PyTorch 2.2 中文官方教程（十九）（2）https://developer.aliyun.com/article/1482622
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：
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。
# 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=["<unk>"]) vocab.set_default_index(vocab["<unk>"]) 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：
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。
# 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，因为 Pipe 依赖于 RPC 框架通过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。
# 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
交叉熵损失用于跟踪损失，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_函数将所有梯度一起缩放，以防止梯度爆炸。
# 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。
# 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。
# 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
[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
使用 Join 上下文管理器进行不均匀输入的分布式训练
原文：
pytorch.org/tutorials/advanced/generic_join.html译者：飞龙
作者：Andrew Gu
注意
在github中查看并编辑此教程。
注意
Join在 PyTorch 1.10 中作为原型功能引入。此 API 可能会更改。
在本教程中，您将看到：
- Join上下文管理器的概述。
- 如何使用
DistributedDataParallel与上下文管理器的示例。 - 如何使用上下文管理器与
DistributedDataParallel和ZeroRedundancyOptimizer的示例。 - 将关键字参数传递给上下文管理器的示例。
- 深入了解Join上下文管理器的工作原理。
- 一个示例，展示如何使一个玩具类与上下文管理器兼容。
PyTorch 2.2 中文官方教程（十九）（4）https://developer.aliyun.com/article/1482627
