• Tutorials >
  • Training Transformer models using Pipeline Parallelism
Shortcuts

Training Transformer models using Pipeline Parallelism

Author: Pritam Damania

This tutorial demonstrates how to train a large Transformer model across multiple GPUs using pipeline parallelism. This tutorial is an extension of the Sequence-to-Sequence Modeling with nn.Transformer and TorchText tutorial and scales up the same model to demonstrate how pipeline parallelism can be used to train Transformer models.

Prerequisites:

Define the model

In this tutorial, we will split a Transformer model across two GPUs and use pipeline parallelism to train the model. The model is exactly the same model used in the Sequence-to-Sequence Modeling with nn.Transformer and TorchText tutorial, but is split into two stages. The largest number of parameters belong to the nn.TransformerEncoder layer. The nn.TransformerEncoder itself consists of nlayers of nn.TransformerEncoderLayer. As a result, our focus is on nn.TransformerEncoder and we split the model such that half of the nn.TransformerEncoderLayer are on one GPU and the other half are on another. To do this, we pull out the Encoder and Decoder sections into separate modules and then build an nn.Sequential representing the original Transformer module.

import sys
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
import tempfile
from torch.nn import TransformerEncoder, TransformerEncoderLayer

if sys.platform == 'win32':
    print('Windows platform is not supported for pipeline parallelism')
    sys.exit(0)
if torch.cuda.device_count() < 2:
    print('Need at least two 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)

PositionalEncoding module injects some information about the relative or absolute position of the tokens in the sequence. The positional encodings have the same dimension as the embeddings so that the two can be summed. Here, we use sine and cosine functions of different frequencies.

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.register_buffer('pe', pe)

    def forward(self, x):
        x = x + self.pe[:x.size(0), :]
        return self.dropout(x)

Load and batch data

The training process uses Wikitext-2 dataset from torchtext. To access torchtext datasets, please install torchdata following instructions at https://github.com/pytorch/data.

The vocab object is built based on the train dataset and is used to numericalize tokens into tensors. Starting from sequential data, the batchify() function arranges the dataset into columns, trimming off any tokens remaining after the data has been divided into batches of size batch_size. For instance, with the alphabet as the sequence (total length of 26) and a batch size of 4, we would divide the alphabet into 4 sequences of length 6:

\[\begin{bmatrix} \text{A} & \text{B} & \text{C} & \ldots & \text{X} & \text{Y} & \text{Z} \end{bmatrix} \Rightarrow \begin{bmatrix} \begin{bmatrix}\text{A} \\ \text{B} \\ \text{C} \\ \text{D} \\ \text{E} \\ \text{F}\end{bmatrix} & \begin{bmatrix}\text{G} \\ \text{H} \\ \text{I} \\ \text{J} \\ \text{K} \\ \text{L}\end{bmatrix} & \begin{bmatrix}\text{M} \\ \text{N} \\ \text{O} \\ \text{P} \\ \text{Q} \\ \text{R}\end{bmatrix} & \begin{bmatrix}\text{S} \\ \text{T} \\ \text{U} \\ \text{V} \\ \text{W} \\ \text{X}\end{bmatrix} \end{bmatrix}\]

These columns are treated as independent by the model, which means that the dependence of G and F can not be learned, but allows more efficient batch processing.

import torch
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("cuda")

def batchify(data, bsz):
    # 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()
    return data.to(device)

batch_size = 20
eval_batch_size = 10
train_data = batchify(train_data, batch_size)
val_data = batchify(val_data, eval_batch_size)
test_data = batchify(test_data, eval_batch_size)

Functions to generate input and target sequence

get_batch() function generates the input and target sequence for the transformer model. It subdivides the source data into chunks of length bptt. For the language modeling task, the model needs the following words as Target. For example, with a bptt value of 2, we’d get the following two Variables for i = 0:

../_images/transformer_input_target.png

It should be noted that the chunks are along dimension 0, consistent with the S dimension in the Transformer model. The batch dimension N is along dimension 1.

bptt = 25
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

Model scale and Pipe initialization

To demonstrate training large Transformer models using pipeline parallelism, we scale up the Transformer layers appropriately. We use an embedding dimension of 4096, hidden size of 4096, 16 attention heads and 12 total transformer layers (nn.TransformerEncoderLayer). This creates a model with ~1.4 billion parameters.

We need to initialize the RPC Framework since Pipe depends on the RPC framework via RRef which allows for future expansion to cross host pipelining. We need to initialize the RPC framework with only a single worker since we’re using a single process to drive multiple GPUs.

The pipeline is then initialized with 8 transformer layers on one GPU and 8 transformer layers on the other GPU.

Note

For efficiency purposes we ensure that the nn.Sequential passed to Pipe only consists of two elements (corresponding to two GPUs), this allows the Pipe to work with only two partitions and avoid any cross-partition overheads.

ntokens = len(vocab) # the size of vocabulary
emsize = 4096 # embedding dimension
nhid = 4096 # the dimension of the feedforward network model in ``nn.TransformerEncoder``
nlayers = 12 # 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"],
    )
)

num_gpus = 2
partition_len = ((nlayers - 1) // num_gpus) + 1

# Add encoder in the beginning.
tmp_list = [Encoder(ntokens, emsize, dropout).cuda(0)]
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(device))

# Add decoder in the end.
tmp_list.append(Decoder(ntokens, emsize).cuda(num_gpus - 1))
module_list.append(nn.Sequential(*tmp_list))

from torch.distributed.pipeline.sync import Pipe

# Build the pipeline.
chunks = 8
model = Pipe(torch.nn.Sequential(*module_list), chunks = chunks)


def get_total_params(module: torch.nn.Module):
    total_params = 0
    for param in module.parameters():
        total_params += param.numel()
    return total_params

print ('Total parameters in model: {:,}'.format(get_total_params(model)))

Run the model

CrossEntropyLoss is applied to track the loss and SGD implements stochastic gradient descent method as the optimizer. The initial learning rate is set to 5.0. StepLR is applied to adjust the learn rate through epochs. During the training, we use nn.utils.clip_grad_norm_ function to scale all the gradient together to prevent exploding.

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(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('| 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_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(1)).item()
    return total_loss / (len(data_source) - 1)

Loop over epochs. Save the model if the validation loss is the best we’ve seen so far. Adjust the learning rate after each epoch.

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('-' * 89)
    print('| 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('-' * 89)

    if val_loss < best_val_loss:
        best_val_loss = val_loss
        best_model = model

    scheduler.step()

Evaluate the model with the test dataset

Apply the best model to check the result with the test dataset.

test_loss = evaluate(best_model, test_data)
print('=' * 89)
print('| End of training | test loss {:5.2f} | test ppl {:8.2f}'.format(
    test_loss, math.exp(test_loss)))
print('=' * 89)

Total running time of the script: ( 0 minutes 0.000 seconds)

Gallery generated by Sphinx-Gallery

Docs

Access comprehensive developer documentation for PyTorch

View Docs

Tutorials

Get in-depth tutorials for beginners and advanced developers

View Tutorials

Resources

Find development resources and get your questions answered

View Resources