异步保存与分布式检查点 (DCP)#

创建日期：2024 年 7 月 22 日 | 最后更新：2025 年 7 月 10 日 | 最后验证：2024 年 11 月 05 日

作者： Lucas Pasqualin, Iris Zhang, Rodrigo Kumpera, Chien-Chin Huang

随着模型和全局规模的不断增长，检查点在分布式训练工作流的关键路径中常常成为瓶颈，成本也越来越高。一种抵消这种成本的绝佳策略是并行、异步地进行检查点。下面，我们将扩展分布式检查点入门教程中的保存示例，展示如何将其轻松集成到 torch.distributed.checkpoint.async_save 中。

您将学到什么

如何使用 DCP 并行生成检查点
优化性能的有效策略

先决条件

PyTorch v2.4.0 或更高版本
分布式检查点入门教程

异步检查点概述#

在开始异步检查点之前，了解其与同步检查点的区别和限制很重要。具体而言：

内存需求 - 异步检查点通过首先将模型复制到内部 CPU 缓冲区来工作。
这很有帮助，因为它确保在模型仍在检查点时模型和优化器权重不会发生变化，但会将 CPU 内存增加 checkpoint_size_per_rank X number_of_ranks 倍。此外，用户应注意了解其系统的内存限制。具体来说，固定内存（pinned memory）意味着使用 page-lock 内存，与 pageable 内存相比，page-lock 内存可能很稀缺。
检查点管理 - 由于检查点是异步的，用户需要负责管理并发运行的检查点。通常，用户可以
通过处理从 async_save 返回的 Future 对象来采用自己的管理策略。对于大多数用户，我们建议一次仅限制一个异步请求，避免为每个请求增加额外的内存压力。

import os

import torch
import torch.distributed as dist
import torch.distributed.checkpoint as dcp
import torch.multiprocessing as mp
import torch.nn as nn

from torch.distributed.fsdp import fully_shard
from torch.distributed.checkpoint.state_dict import get_state_dict, set_state_dict
from torch.distributed.checkpoint.stateful import Stateful

CHECKPOINT_DIR = "checkpoint"


class AppState(Stateful):
    """This is a useful wrapper for checkpointing the Application State. Since this object is compliant
    with the Stateful protocol, DCP will automatically call state_dict/load_stat_dict as needed in the
    dcp.save/load APIs.

    Note: We take advantage of this wrapper to hande calling distributed state dict methods on the model
    and optimizer.
    """

    def __init__(self, model, optimizer=None):
        self.model = model
        self.optimizer = optimizer

    def state_dict(self):
        # this line automatically manages FSDP FQN's, as well as sets the default state dict type to FSDP.SHARDED_STATE_DICT
        model_state_dict, optimizer_state_dict = get_state_dict(self.model, self.optimizer)
        return {
            "model": model_state_dict,
            "optim": optimizer_state_dict
        }

    def load_state_dict(self, state_dict):
        # sets our state dicts on the model and optimizer, now that we've loaded
        set_state_dict(
            self.model,
            self.optimizer,
            model_state_dict=state_dict["model"],
            optim_state_dict=state_dict["optim"]
        )

class ToyModel(nn.Module):
    def __init__(self):
        super(ToyModel, self).__init__()
        self.net1 = nn.Linear(16, 16)
        self.relu = nn.ReLU()
        self.net2 = nn.Linear(16, 8)

    def forward(self, x):
        return self.net2(self.relu(self.net1(x)))


def setup(rank, world_size):
    os.environ["MASTER_ADDR"] = "localhost"
    os.environ["MASTER_PORT"] = "12355 "

    # initialize the process group
    dist.init_process_group("gloo", rank=rank, world_size=world_size)
    torch.cuda.set_device(rank)


def cleanup():
    dist.destroy_process_group()


def run_fsdp_checkpoint_save_example(rank, world_size):
    print(f"Running basic FSDP checkpoint saving example on rank {rank}.")
    setup(rank, world_size)

    # create a model and move it to GPU with id rank
    model = ToyModel().to(rank)
    model = fully_shard(model)

    loss_fn = nn.MSELoss()
    optimizer = torch.optim.Adam(model.parameters(), lr=0.1)

    checkpoint_future = None
    for step in range(10):
        optimizer.zero_grad()
        model(torch.rand(8, 16, device="cuda")).sum().backward()
        optimizer.step()

        # waits for checkpointing to finish if one exists, avoiding queuing more then one checkpoint request at a time
        if checkpoint_future is not None:
            checkpoint_future.result()

        state_dict = { "app": AppState(model, optimizer) }
        checkpoint_future = dcp.async_save(state_dict, checkpoint_id=f"{CHECKPOINT_DIR}_step{step}")

    cleanup()


if __name__ == "__main__":
    world_size = torch.cuda.device_count()
    print(f"Running async checkpoint example on {world_size} devices.")
    mp.spawn(
        run_fsdp_checkpoint_save_example,
        args=(world_size,),
        nprocs=world_size,
        join=True,
    )

