► Developer guides / Multi-GPU distributed training with PyTorch

Multi-GPU distributed training with PyTorch

Author: fchollet
Date created: 2023/06/29
Last modified: 2023/06/29
Description: Guide to multi-GPU training for Keras models with PyTorch.

View in Colab • GitHub source

Introduction

There are generally two ways to distribute computation across multiple devices:

Data parallelism, where a single model gets replicated on multiple devices or multiple machines. Each of them processes different batches of data, then they merge their results. There exist many variants of this setup, that differ in how the different model replicas merge results, in whether they stay in sync at every batch or whether they are more loosely coupled, etc.

Model parallelism, where different parts of a single model run on different devices, processing a single batch of data together. This works best with models that have a naturally-parallel architecture, such as models that feature multiple branches.

This guide focuses on data parallelism, in particular synchronous data parallelism, where the different replicas of the model stay in sync after each batch they process. Synchronicity keeps the model convergence behavior identical to what you would see for single-device training.

Specifically, this guide teaches you how to use PyTorch's DistributedDataParallel module wrapper to train Keras, with minimal changes to your code, on multiple GPUs (typically 2 to 16) installed on a single machine (single host, multi-device training). This is the most common setup for researchers and small-scale industry workflows.

Setup

Let's start by defining the function that creates the model that we will train, and the function that creates the dataset we will train on (MNIST in this case).

import os

os.environ["KERAS_BACKEND"] = "torch"

import torch
import numpy as np
import keras


def get_model():
    # Make a simple convnet with batch normalization and dropout.
    inputs = keras.Input(shape=(28, 28, 1))
    x = keras.layers.Rescaling(1.0 / 255.0)(inputs)
    x = keras.layers.Conv2D(filters=12, kernel_size=3, padding="same", use_bias=False)(
        x
    )
    x = keras.layers.BatchNormalization(scale=False, center=True)(x)
    x = keras.layers.ReLU()(x)
    x = keras.layers.Conv2D(
        filters=24,
        kernel_size=6,
        use_bias=False,
        strides=2,
    )(x)
    x = keras.layers.BatchNormalization(scale=False, center=True)(x)
    x = keras.layers.ReLU()(x)
    x = keras.layers.Conv2D(
        filters=32,
        kernel_size=6,
        padding="same",
        strides=2,
        name="large_k",
    )(x)
    x = keras.layers.BatchNormalization(scale=False, center=True)(x)
    x = keras.layers.ReLU()(x)
    x = keras.layers.GlobalAveragePooling2D()(x)
    x = keras.layers.Dense(256, activation="relu")(x)
    x = keras.layers.Dropout(0.5)(x)
    outputs = keras.layers.Dense(10)(x)
    model = keras.Model(inputs, outputs)
    return model


def get_dataset():
    # Load the data and split it between train and test sets
    (x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()

    # Scale images to the [0, 1] range
    x_train = x_train.astype("float32")
    x_test = x_test.astype("float32")
    # Make sure images have shape (28, 28, 1)
    x_train = np.expand_dims(x_train, -1)
    x_test = np.expand_dims(x_test, -1)
    print("x_train shape:", x_train.shape)

    # Create a TensorDataset
    dataset = torch.utils.data.TensorDataset(
        torch.from_numpy(x_train), torch.from_numpy(y_train)
    )
    return dataset

Next, let's define a simple PyTorch training loop that targets a GPU (note the calls to .cuda()).

def train_model(model, dataloader, num_epochs, optimizer, loss_fn):
    for epoch in range(num_epochs):
        running_loss = 0.0
        running_loss_count = 0
        for batch_idx, (inputs, targets) in enumerate(dataloader):
            inputs = inputs.cuda(non_blocking=True)
            targets = targets.cuda(non_blocking=True)

            # Forward pass
            outputs = model(inputs)
            loss = loss_fn(outputs, targets)

            # Backward and optimize
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()

            running_loss += loss.item()
            running_loss_count += 1

        # Print loss statistics
        print(
            f"Epoch {epoch + 1}/{num_epochs}, "
            f"Loss: {running_loss / running_loss_count}"
        )

Single-host, multi-device synchronous training

In this setup, you have one machine with several GPUs on it (typically 2 to 16). Each device will run a copy of your model (called a replica). For simplicity, in what follows, we'll assume we're dealing with 8 GPUs, at no loss of generality.

