# 使用 Distributed Data Parallel 和 Pipeline Parallelism 训练 Transformer 模型 > 原文：[`pytorch.org/tutorials/advanced/ddp_pipeline.html`](https://pytorch.org/tutorials/advanced/ddp_pipeline.html) > > 译者：[飞龙](https://github.com/wizardforcel) > > 协议：[CC BY-NC-SA 4.0](http://creativecommons.org/licenses/by-nc-sa/4.0/) 注意 点击这里下载完整示例代码 **作者**：[Pritam Damania](https://github.com/pritamdamania87) 本教程演示了如何使用[Distributed Data Parallel](https://pytorch.org/docs/stable/generated/torch.nn.parallel.DistributedDataParallel.html)和[Pipeline Parallelism](https://pytorch.org/docs/stable/pipeline.html)在多个 GPU 上训练大型 Transformer 模型。本教程是[使用 nn.Transformer 和 TorchText 进行序列到序列建模](https://pytorch.org/tutorials/beginner/transformer_tutorial.html)教程的延伸，扩展了相同的模型以演示如何使用 Distributed Data Parallel 和 Pipeline Parallelism 来训练 Transformer 模型。 先决条件： > + [管道并行](https://pytorch.org/docs/stable/pipeline.html) > + > + [使用 nn.Transformer 和 TorchText 进行序列到序列建模](https://pytorch.org/tutorials/beginner/transformer_tutorial.html) > + > + [使用分布式数据并行开始](https://pytorch.org/tutorials/intermediate/ddp_tutorial.html) ## 定义模型 `PositionalEncoding` 模块向序列中的令牌注入了一些关于相对或绝对位置的信息。位置编码与嵌入的维度相同，因此可以将两者相加。在这里，我们使用不同频率的 `sine` 和 `cosine` 函数。 ```py import sys import os import math import torch import torch.nn as nn import torch.nn.functional as F import tempfile from torch.nn import TransformerEncoder, TransformerEncoderLayer class PositionalEncoding(nn.Module): def __init__(self, d_model, dropout=0.1, max_len=5000): super(PositionalEncoding, self).__init__() self.dropout = nn.Dropout(p=dropout) pe = torch.zeros(max_len, d_model) position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1) div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model)) pe[:, 0::2] = torch.sin(position * div_term) pe[:, 1::2] = torch.cos(position * div_term) pe = pe.unsqueeze(0).transpose(0, 1) self.pe = nn.Parameter(pe, requires_grad=False) def forward(self, x): x = x + self.pe[:x.size(0), :] return self.dropout(x) ``` 在本教程中，我们将一个 Transformer 模型分割到两个 GPU 上，并使用管道并行来训练模型。除此之外，我们使用[Distributed Data Parallel](https://pytorch.org/docs/stable/generated/torch.nn.parallel.DistributedDataParallel.html)来训练这个管道的两个副本。我们有一个进程在 GPU 0 和 1 之间驱动一个管道，另一个进程在 GPU 2 和 3 之间驱动一个管道。然后，这两个进程使用 Distributed Data Parallel 来训练这两个副本。模型与[使用 nn.Transformer 和 TorchText 进行序列到序列建模](https://pytorch.org/tutorials/beginner/transformer_tutorial.html)教程中使用的模型完全相同，但被分成了两个阶段。最多的参数属于[nn.TransformerEncoder](https://pytorch.org/docs/stable/generated/torch.nn.TransformerEncoder.html)层。[nn.TransformerEncoder](https://pytorch.org/docs/stable/generated/torch.nn.TransformerEncoder.html)本身由`nlayers`个[nn.TransformerEncoderLayer](https://pytorch.org/docs/stable/generated/torch.nn.TransformerEncoderLayer.html)组成。因此，我们的重点是`nn.TransformerEncoder`，我们将模型分割成一半的`nn.TransformerEncoderLayer`在一个 GPU 上，另一半在另一个 GPU 上。为此，我们将`Encoder`和`Decoder`部分提取到单独的模块中，然后构建一个代表原始 Transformer 模块的`nn.Sequential`。 ```py if sys.platform == 'win32': print('Windows platform is not supported for pipeline parallelism') sys.exit(0) if torch.cuda.device_count() < 4: print('Need at least four GPU devices for this tutorial') sys.exit(0) class Encoder(nn.Module): def __init__(self, ntoken, ninp, dropout=0.5): super(Encoder, self).__init__() self.pos_encoder = PositionalEncoding(ninp, dropout) self.encoder = nn.Embedding(ntoken, ninp) self.ninp = ninp self.init_weights() def init_weights(self): initrange = 0.1 self.encoder.weight.data.uniform_(-initrange, initrange) def forward(self, src): # Need (S, N) format for encoder. src = src.t() src = self.encoder(src) * math.sqrt(self.ninp) return self.pos_encoder(src) class Decoder(nn.Module): def __init__(self, ntoken, ninp): super(Decoder, self).__init__() self.decoder = nn.Linear(ninp, ntoken) self.init_weights() def init_weights(self): initrange = 0.1 self.decoder.bias.data.zero_() self.decoder.weight.data.uniform_(-initrange, initrange) def forward(self, inp): # Need batch dimension first for output of pipeline. return self.decoder(inp).permute(1, 0, 2) ``` ## 启动多个进程进行训练 我们启动两个进程，每个进程在两个 GPU 上驱动自己的管道。对于每个进程，都会执行`run_worker`。 ```py def run_worker(rank, world_size): ``` ## 加载和批处理数据 训练过程使用了来自`torchtext`的 Wikitext-2 数据集。要访问 torchtext 数据集，请按照[`github.com/pytorch/data`](https://github.com/pytorch/data)上的说明安装 torchdata。 vocab 对象是基于训练数据集构建的，并用于将令牌数值化为张量。从顺序数据开始，`batchify()` 函数将数据集排列成列，将数据分成大小为 `batch_size` 的批次后，修剪掉任何剩余的令牌。例如，对于字母表作为序列（总长度为 26）和批次大小为 4，我们将字母表分成长度为 6 的 4 个序列： $$ \begin{bmatrix} \text{A} & \text{B} & \text{C} & \ldots & \text{X} & \text{Y} & \text{Z} \end{bmatrix} \Rightarrow \begin{bmatrix} \begin{bmatrix}\text{A} \\ \text{B} \\ \text{C} \\ \text{D} \\ \text{E} \\ \text{F}\end{bmatrix} & \begin{bmatrix}\text{G} \\ \text{H} \\ \text{I} \\ \text{J} \\ \text{K} \\ \text{L}\end{bmatrix} & \begin{bmatrix}\text{M} \\ \text{N} \\ \text{O} \\ \text{P} \\ \text{Q} \\ \text{R}\end{bmatrix} & \begin{bmatrix}\text{S} \\ \text{T} \\ \text{U} \\ \text{V} \\ \text{W} \\ \text{X}\end{bmatrix} \end{bmatrix}$$ 这些列被模型视为独立的，这意味着`G`和`F`之间的依赖关系无法被学习，但可以实现更高效的批处理。 ```py # In 'run_worker' def print_with_rank(msg): print('[RANK {}]: {}'.format(rank, msg)) from torchtext.datasets import WikiText2 from torchtext.data.utils import get_tokenizer from torchtext.vocab import build_vocab_from_iterator train_iter = WikiText2(split='train') tokenizer = get_tokenizer('basic_english') vocab = build_vocab_from_iterator(map(tokenizer, train_iter), specials=[""]) vocab.set_default_index(vocab[""]) def data_process(raw_text_iter): data = [torch.tensor(vocab(tokenizer(item)), dtype=torch.long) for item in raw_text_iter] return torch.cat(tuple(filter(lambda t: t.numel() > 0, data))) train_iter, val_iter, test_iter = WikiText2() train_data = data_process(train_iter) val_data = data_process(val_iter) test_data = data_process(test_iter) device = torch.device(2 * rank) def batchify(data, bsz, rank, world_size, is_train=False): # Divide the dataset into ``bsz`` parts. nbatch = data.size(0) // bsz # Trim off any extra elements that wouldn't cleanly fit (remainders). data = data.narrow(0, 0, nbatch * bsz) # Evenly divide the data across the ``bsz`` batches. data = data.view(bsz, -1).t().contiguous() # Divide the data across the ranks only for training data. if is_train: data_per_rank = data.size(0) // world_size data = data[rank * data_per_rank : (rank + 1) * data_per_rank] return data.to(device) batch_size = 20 eval_batch_size = 10 train_data = batchify(train_data, batch_size, rank, world_size, True) val_data = batchify(val_data, eval_batch_size, rank, world_size) test_data = batchify(test_data, eval_batch_size, rank, world_size) ``` ### 生成输入和目标序列的函数 `get_batch()`函数为变压器模型生成输入和目标序列。它将源数据细分为长度为`bptt`的块。对于语言建模任务，模型需要以下单词作为`目标`。例如，对于`bptt`值为 2，我们会得到以下两个变量，对于`i` = 0： ![../_images/transformer_input_target.png](img/20ef8681366b44461cf49d1ab98ab8f2.png) 值得注意的是，块沿着维度 0，与变压器模型中的`S`维度一致。批处理维度`N`沿着维度 1。 ```py # In 'run_worker' bptt = 35 def get_batch(source, i): seq_len = min(bptt, len(source) - 1 - i) data = source[i:i+seq_len] target = source[i+1:i+1+seq_len].view(-1) # Need batch dimension first for pipeline parallelism. return data.t(), target ``` ## 模型规模和 Pipe 初始化 为了演示使用管道并行性训练大型 Transformer 模型，我们适当扩展 Transformer 层。我们使用 4096 的嵌入维度，4096 的隐藏大小，16 个注意力头和 8 个总变压器层（`nn.TransformerEncoderLayer`）。这创建了一个具有**~10 亿**参数的模型。 我们需要初始化[RPC 框架](https://pytorch.org/docs/stable/rpc.html)，因为 Pipe 依赖于 RPC 框架通过[RRef](https://pytorch.org/docs/stable/rpc.html#rref)允许未来扩展到跨主机流水线。我们需要使用单个 worker 初始化 RPC 框架，因为我们使用单个进程驱动多个 GPU。 然后，在一个 GPU 上初始化 8 个变压器层，并在另一个 GPU 上初始化 8 个变压器层。一个管道设置在 GPU 0 和 1 之间，另一个设置在 GPU 2 和 3 之间。然后使用`DistributedDataParallel`复制这两个管道。 ```py # In 'run_worker' ntokens = len(vocab) # the size of vocabulary emsize = 4096 # embedding dimension nhid = 4096 # the dimension of the feedforward network model in ``nn.TransformerEncoder`` nlayers = 8 # the number of ``nn.TransformerEncoderLayer`` in ``nn.TransformerEncoder`` nhead = 16 # the number of heads in the Multihead Attention models dropout = 0.2 # the dropout value from torch.distributed import rpc tmpfile = tempfile.NamedTemporaryFile() rpc.init_rpc( name="worker", rank=0, world_size=1, rpc_backend_options=rpc.TensorPipeRpcBackendOptions( init_method="file://{}".format(tmpfile.name), # Specifying _transports and _channels is a workaround and we no longer # will have to specify _transports and _channels for PyTorch # versions >= 1.8.1 _transports=["ibv", "uv"], _channels=["cuda_ipc", "cuda_basic"], ) ) # Number of GPUs for model parallelism. num_gpus = 2 partition_len = ((nlayers - 1) // num_gpus) + 1 # Add encoder in the beginning. tmp_list = [Encoder(ntokens, emsize, dropout).cuda(2 * rank)] module_list = [] # Add all the necessary transformer blocks. for i in range(nlayers): transformer_block = TransformerEncoderLayer(emsize, nhead, nhid, dropout) if i != 0 and i % (partition_len) == 0: module_list.append(nn.Sequential(*tmp_list)) tmp_list = [] device = i // (partition_len) tmp_list.append(transformer_block.to(2 * rank + device)) # Add decoder in the end. tmp_list.append(Decoder(ntokens, emsize).cuda(2 * rank + num_gpus - 1)) module_list.append(nn.Sequential(*tmp_list)) # Need to use 'checkpoint=never' since as of PyTorch 