group_sharded.py

# Copyright (c) 2022 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 logging
import os

import paddle
from paddle.distributed.fleet.meta_parallel.sharding.group_sharded_optimizer_stage2 import (
    GroupShardedOptimizerStage2,
)
from paddle.distributed.fleet.meta_parallel.sharding.group_sharded_stage2 import (
    GroupShardedStage2,
)
from paddle.distributed.fleet.meta_parallel.sharding.group_sharded_stage3 import (
    GroupShardedStage3,
)
from paddle.distributed.fleet.meta_parallel.sharding.group_sharded_utils import (
    GroupShardedScaler,
)
from paddle.distributed.fleet.utils.mix_precision_utils import (
    MixPrecisionOptimizer,
)
from paddle.distributed.utils.log_utils import get_logger
from paddle.optimizer import Optimizer

logger_ = get_logger(logging.WARNING)


def group_sharded_parallel(
    model,
    optimizer,
    level,
    scaler=None,
    group=None,
    offload=False,
    sync_buffers=False,
    buffer_max_size=2**23,
    segment_size=2**20,
    sync_comm=False,
    dp_group=None,
    exclude_layer=None,
):
    """
    Use group_sharded_parallel can perform group shared configuration on the model, optimizer and GradScaler. Level has three string options, 'os', 'os_g' and 'p_g_os' corresponds to three different usage scenarios: optimizer state segmentation, optimizer state + gradient segmentation, and parameter + gradient + optimizer state segmentation.
    Usually, optimizer state + gradient segmentation is actually a re optimization of optimizer state segmentation, so optimizer state + gradient segmentation can be used to realize optimizer state segmentation.

    Args:
        model (Layer): The layer to be wrapped with group_sharded_parallel.
        optimizer (Optimizer): The optimizer to be wrapped with group_sharded_parallel.
        level (str): The different level of the group sharded. Such as `os`, `os_g`, `p_g_os`.
        scaler (GradScaler, optional): If AMP is used, you need to pass GradScaler. Defaults to None, indicating that GradScaler is not used.
        group (Group, optional): The group instance. Defaults to None, indicating that the default environment group is used.
        offload (bool, optional): Whether to use the offload function. Defaults to False, which means that the offload function is not used.
        sync_buffers (bool, optional): Whether to broadcast model buffers. It is generally used when there are registered model buffers. Defaults to False, indicating that model buffers are not used.
        buffer_max_size (int, optional): The max size of the buffer used to integrate gradient in `os_g`. The larger the size, the more GPU memory will be used. Defaults to 2**23, which means that the dimension of the buffer is 2**23.
        segment_size (int, optional): The smallest size of parameter to be sharded in `p_g_os`. Defaults to 2**20, indicating that the dimension of the minimum segmented parameter is 2**20.
        sync_comm (bool, optional): Whether to use synchronous communication, only in `p_g_os` used. Defaults to False, indicating that asynchronous communication is used.
        dp_group(Group, optional): dp communication group, support to combine stage2 or stage3 with dp hybrid communication.
        exclude_layer(list, optional): exclude some layers for slicing for sharding stage3, for example, exclude_layer=["GroupNorm", id(model.gpt.linear)], exclude_layer must contain the layers' name or one layer's id.

    Returns:
        model: A wrapper for group sharded given model.
        optimizer: A wrapper for group sharded given optimizer.
        scaler: A wrapper for group sharded given scaler.

    Examples:
        .. code-block:: python

            >>> # doctest: +REQUIRES(env:DISTRIBUTED)
            >>> import paddle
            >>> from paddle.nn import Linear
            >>> from paddle.distributed import fleet
            >>> from paddle.distributed.sharding import group_sharded_parallel

            >>> fleet.init(is_collective=True)
            >>> group = paddle.distributed.new_group([0, 1])
            >>> model = Linear(1000, 1000)

            >>> clip = paddle.nn.ClipGradByGlobalNorm(clip_norm=1.0)
            >>> optimizer = paddle.optimizer.AdamW(learning_rate=0.001, parameters=model.parameters(), weight_decay=0.00001, grad_clip=clip)

            >>> # wrap sharding model, optimizer and scaler
            >>> model, optimizer, scaler = group_sharded_parallel(model, optimizer, "p_g", scaler=scaler)

            >>> img, label = data
            >>> label.stop_gradient = True
            >>> img.stop_gradient = True

            >>> out = model(img)
            >>> loss = paddle.nn.functional.cross_entropy(input=out, label=label)

            >>> loss.backward()
            >>> optimizer.step()
            >>> optimizer.clear_grad()

    """

    device = paddle.get_device().split(":")[0]
    assert (
        device
        in [
            "gpu",
            "xpu",
        ]
        or device in paddle.device.get_all_custom_device_type()
    ), "group_sharded_parallel only support gpu, xpu and custom_device now"
    # check option type
    assert isinstance(
        model, paddle.nn.Layer
    ), "The model must be the instance of paddle.nn.Layer."
    assert isinstance(optimizer, (MixPrecisionOptimizer, Optimizer)), (
        "The optimizer must be the instance of paddle.optimizer.Optimizer "
        "or MixPrecisionOptimizer for main grad."
    )
    assert level in [
        'os',
        'os_g',
        'p_g_os',
    ], "The level must be os, os_g or p_g_os."