使用固定内存（Pinned Memory）获得更多性能#

如果上述优化仍然不够高效，您可以利用 GPU 模型的一项额外优化，该优化使用固定内存缓冲区进行检查点暂存。具体来说，这项优化解决了异步检查点的主要开销，即内存复制到检查点缓冲区。通过在检查点请求之间维护一个固定内存缓冲区，用户可以利用直接内存访问来加速此复制。

注意

此优化的主要缺点是缓冲区在检查点步骤之间的持续存在。如果没有固定内存优化（如上所示），任何检查点缓冲区都会在检查点完成后立即释放。通过固定内存实现，此缓冲区会在步骤之间保持，从而导致在应用程序生命周期内持续相同的峰值内存压力。

import os

import torch
import torch.distributed as dist
import torch.distributed.checkpoint as dcp
import torch.multiprocessing as mp
import torch.nn as nn

from torch.distributed.fsdp import fully_shard
from torch.distributed.checkpoint.state_dict import get_state_dict, set_state_dict
from torch.distributed.checkpoint.stateful import Stateful
from torch.distributed.checkpoint import FileSystemWriter as StorageWriter

CHECKPOINT_DIR = "checkpoint"


class AppState(Stateful):
    """This is a useful wrapper for checkpointing the Application State. Since this object is compliant
    with the Stateful protocol, DCP will automatically call state_dict/load_stat_dict as needed in the
    dcp.save/load APIs.

    Note: We take advantage of this wrapper to hande calling distributed state dict methods on the model
    and optimizer.
    """

    def __init__(self, model, optimizer=None):
        self.model = model
        self.optimizer = optimizer

    def state_dict(self):
        # this line automatically manages FSDP FQN's, as well as sets the default state dict type to FSDP.SHARDED_STATE_DICT
        model_state_dict, optimizer_state_dict = get_state_dict(self.model, self.optimizer)
        return {
            "model": model_state_dict,
            "optim": optimizer_state_dict
        }

    def load_state_dict(self, state_dict):
        # sets our state dicts on the model and optimizer, now that we've loaded
        set_state_dict(
            self.model,
            self.optimizer,
            model_state_dict=state_dict["model"],
            optim_state_dict=state_dict["optim"]
        )

class ToyModel(nn.Module):
    def __init__(self):
        super(ToyModel, self).__init__()
        self.net1 = nn.Linear(16, 16)
        self.relu = nn.ReLU()
        self.net2 = nn.Linear(16, 8)

    def forward(self, x):
        return self.net2(self.relu(self.net1(x)))


def setup(rank, world_size):
    os.environ["MASTER_ADDR"] = "localhost"
    os.environ["MASTER_PORT"] = "12355 "

    # initialize the process group
    dist.init_process_group("gloo", rank=rank, world_size=world_size)
    torch.cuda.set_device(rank)


def cleanup():
    dist.destroy_process_group()


def run_fsdp_checkpoint_save_example(rank, world_size):
    print(f"Running basic FSDP checkpoint saving example on rank {rank}.")
    setup(rank, world_size)

    # create a model and move it to GPU with id rank
    model = ToyModel().to(rank)
    model = fully_shard(model)

    loss_fn = nn.MSELoss()
    optimizer = torch.optim.Adam(model.parameters(), lr=0.1)

    # The storage writer defines our 'staging' strategy, where staging is considered the process of copying
    # checkpoints to in-memory buffers. By setting `cached_state_dict=True`, we enable efficient memory copying
    # into a persistent buffer with pinned memory enabled.
    # Note: It's important that the writer persists in between checkpointing requests, since it maintains the
    # pinned memory buffer.
    writer = StorageWriter(cache_staged_state_dict=True, path=CHECKPOINT_DIR)
    checkpoint_future = None
    for step in range(10):
        optimizer.zero_grad()
        model(torch.rand(8, 16, device="cuda")).sum().backward()
        optimizer.step()

        state_dict = { "app": AppState(model, optimizer) }
        if checkpoint_future is not None:
            # waits for checkpointing to finish, avoiding queuing more then one checkpoint request at a time
            checkpoint_future.result()
        dcp.async_save(state_dict, storage_writer=writer, checkpoint_id=f"{CHECKPOINT_DIR}_step{step}")

    cleanup()


if __name__ == "__main__":
    world_size = torch.cuda.device_count()
    print(f"Running fsdp checkpoint example on {world_size} devices.")
    mp.spawn(
        run_fsdp_checkpoint_save_example,
        args=(world_size,),
        nprocs=world_size,
        join=True,
    )

结论#

总之，我们学习了如何使用 DCP 的 async_save() API 在关键训练路径之外生成检查点。我们还了解了使用此 API 引入的额外内存和并发开销，以及利用固定内存（pinned memory）进一步提高速度的额外优化。

异步保存与分布式检查点 (DCP)#

异步检查点概述#

使用固定内存（Pinned Memory）获得更多性能#

结论#

文档

教程

资源