size, is_last=False)
                    self.runfunc1 = get_fc_block(1, input_size, is_last=False)
                    self.runfunc2 = get_fc_block(2, input_size, is_last=False)
                    self.runfunc3 = get_fc_block(3, input_size, is_last=False)
                    self.runfunc4 = get_fc_block(4, input_size, is_last=True)
                    self.total_func = [self.runfunc0, self.runfunc1, self.runfunc2, self.runfunc3, self.runfunc4]
                def forward(self, inputs):
                    nums = len(self.total_func)
                    for i in range(nums):
                        if i in self.recompute_blocks:
                            inputs = recompute(self.total_func[i], inputs, **{"preserve_rng_state": True})
                        else:
                            inputs = self.total_func[i](inputs)
                    return inputs
            def run_model(cuda_state, recompute_block=[], recompute_kwargs={}):
                gen = paddle.seed(10)
                gen.manual_seed(10)
                random.seed(10)
                if cuda_state:
                    paddle.set_cuda_rng_state(cuda_state)
                batch_size, input_size = 1, 10
                model = Naive_fc_net(
                    input_size,
                    recompute_blocks=recompute_block,
                    recompute_kwargs=recompute_kwargs)
                optimizer = paddle.optimizer.SGD(learning_rate=0.01, parameters=model.parameters())
                loss_ = []
                param_ = []
                grad_ = []
                for _ in range(5):
                    x = paddle.rand(shape=[batch_size, input_size], dtype="float32")
                    y_pred = model(x)
                    loss = y_pred.mean()
                    loss_.append(loss.item())
                    loss.backward()
                    optimizer.step()
                    param_.append(model.parameters()[9])
                    grad_.append(model.parameters()[3]._grad_ivar())
                    optimizer.clear_grad()
                return loss_, param_, grad_
            cuda_state = paddle.get_cuda_rng_state()
            # without recompute
            loss_ref, param_ref, grad_ref = run_model(
                cuda_state, recompute_block=[]
            )
            loss, param, grad = run_model(cuda_state, recompute_block=[1, 2])
            print("normal_loss: {}, recompute_loss: {}".format(loss_ref, loss))
            # The result of the recompute_loss should be the same as the normal_loss.
    """
    # Hack to mix *args with **kwargs in a python 2.7-compliant way
    preserve = kwargs.pop('preserve_rng_state', True)

    # whether to use reentrant method to implement recompute
    use_reentrant = kwargs.pop('use_reentrant', True)

    if kwargs and use_reentrant:
        raise ValueError(
            "Error, if you want to send kwargs(dict parameter) to function, please set use_reentrant=False."
        )

    if framework._dygraph_tracer()._has_grad:
        check_recompute_necessary(args)

    if use_reentrant:
        return RecomputeFunction.apply(function, preserve, *args)
    else:
        return _recompute_without_reentrant(function, preserve, *args, **kwargs)


def recompute_sequential(ctx, functions, *args, **kwargs):
    """
    recompute intermediate activations to save then memory for 'Sequential' models.

    Parameters:
        ctx(dict): include 'segments' and  'preserve_rng_state' keys, the key 'segments' (int, default 1), represents the number of chunks to create in the model,
                   the key 'preserve_rng_state' (bool, optional, default=True) indicate whether to save the forward rng. If it is True, then the last forward rng value will be
                   restored when the forward recalculation of backpropagation is performed. and some keys such as 'mp_group', 'offload' and 'partition' are invalid here,
                   they are useful in 'recompute_hybrid' API.
        functions(paddle.nn.Sequential): layer of sequence of layers that describes part of forward pass of the model
              whose intermediate activations will be released to save memory in forward stage and will be recomputed
              in backward stage for gradient calculation.
        *args(Tensor): inputs(tuple) to the function.
        **kwargs(Dict): inputs(dict) to the function.

    Returns:
        Output of function on args and kwargs.

    Examples:
        .. code-block:: python

            model = paddle.nn.Sequential(...)
            input = recompute_sequential({'segments' : 1}, model, input)
    """
    segments = ctx.get('segments', 1)
    preserve_rng_state = ctx.get('preserve_rng_state', True)

    def _run_func(begin, end, funcs):
        def do_run(input):
            for i in range(begin, end + 1):
                input = funcs[i](input)
            return input

        return do_run

    if isinstance(functions, paddle.nn.Sequential):
        functions = list(functions.children())

    segment_size = len(functions) // segments

    end = -1
    for begin in range(0, segment_size * (segments - 1), segment_size):
        end = begin + segment_size - 1
        args = recompute(
            _run_func(begin, end, functions),
            *args,
            preserve_rng_state=preserve_rng_state,
            **kwargs
        )
    return _run_func(end + 1, len(functions) - 1, functions)(args)