1.8, Pipe checkpointing # doesn't work with DDP. from torch.distributed.pipeline.sync import Pipe chunks = 8 model = Pipe(torch.nn.Sequential( *module_list), chunks = chunks, checkpoint="never") # Initialize process group and wrap model in DDP. from torch.nn.parallel import DistributedDataParallel import torch.distributed as dist os.environ['MASTER_ADDR'] = 'localhost' os.environ['MASTER_PORT'] = '29500' dist.init_process_group( backend="nccl", rank=rank, world_size=world_size) model = DistributedDataParallel(model) def get_total_params(module: torch.nn.Module): total_params = 0 for param in module.parameters(): total_params += param.numel() return total_params print_with_rank('Total parameters in model: {:,}'.format(get_total_params(model))) ``` ## 运行模型 [交叉熵损失](https://pytorch.org/docs/master/nn.html?highlight=crossentropyloss#torch.nn.CrossEntropyLoss)用于跟踪损失，[SGD](https://pytorch.org/docs/master/optim.html?highlight=sgd#torch.optim.SGD)实现随机梯度下降方法作为优化器。初始学习率设置为 5.0。[StepLR](https://pytorch.org/docs/master/optim.html?highlight=steplr#torch.optim.lr_scheduler.StepLR)用于通过 epochs 调整学习率。在训练过程中，我们使用[nn.utils.clip_grad_norm_](https://pytorch.org/docs/master/nn.html?highlight=nn%20utils%20clip_grad_norm#torch.nn.utils.clip_grad_norm_)函数将所有梯度一起缩放，以防止梯度爆炸。 ```py # In 'run_worker' criterion = nn.CrossEntropyLoss() lr = 5.0 # learning rate optimizer = torch.optim.SGD(model.parameters(), lr=lr) scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 1.0, gamma=0.95) import time def train(): model.train() # Turn on the train mode total_loss = 0. start_time = time.time() ntokens = len(vocab) # Train only for 50 batches to keep script execution time low. nbatches = min(50 * bptt, train_data.size(0) - 1) for batch, i in enumerate(range(0, nbatches, bptt)): data, targets = get_batch(train_data, i) optimizer.zero_grad() # Since the Pipe is only within a single host and process the ``RRef`` # returned by forward method is local to this node and can simply # retrieved via ``RRef.local_value()``. output = model(data).local_value() # Need to move targets to the device where the output of the # pipeline resides. loss = criterion(output.view(-1, ntokens), targets.cuda(2 * rank + 1)) loss.backward() torch.nn.utils.clip_grad_norm_(model.parameters(), 0.5) optimizer.step() total_loss += loss.item() log_interval = 10 if batch % log_interval == 0 and batch > 0: cur_loss = total_loss / log_interval elapsed = time.time() - start_time print_with_rank('| epoch {:3d} | {:5d}/{:5d} batches | ' 'lr {:02.2f} | ms/batch {:5.2f} | ' 'loss {:5.2f} | ppl {:8.2f}'.format( epoch, batch, nbatches // bptt, scheduler.get_last_lr()[0], elapsed * 1000 / log_interval, cur_loss, math.exp(cur_loss))) total_loss = 0 start_time = time.time() def evaluate(eval_model, data_source): eval_model.eval() # Turn on the evaluation mode total_loss = 0. ntokens = len(vocab) # Evaluate only for 50 batches to keep script execution time low. nbatches = min(50 * bptt, data_source.size(0) - 1) with torch.no_grad(): for i in range(0, nbatches, bptt): data, targets = get_batch(data_source, i) output = eval_model(data).local_value() output_flat = output.view(-1, ntokens) # Need to move targets to the device where the output of the # pipeline resides. total_loss += len(data) * criterion(output_flat, targets.cuda(2 * rank + 1)).item() return total_loss / (len(data_source) - 1) ``` 循环遍历 epochs。