Pytorch实现并行训练通常有两个接口：DP（DataParallel）和DDP（DistributedDataParallel）。目前DP（DataParallel）已经被Pytorch官方deprecate掉了，原因有二：1，DP（DataParallel）只支持单机多卡，无法支持多机多卡；2，DP（DataParallel）即便在单机多卡模式下效率也不及DDP（DistributedDataParallel）。我们可以看一下官方文档里的描述：

DistributedDataParallel is proven to be significantly faster than torch.nn.DataParallel for single-node multi-GPU data parallel training.

基于这两个缺点，DP（DataParallel）本文就不介绍了，即使DP（DataParallel）更好上手。本文重点讲解如何使用DDP（DistributedDataParallel）来完成并行训练。

官方文档：https://pytorch.org/docs/stable/generated/torch.nn.parallel.DistributedDataParallel.html
官方视频教程：https://pytorch.org/tutorials/beginner/ddp_series_intro.html

---

DDP

DDP是基于多进程来实现并行训练，每个GPU依靠独立的进程来驱动，进程之间有特殊的通信机制。先介绍几个变量：

变量名	值	意义
rank	0	在节点中的编号，比如当前是第0张GPU
world_size	4	整个节点数量，比如一共4张GPU

我们先看一个单GPU代码，然后将其转换成DDP并行训练代码，观察变化来进行学习。
单GPU版本：（实验代码来自https://github.com/pytorch/examples/tree/main/distributed/ddp-tutorial-series）

import torch
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader

class MyTrainDataset(Dataset):
    def __init__(self, size):
        self.size = size
        self.data = [(torch.rand(20), torch.rand(1)) for _ in range(size)]

    def __len__(self):
        return self.size
    
    def __getitem__(self, index):
        return self.data[index]


class Trainer:
    def __init__(
        self,
        model: torch.nn.Module,
        train_data: DataLoader,
        optimizer: torch.optim.Optimizer,
        gpu_id: int,
        save_every: int,
    ) -> None:
        self.gpu_id = gpu_id
        self.model = model.to(gpu_id)
        self.train_data = train_data
        self.optimizer = optimizer
        self.save_every = save_every

    def _run_batch(self, source, targets):
        self.optimizer.zero_grad()
        output = self.model(source)
        loss = F.cross_entropy(output, targets)
        loss.backward()
        self.optimizer.step()

    def _run_epoch(self, epoch):
        b_sz = len(next(iter(self.train_data))[0])
        print(f"[GPU{self.gpu_id}] Epoch {epoch} | Batchsize: {b_sz} | Steps: {len(self.train_data)}")
        for source, targets in self.train_data:
            source = source.to(self.gpu_id)
            targets = targets.to(self.gpu_id)
            self._run_batch(source, targets)

    def _save_checkpoint(self, epoch):
        ckp = self.model.state_dict()
        PATH = "checkpoint.pt"
        torch.save(ckp, PATH)
        print(f"Epoch {epoch} | Training checkpoint saved at {PATH}")

    def train(self, max_epochs: int):
        for epoch in range(max_epochs):
            self._run_epoch(epoch)
            if epoch % self.save_every == 0:
                self._save_checkpoint(epoch)


def load_train_objs():
    train_set = MyTrainDataset(2048)  # load your dataset
    model = torch.nn.Linear(20, 1)  # load your model
    optimizer = torch.optim.SGD(model.parameters(), lr=1e-3)
    return train_set, model, optimizer


def prepare_dataloader(dataset: Dataset, batch_size: int):
    return DataLoader(
        dataset,
        batch_size=batch_size,
        pin_memory=True,
        shuffle=True
    )


def main(device, total_epochs, save_every, batch_size):
    dataset, model, optimizer = load_train_objs()
    train_data = prepare_dataloader(dataset, batch_size)
    trainer = Trainer(model, train_data, optimizer, device, save_every)
    trainer.train(total_epochs)


if __name__ == "__main__":
    import argparse
    parser = argparse.ArgumentParser(description='simple distributed training job')
    parser.add_argument('total_epochs', type=int, help='Total epochs to train the model')
    parser.add_argument('save_every', type=int, help='How often to save a snapshot')
    parser.add_argument('--batch_size', default=32, type=int, help='Input batch size on each device (default: 32)')
    args = parser.parse_args()
    
    device = 0  # shorthand for cuda:0
    main(device, args.total_epochs, args.save_every, args.batch_size)

