Getting Started with Distributed Data Parallel

Author: Shen Li

Edited by: Joe Zhu

Note

View and edit this tutorial in github.

Prerequisites:

• PyTorch Distributed Overview

• DistributedDataParallel API documents

• DistributedDataParallel notes

DistributedDataParallel (DDP) implements data parallelism at the module level which can run across multiple machines. Applications using DDP should spawn multiple processes and create a single DDP instance per process. DDP uses collective communications in the torch.distributed package to synchronize gradients and buffers. More specifically, DDP registers an autograd hook for each parameter given by model.parameters() and the hook will fire when the corresponding gradient is computed in the backward pass. Then DDP uses that signal to trigger gradient synchronization across processes. Please refer to DDP design note for more details.

The recommended way to use DDP is to spawn one process for each model replica, where a model replica can span multiple devices. DDP processes can be placed on the same machine or across machines, but GPU devices cannot be shared across processes. This tutorial starts from a basic DDP use case and then demonstrates more advanced use cases including checkpointing models and combining DDP with model parallel.

Note

The code in this tutorial runs on an 8-GPU server, but it can be easily generalized to other environments.

Comparison between DataParallel and DistributedDataParallel

Before we dive in, let’s clarify why, despite the added complexity, you would consider using DistributedDataParallel over DataParallel:

• First, DataParallel is single-process, multi-thread, and only works on a single machine, while DistributedDataParallel is multi-process and works for both single- and multi- machine training. DataParallel is usually slower than DistributedDataParallel even on a single machine due to GIL contention across threads, per-iteration replicated model, and additional overhead introduced by scattering inputs and gathering outputs.

• Recall from the prior tutorial that if your model is too large to fit on a single GPU, you must use model parallel to split it across multiple GPUs. DistributedDataParallel works with model parallel; DataParallel does not at this time. When DDP is combined with model parallel, each DDP process would use model parallel, and all processes collectively would use data parallel.

• If your model needs to span multiple machines or if your use case does not fit into data parallelism paradigm, please see the RPC API for more generic distributed training support.

Basic Use Case

To create a DDP module, you must first set up process groups properly. More details can be found in Writing Distributed Applications with PyTorch.

import os
import sys
import tempfile
import torch
import torch.distributed as dist
import torch.nn as nn
import torch.optim as optim
import torch.multiprocessing as mp

from torch.nn.parallel import DistributedDataParallel as DDP

# On Windows platform, the torch.distributed package only
# supports Gloo backend, FileStore and TcpStore.
# For FileStore, set init_method parameter in init_process_group
# to a local file. Example as follow:
# init_method="file:///f:/libtmp/some_file"
# dist.init_process_group(
#    "gloo",
#    rank=rank,
#    init_method=init_method,
#    world_size=world_size)
# For TcpStore, same way as on Linux.

def setup(rank, world_size):
    os.environ['MASTER_ADDR'] = 'localhost'
    os.environ['MASTER_PORT'] = '12355'

    # initialize the process group
    dist.init_process_group("gloo", rank=rank, world_size=world_size)

def cleanup():
    dist.destroy_process_group()

Now, let’s create a toy module, wrap it with DDP, and feed it some dummy input data. Please note, as DDP broadcasts model states from rank 0 process to all other processes in the DDP constructor, you do not need to worry about different DDP processes starting from different initial model parameter values.

class ToyModel(nn.Module):
    def __init__(self):
        super(ToyModel, self).__init__()
        self.net1 = nn.Linear(10, 10)
        self.relu = nn.ReLU()
        self.net2 = nn.Linear(10, 5)

    def forward(self, x):
        return self.net2(self.relu(self.net1(x)))


def demo_basic(rank, world_size):
    print(f"Running basic DDP example on rank {rank}.")
    setup(rank, world_size)