    def check_dtype(param):
        return param.dtype == paddle.float16

    params_fp16 = list(filter(check_dtype, model.parameters()))
    if scaler is None and len(params_fp16) > 0:
        logger_.warning(
            "the input of scaler is None, please ensure the logic of your scaler outside is same as GroupShardedScaler."
        )
    # convert model/optimizer/scaler
    if level in ['os', 'os_g']:
        logger_.info("*" * 30)
        logger_.info("Sharded level os uses sharded level os_g achieved now.")
        logger_.info("*" * 30)
        optimizer = GroupShardedOptimizerStage2(
            params=optimizer._parameter_list,
            optim=optimizer,
            group=group,
            offload=offload,
            dp_group=dp_group,
            device=device,
        )
        model = GroupShardedStage2(
            model,
            optimizer,
            group=group,
            sync_buffers=sync_buffers,
            buffer_max_size=buffer_max_size,
            dp_group=dp_group,
            device=device,
        )
    elif level == 'p_g_os':
        model = GroupShardedStage3(
            model,
            optimizer=optimizer,
            group=group,
            sync_buffers=sync_buffers,
            segment_size=segment_size,
            offload=offload,
            sync_comm=sync_comm,
            dp_group=dp_group,
            device=device,
            exclude_layer=exclude_layer,
        )
    else:
        raise ValueError("Please enter the correct level.")
    if isinstance(scaler, paddle.amp.GradScaler):
        scaler = GroupShardedScaler(scaler)
    logger_.info("*" * 30)
    logger_.info(
        "If there is a communication hang using group sharded, please check whether the communication operations of each process are unified."
    )
    logger_.info("*" * 30)

    return model, optimizer, scaler


def save_group_sharded_model(model, output, optimizer=None):
    """
    Group sharded encapsulated model and optimizer state saving module.

    Note:
        If using save_group_sharded_model saves the model. When loading again, you need to set the model or optimizer state before using group_sharded_parallel.

    Args:
        model (Layer): A wrapper for group sharded given model.
        output (str): Save directory.
        optimizer (Optimizer, optional): Group sharded encapsulated optimizer. Defaults to None, indicating that the optimizer state is not saved.

    Examples:
        .. code-block:: python

            >>> # doctest: +REQUIRES(env:DISTRIBUTED)
            >>> import paddle
            >>> from paddle.nn import Linear
            >>> from paddle.distributed import fleet
            >>> from paddle.distributed.sharding import group_sharded_parallel, save_group_sharded_model

            >>> fleet.init(is_collective=True)
            >>> group = paddle.distributed.new_group([0, 1])
            >>> model = Linear(1000, 1000)

            >>> clip = paddle.nn.ClipGradByGlobalNorm(clip_norm=1.0)
            >>> optimizer = paddle.optimizer.AdamW(learning_rate=0.001, parameters=model.parameters(), weight_decay=0.00001, grad_clip=clip)

            >>> # wrap sharding model, optimizer and scaler
            >>> model, optimizer, scaler = group_sharded_parallel(model, optimizer, "p_g", scaler=scaler)

            >>> img, label = data
            >>> label.stop_gradient = True
            >>> img.stop_gradient = True

            >>> out = model(img)
            >>> loss = paddle.nn.functional.cross_entropy(input=out, label=label)

            >>> loss.backward()
            >>> optimizer.step()
            >>> optimizer.clear_grad()

            >>> # save model and optimizer state_dict
            >>> save_group_sharded_model(model, optimizer, output=output_dir)

    """
    logger_.info(
        "==========Begin to save group sharded model and optimizer=========="
    )
    assert not os.path.isfile(
        output
    ), f"Saving directory ({output}) should be a directory, not a file"
    os.makedirs(output, exist_ok=True)
    output_model = os.path.join(output, "model.pdmodel")
    if isinstance(model, GroupShardedStage2):
        paddle.save(model._layer.state_dict(), output_model)
    elif isinstance(model, GroupShardedStage3):
        convert2cpu = True if model._offload else False
        model.get_all_parameters(convert2cpu=convert2cpu)
        paddle.save(model._layer.state_dict(), output_model)
    else:
        raise ValueError(
            "Please use the layer which is wrapped with group_sharded_parallel."
        )

    if optimizer is not None:
        assert hasattr(
            optimizer, "_optim"
        ), "Please use the optimizer which is wrapped with group_sharded_parallel."
        output_opt = os.path.join(output, "model.pdopt")
        paddle.save(optimizer._optim.state_dict(), output_opt)
    logger_.info(
        "==========End to save group sharded model and optimizer=========="
    )