直接复制上面代码，保存为single_gpu.py，然后用以下命令就可以让代码运行：

python single_gpu.py 10000 1000

这一步如果运行失败，请留言告知。如果运行成功，我们观察GPU，会发现该程序只会调用第一张GPU。我们将上述代码稍微改进，就可以得到多GPU版本了：

import torch
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader

import torch.multiprocessing as mp
from torch.utils.data.distributed import DistributedSampler
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.distributed import init_process_group, destroy_process_group
import os

class MyTrainDataset(Dataset):
    def __init__(self, size):
        self.size = size
        self.data = [(torch.rand(20), torch.rand(1)) for _ in range(size)]

    def __len__(self):
        return self.size
    
    def __getitem__(self, index):
        return self.data[index]

def ddp_setup(rank, world_size):
    """
    Args:
        rank: Unique identifier of each process
        world_size: Total number of processes
    """
    os.environ["MASTER_ADDR"] = "localhost"
    os.environ["MASTER_PORT"] = "12355"
    init_process_group(backend="nccl", rank=rank, world_size=world_size)
    torch.cuda.set_device(rank)

class Trainer:
    def __init__(
        self,
        model: torch.nn.Module,
        train_data: DataLoader,
        optimizer: torch.optim.Optimizer,
        gpu_id: int,
        save_every: int,
    ) -> None:
        self.gpu_id = gpu_id
        self.model = model.to(gpu_id)
        self.train_data = train_data
        self.optimizer = optimizer
        self.save_every = save_every
        self.model = DDP(model, device_ids=[gpu_id])

    def _run_batch(self, source, targets):
        self.optimizer.zero_grad()
        output = self.model(source)
        loss = F.cross_entropy(output, targets)
        loss.backward()
        self.optimizer.step()

    def _run_epoch(self, epoch):
        b_sz = len(next(iter(self.train_data))[0])
        print(f"[GPU{self.gpu_id}] Epoch {epoch} | Batchsize: {b_sz} | Steps: {len(self.train_data)}")
        self.train_data.sampler.set_epoch(epoch)
        for source, targets in self.train_data:
            source = source.to(self.gpu_id)
            targets = targets.to(self.gpu_id)
            self._run_batch(source, targets)

    def _save_checkpoint(self, epoch):
        ckp = self.model.module.state_dict()
        PATH = "checkpoint.pt"
        torch.save(ckp, PATH)
        print(f"Epoch {epoch} | Training checkpoint saved at {PATH}")

    def train(self, max_epochs: int):
        for epoch in range(max_epochs):
            self._run_epoch(epoch)
            if self.gpu_id == 0 and epoch % self.save_every == 0:
                self._save_checkpoint(epoch)


def load_train_objs():
    train_set = MyTrainDataset(2048)  # load your dataset
    model = torch.nn.Linear(20, 1)  # load your model
    optimizer = torch.optim.SGD(model.parameters(), lr=1e-3)
    return train_set, model, optimizer


def prepare_dataloader(dataset: Dataset, batch_size: int):
    return DataLoader(
        dataset,
        batch_size=batch_size,
        pin_memory=True,
        shuffle=False,
        sampler=DistributedSampler(dataset)
    )


def main(rank: int, world_size: int, save_every: int, total_epochs: int, batch_size: int):
    ddp_setup(rank, world_size)
    dataset, model, optimizer = load_train_objs()
    train_data = prepare_dataloader(dataset, batch_size)
    trainer = Trainer(model, train_data, optimizer, rank, save_every)
    trainer.train(total_epochs)
    destroy_process_group()


if __name__ == "__main__":
    import argparse
    parser = argparse.ArgumentParser(description='simple distributed training job')
    parser.add_argument('total_epochs', type=int, help='Total epochs to train the model')
    parser.add_argument('save_every', type=int, help='How often to save a snapshot')
    parser.add_argument('--batch_size', default=32, type=int, help='Input batch size on each device (default: 32)')
    args = parser.parse_args()
    
    world_size = torch.cuda.device_count()
    mp.spawn(main, args=(world_size, args.save_every, args.total_epochs, args.batch_size), nprocs=world_size)

将上述代码保存为mp_train.py，然后直接运行：

python mp_train.py 10000 1000

我们再观察GPU，会发现所有GPU均被调用。到此，实验部分已经结束。下面进入分析阶段。

---

从单卡变为多卡，需要进行哪些修改？

我们可以简单比对以下差别：

vimdiff single_gpu.py mp_train.py

vimdiff可以比较两个文档的差别，这是一个小trick，分享给大家。

• 第一步：初始化

import torch.multiprocessing as mp
from torch.utils.data.distributed import DistributedSampler
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.distributed import init_process_group, destroy_process_group
import os

def ddp_setup(rank, world_size):
   """
   Args:
       rank: Unique identifier of each process
       world_size: Total number of processes
   """
   os.environ["MASTER_ADDR"] = "localhost"
   os.environ["MASTER_PORT"] = "12355"
   init_process_group(backend="nccl", rank=rank, world_size=world_size)
   torch.cuda.set_device(rank)

init_process_group就是初始化群组，相当于给GPU们打了个招呼，我们要进行并行训练啦。这里的"os.environ"可以不用修改，因为只是单机多卡，不涉及多机多卡。"nccl"表示Nvidia Collective Communications Library，是一个跨GPU的通信后端类型，一般也不用修改。这里需要注意的是rank和world_size，比如咱们是4卡机，rank的取值范围是[0,1,2,3]，而world_size=4。

• 第二步：改造模型和数据加载器
  
  
  左边是单卡模式，右边是多卡模式。
• 步骤三：用多进程启动

world_size = torch.cuda.device_count()
mp.spawn(main, args=(world_size, args.save_every, args.total_epochs, args.batch_size), nprocs=world_size)

如有不解，欢迎留言讨论～