    # create model and move it to GPU with id rank
    model = ToyModel().to(rank)
    ddp_model = DDP(model, device_ids=[rank])

    loss_fn = nn.MSELoss()
    optimizer = optim.SGD(ddp_model.parameters(), lr=0.001)

    optimizer.zero_grad()
    outputs = ddp_model(torch.randn(20, 10))
    labels = torch.randn(20, 5).to(rank)
    loss_fn(outputs, labels).backward()
    optimizer.step()

    cleanup()


def run_demo(demo_fn, world_size):
    mp.spawn(demo_fn,
             args=(world_size,),
             nprocs=world_size,
             join=True)

As you can see, DDP wraps lower-level distributed communication details and provides a clean API as if it were a local model. Gradient synchronization communications take place during the backward pass and overlap with the backward computation. When the backward() returns, param.grad already contains the synchronized gradient tensor. For basic use cases, DDP only requires a few more LoCs to set up the process group. When applying DDP to more advanced use cases, some caveats require caution.

Skewed Processing Speeds

In DDP, the constructor, the forward pass, and the backward pass are distributed synchronization points. Different processes are expected to launch the same number of synchronizations and reach these synchronization points in the same order and enter each synchronization point at roughly the same time. Otherwise, fast processes might arrive early and timeout while waiting for stragglers. Hence, users are responsible for balancing workload distributions across processes. Sometimes, skewed processing speeds are inevitable due to, e.g., network delays, resource contentions, or unpredictable workload spikes. To avoid timeouts in these situations, make sure that you pass a sufficiently large timeout value when calling init_process_group.

Save and Load Checkpoints

It’s common to use torch.save and torch.load to checkpoint modules during training and recover from checkpoints. See SAVING AND LOADING MODELS for more details. When using DDP, one optimization is to save the model in only one process and then load it to all processes, reducing write overhead. This is correct because all processes start from the same parameters and gradients are synchronized in backward passes, and hence optimizers should keep setting parameters to the same values. If you use this optimization, make sure no process starts loading before the saving is finished. Additionally, when loading the module, you need to provide an appropriate map_location argument to prevent a process from stepping into others’ devices. If map_location is missing, torch.load will first load the module to CPU and then copy each parameter to where it was saved, which would result in all processes on the same machine using the same set of devices. For more advanced failure recovery and elasticity support, please refer to TorchElastic.

def demo_checkpoint(rank, world_size):
    print(f"Running DDP checkpoint example on rank {rank}.")
    setup(rank, world_size)

    model = ToyModel().to(rank)
    ddp_model = DDP(model, device_ids=[rank])


    CHECKPOINT_PATH = tempfile.gettempdir() + "/model.checkpoint"
    if rank == 0:
        # All processes should see same parameters as they all start from same
        # random parameters and gradients are synchronized in backward passes.
        # Therefore, saving it in one process is sufficient.
        torch.save(ddp_model.state_dict(), CHECKPOINT_PATH)

    # Use a barrier() to make sure that process 1 loads the model after process
    # 0 saves it.
    dist.barrier()
    # configure map_location properly
    map_location = {'cuda:%d' % 0: 'cuda:%d' % rank}
    ddp_model.load_state_dict(
        torch.load(CHECKPOINT_PATH, map_location=map_location))

    loss_fn = nn.MSELoss()
    optimizer = optim.SGD(ddp_model.parameters(), lr=0.001)

    optimizer.zero_grad()
    outputs = ddp_model(torch.randn(20, 10))
    labels = torch.randn(20, 5).to(rank)

    loss_fn(outputs, labels).backward()
    optimizer.step()

    # Not necessary to use a dist.barrier() to guard the file deletion below
    # as the AllReduce ops in the backward pass of DDP already served as
    # a synchronization.

    if rank == 0:
        os.remove(CHECKPOINT_PATH)

    cleanup()

Combining DDP with Model Parallelism

DDP also works with multi-GPU models. DDP wrapping multi-GPU models is especially helpful when training large models with a huge amount of data.

class ToyMpModel(nn.Module):
    def __init__(self, dev0, dev1):
        super(ToyMpModel, self).__init__()
        self.dev0 = dev0
        self.dev1 = dev1
        self.net1 = torch.nn.Linear(10, 10).to(dev0)
        self.relu = torch.nn.ReLU()
        self.net2 = torch.nn.Linear(10, 5).to(dev1)

    def forward(self, x):
        x = x.to(self.dev0)
        x = self.relu(self.net1(x))
        x = x.to(self.dev1)
        return self.net2(x)