How it works

At each step of training:

  • The current batch of data (called global batch) is split into 8 different sub-batches (called local batches). For instance, if the global batch has 512 samples, each of the 8 local batches will have 64 samples.
  • Each of the 8 replicas independently processes a local batch: they run a forward pass, then a backward pass, outputting the gradient of the weights with respect to the loss of the model on the local batch.
  • The weight updates originating from local gradients are efficiently merged across the 8 replicas. Because this is done at the end of every step, the replicas always stay in sync.

In practice, the process of synchronously updating the weights of the model replicas is handled at the level of each individual weight variable. This is done through a mirrored variable object.

How to use it

To do single-host, multi-device synchronous training with a Keras model, you would use the torch.nn.parallel.DistributedDataParallel module wrapper. Here's how it works:

  • We use torch.multiprocessing.start_processes to start multiple Python processes, one per device. Each process will run the per_device_launch_fn function.
  • The per_device_launch_fn function does the following: - It uses torch.distributed.init_process_group and torch.cuda.set_device to configure the device to be used for that process. - It uses torch.utils.data.distributed.DistributedSampler and torch.utils.data.DataLoader to turn our data into a distributed data loader. - It also uses torch.nn.parallel.DistributedDataParallel to turn our model into a distributed PyTorch module. - It then calls the train_model function.
  • The train_model function will then run in each process, with the model using a separate device in each process.

Here's the flow, where each step is split into its own utility function:

# Config
num_gpu = torch.cuda.device_count()
num_epochs = 2
batch_size = 64
print(f"Running on {num_gpu} GPUs")


def setup_device(current_gpu_index, num_gpus):
    # Device setup
    os.environ["MASTER_ADDR"] = "localhost"
    os.environ["MASTER_PORT"] = "56492"
    device = torch.device("cuda:{}".format(current_gpu_index))
    torch.distributed.init_process_group(
        backend="nccl",
        init_method="env://",
        world_size=num_gpus,
        rank=current_gpu_index,
    )
    torch.cuda.set_device(device)


def cleanup():
    torch.distributed.destroy_process_group()


def prepare_dataloader(dataset, current_gpu_index, num_gpus, batch_size):
    sampler = torch.utils.data.distributed.DistributedSampler(
        dataset,
        num_replicas=num_gpus,
        rank=current_gpu_index,
        shuffle=False,
    )
    dataloader = torch.utils.data.DataLoader(
        dataset,
        sampler=sampler,
        batch_size=batch_size,
        shuffle=False,
    )
    return dataloader


def per_device_launch_fn(current_gpu_index, num_gpu):
    # Setup the process groups
    setup_device(current_gpu_index, num_gpu)

    dataset = get_dataset()
    model = get_model()

    # prepare the dataloader
    dataloader = prepare_dataloader(dataset, current_gpu_index, num_gpu, batch_size)

    # Instantiate the torch optimizer
    optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)

    # Instantiate the torch loss function
    loss_fn = torch.nn.CrossEntropyLoss()

    # Put model on device
    model = model.to(current_gpu_index)
    ddp_model = torch.nn.parallel.DistributedDataParallel(
        model, device_ids=[current_gpu_index], output_device=current_gpu_index
    )

    train_model(ddp_model, dataloader, num_epochs, optimizer, loss_fn)

    cleanup()

Running on 0 GPUs

/opt/conda/envs/keras-torch/lib/python3.10/site-packages/torch/cuda/__init__.py:611: UserWarning: Can't initialize NVML
  warnings.warn("Can't initialize NVML")

Time to start multiple processes:

if __name__ == "__main__":
    # We use the "fork" method rather than "spawn" to support notebooks
    torch.multiprocessing.start_processes(
        per_device_launch_fn,
        args=(num_gpu,),
        nprocs=num_gpu,
        join=True,
        start_method="fork",
    )

That's it!

Multi-GPU distributed training with PyTorch
◆ Introduction
◆ Setup
◆ Single-host, multi-device synchronous training