如果验证损失是迄今为止看到的最佳损失，则保存模型。每个 epoch 后调整学习率。 ```py # In 'run_worker' best_val_loss = float("inf") epochs = 3 # The number of epochs best_model = None for epoch in range(1, epochs + 1): epoch_start_time = time.time() train() val_loss = evaluate(model, val_data) print_with_rank('-' * 89) print_with_rank('| end of epoch {:3d} | time: {:5.2f}s | valid loss {:5.2f} | ' 'valid ppl {:8.2f}'.format(epoch, (time.time() - epoch_start_time), val_loss, math.exp(val_loss))) print_with_rank('-' * 89) if val_loss < best_val_loss: best_val_loss = val_loss best_model = model scheduler.step() ``` ## 用测试数据集评估模型 将最佳模型应用于测试数据集以检查结果。 ```py # In 'run_worker' test_loss = evaluate(best_model, test_data) print_with_rank('=' * 89) print_with_rank('| End of training | test loss {:5.2f} | test ppl {:8.2f}'.format( test_loss, math.exp(test_loss))) print_with_rank('=' * 89) # Main execution import torch.multiprocessing as mp if __name__=="__main__": world_size = 2 mp.spawn(run_worker, args=(world_size, ), nprocs=world_size, join=True) ``` ## 输出 ```py [RANK 0]: | epoch 1 | 10/ 50 batches | lr 5.00 | ms/batch 778.97 | loss 43.31 | ppl 6432469059895903232.00 [RANK 1]: | epoch 1 | 10/ 50 batches | lr 5.00 | ms/batch 778.90 | loss 44.50 | ppl 21245447128217366528.00 [RANK 0]: | epoch 1 | 20/ 50 batches | lr 5.00 | ms/batch 699.89 | loss 44.50 | ppl 21176949187407757312.00 [RANK 1]: | epoch 1 | 20/ 50 batches | lr 5.00 | ms/batch 699.87 | loss 44.62 | ppl 23975861229620961280.00 [RANK 0]: | epoch 1 | 30/ 50 batches | lr 5.00 | ms/batch 698.86 | loss 41.62 | ppl 1193312915629888256.00 [RANK 1]: | epoch 1 | 30/ 50 batches | lr 5.00 | ms/batch 698.87 | loss 40.69 | ppl 471605759847546240.00 [RANK 0]: | epoch 1 | 40/ 50 batches | lr 5.00 | ms/batch 698.34 | loss 45.20 | ppl 42812308420836458496.00 [RANK 1]: | epoch 1 | 40/ 50 batches | lr 5.00 | ms/batch 698.33 | loss 45.68 | ppl 68839569686012223488.00 [RANK 1]: ----------------------------------------------------------------------------------------- [RANK 1]: | end of epoch 1 | time: 40.08s | valid loss 0.80 | valid ppl 2.22 [RANK 1]: ----------------------------------------------------------------------------------------- [RANK 0]: ----------------------------------------------------------------------------------------- [RANK 0]: | end of epoch 1 | time: 40.09s | valid loss 0.80 | valid ppl 2.22 [RANK 0]: ----------------------------------------------------------------------------------------- [RANK 0]: | epoch 2 | 10/ 50 batches | lr 4.75 | ms/batch 768.51 | loss 36.34 | ppl 6063529544668166.00 [RANK 1]: | epoch 2 | 10/ 50 batches | lr 4.75 | ms/batch 769.23 | loss 37.41 | ppl 17651211266236086.00 [RANK 0]: | epoch 2 | 20/ 50 batches | lr 4.75 | ms/batch 699.57 | loss 28.97 | ppl 3798441739584.11 [RANK 1]: | epoch 2 | 20/ 50 batches | lr 4.75 | ms/batch 699.56 | loss 29.28 | ppl 5203636967575.47 [RANK 0]: | epoch 2 | 30/ 50 batches | lr 4.75 | ms/batch 699.04 | loss 28.43 | ppl 2212498693571.25 [RANK 1]: | epoch 2 | 30/ 50 batches | lr 4.75 | ms/batch 699.05 | loss 28.33 | ppl 2015144761281.48 [RANK 0]: | epoch 2 | 40/ 50 batches | lr 4.75 | ms/batch 699.10 | loss 23.30 | ppl 13121380184.92 [RANK 1]: | epoch 2 | 40/ 50 batches | lr 4.75 | ms/batch 699.09 | loss 23.41 | ppl 14653799192.87 [RANK 0]: ----------------------------------------------------------------------------------------- [RANK 0]: | end of epoch 2 | time: 39.97s | valid loss 0.24 | valid ppl 1.27 [RANK 0]: ----------------------------------------------------------------------------------------- [RANK 1]: ----------------------------------------------------------------------------------------- [RANK 1]: | end of epoch 2 | time: 39.98s | valid loss 0.24 | valid ppl 1.27 [RANK 1]: ----------------------------------------------------------------------------------------- [RANK 0]: | epoch 3 | 10/ 50 batches | lr 4.51 | ms/batch 769.36 | loss 12.80 | ppl 361681.11 [RANK 1]: | epoch 3 | 10/ 50 batches | lr 4.51 | ms/batch 768.97 | loss 12.57 | ppl 287876.61 [RANK 0]: | epoch 3 | 20/ 50 batches | lr 4.51 | ms/batch 698.27 | loss 12.01 | ppl 164364.60 [RANK 1]: | epoch 3 | 20/ 50 batches | lr 4.51 | ms/batch 698.30 | loss 11.98 | ppl 159095.89 [RANK 0]: | epoch 3 | 30/ 50 batches | lr 4.51 | ms/batch 697.75 | loss 10.90 | ppl 54261.91 [RANK 1]: | epoch 3 | 30/ 50 batches | lr 4.51 | ms/batch 697.72 | loss 10.89 | ppl 53372.39 [RANK 0]: | epoch 3 | 40/ 50 batches | lr 4.51 | ms/batch 699.49 | loss 10.78 | ppl 47948.35 [RANK 1]: | epoch 3 | 40/ 50 batches | lr 4.51 | ms/batch 699.50 | loss 10.79 | ppl 48664.42 [RANK 0]: ----------------------------------------------------------------------------------------- [RANK 0]: | end of epoch 3 | time: 39.96s | valid loss 0.38 | valid ppl 1.46 [RANK 0]: ----------------------------------------------------------------------------------------- [RANK 1]: ----------------------------------------------------------------------------------------- [RANK 1]: | end of epoch 3 | time: 39.96s | valid loss 0.38 | valid ppl 1.46 [RANK 1]: ----------------------------------------------------------------------------------------- [RANK 0]: ========================================================================================= [RANK 0]: | End of training | test loss 0.33 | test ppl 1.39 [RANK 0]: ========================================================================================= [RANK 1]: ========================================================================================= [RANK 1]: | End of training | test loss 0.33 | test ppl 1.39 [RANK 1]: ========================================================================================= ``` **脚本的总运行时间：**（0 分钟 0.000 秒） 下载 Python 源代码：ddp_pipeline.py 下载 Jupyter 笔记本：ddp_pipeline.ipynb [Sphinx-Gallery 生成的图库](https://sphinx-gallery.github.io)