When passing a multi-GPU model to DDP, device_ids and output_device must NOT be set. Input and output data will be placed in proper devices by either the application or the model forward() method.

def demo_model_parallel(rank, world_size):
    print(f"Running DDP with model parallel example on rank {rank}.")
    setup(rank, world_size)

    # setup mp_model and devices for this process
    dev0 = rank * 2
    dev1 = rank * 2 + 1
    mp_model = ToyMpModel(dev0, dev1)
    ddp_mp_model = DDP(mp_model)

    loss_fn = nn.MSELoss()
    optimizer = optim.SGD(ddp_mp_model.parameters(), lr=0.001)

    optimizer.zero_grad()
    # outputs will be on dev1
    outputs = ddp_mp_model(torch.randn(20, 10))
    labels = torch.randn(20, 5).to(dev1)
    loss_fn(outputs, labels).backward()
    optimizer.step()

    cleanup()


if __name__ == "__main__":
    n_gpus = torch.cuda.device_count()
    assert n_gpus >= 2, f"Requires at least 2 GPUs to run, but got {n_gpus}"
    world_size = n_gpus
    run_demo(demo_basic, world_size)
    run_demo(demo_checkpoint, world_size)
    world_size = n_gpus//2
    run_demo(demo_model_parallel, world_size)

Initialize DDP with torch.distributed.run/torchrun

We can leverage PyTorch Elastic to simplify the DDP code and initialize the job more easily. Let’s still use the Toymodel example and create a file named elastic_ddp.py.

import torch
import torch.distributed as dist
import torch.nn as nn
import torch.optim as optim

from torch.nn.parallel import DistributedDataParallel as DDP

class ToyModel(nn.Module):
    def __init__(self):
        super(ToyModel, self).__init__()
        self.net1 = nn.Linear(10, 10)
        self.relu = nn.ReLU()
        self.net2 = nn.Linear(10, 5)

    def forward(self, x):
        return self.net2(self.relu(self.net1(x)))


def demo_basic():
    dist.init_process_group("nccl")
    rank = dist.get_rank()
    print(f"Start running basic DDP example on rank {rank}.")

    # create model and move it to GPU with id rank
    device_id = rank % torch.cuda.device_count()
    model = ToyModel().to(device_id)
    ddp_model = DDP(model, device_ids=[device_id])

    loss_fn = nn.MSELoss()
    optimizer = optim.SGD(ddp_model.parameters(), lr=0.001)

    optimizer.zero_grad()
    outputs = ddp_model(torch.randn(20, 10))
    labels = torch.randn(20, 5).to(device_id)
    loss_fn(outputs, labels).backward()
    optimizer.step()

if __name__ == "__main__":
    demo_basic()

One can then run a `torch elastic/torchrun<https://pytorch.org/docs/stable/elastic/quickstart.html>`__ command on all nodes to initialize the DDP job created above:

torchrun --nnodes=2 --nproc_per_node=8 --rdzv_id=100 --rdzv_backend=c10d --rdzv_endpoint=$MASTER_ADDR:29400 elastic_ddp.py

We are running the DDP script on two hosts, and each host we run with 8 processes, aka, we are running it on 16 GPUs. Note that $MASTER_ADDR must be the same across all nodes.

Here torchrun will launch 8 process and invoke elastic_ddp.py on each process on the node it is launched on, but user also needs to apply cluster management tools like slurm to actually run this command on 2 nodes.

For example, on a SLURM enabled cluster, we can write a script to run the command above and set MASTER_ADDR as:

export MASTER_ADDR=$(scontrol show hostname ${SLURM_NODELIST} | head -n 1)

Then we can just run this script using the SLURM command: srun --nodes=2 ./torchrun_script.sh. Of course, this is just an example; you can choose your own cluster scheduling tools to initiate the torchrun job.

For more information about Elastic run, one can check this quick start document to learn more.