Introduction || What is DDP || Single-Node Multi-GPU Training || Fault Tolerance || Multi-Node training || minGPT Training

使用 DDP 进行多 GPU 训练

Authors: Suraj Subramanian

What you will learn

• How to migrate a single-GPU training script to multi-GPU via DDP

• Setting up the distributed process group

• Saving and loading models in a distributed setup

View the code used in this tutorial on GitHub

Prerequisites

• High-level overview of how DDP works

• A machine with multiple GPUs (this tutorial uses an AWS p3.8xlarge instance)

• PyTorch installed with CUDA

Follow along with the video below or on youtube.

In the previous tutorial, we got a high-level overview of how DDP works; now we see how to use DDP in code. In this tutorial, we start with a single-GPU training script and migrate that to running it on 4 GPUs on a single node. Along the way, we will talk through important concepts in distributed training while implementing them in our code.

Note

If your model contains any BatchNorm layers, it needs to be converted to SyncBatchNorm to sync the running stats of BatchNorm layers across replicas.

Use the helper function torch.nn.SyncBatchNorm.convert_sync_batchnorm(model) to convert all BatchNorm layers in the model to SyncBatchNorm.

Diff for single_gpu.py v/s multigpu.py

These are the changes you typically make to a single-GPU training script to enable DDP.

Imports

• torch.multiprocessing is a PyTorch wrapper around Python’s native multiprocessing

• The distributed process group contains all the processes that can communicate and synchronize with each other.

import torch
import torch.nn.functional as F
from utils import MyTrainDataset

+ import torch.multiprocessing as mp
+ from torch.utils.data.distributed import DistributedSampler
+ from torch.nn.parallel import DistributedDataParallel as DDP
+ from torch.distributed import init_process_group, destroy_process_group
+ import os

Constructing the process group

• First, before initializing the group process, call set_device, which sets the default GPU for each process. This is important to prevent hangs or excessive memory utilization on GPU:0

• The process group can be initialized by TCP (default) or from a shared file-system. Read more on process group initialization

• init_process_group initializes the distributed process group.

• Read more about choosing a DDP backend

+ def ddp_setup(rank: int, world_size: int):
+   """
+   Args:
+       rank: Unique identifier of each process
+      world_size: Total number of processes
+   """
+   os.environ["MASTER_ADDR"] = "localhost"
+   os.environ["MASTER_PORT"] = "12355"
+   torch.cuda.set_device(rank)
+   init_process_group(backend="nccl", rank=rank, world_size=world_size)

Constructing the DDP model

- self.model = model.to(gpu_id)
+ self.model = DDP(model, device_ids=[gpu_id])

Distributing input data

• DistributedSampler chunks the input data across all distributed processes.

• Each process will receive an input batch of 32 samples; the effective batch size is 32 * nprocs, or 128 when using 4 GPUs.

train_data = torch.utils.data.DataLoader(
    dataset=train_dataset,
    batch_size=32,
-   shuffle=True,
+   shuffle=False,
+   sampler=DistributedSampler(train_dataset),
)

• Calling the set_epoch() method on the DistributedSampler at the beginning of each epoch is necessary to make shuffling work properly across multiple epochs. Otherwise, the same ordering will be used in each epoch.

def _run_epoch(self, epoch):
    b_sz = len(next(iter(self.train_data))[0])
+   self.train_data.sampler.set_epoch(epoch)
    for source, targets in self.train_data:
      ...
      self._run_batch(source, targets)

Saving model checkpoints

• We only need to save model checkpoints from one process. Without this condition, each process would save its copy of the identical mode. Read more on saving and loading models with DDP here

- ckp = self.model.state_dict()
+ ckp = self.model.module.state_dict()
...
...
- if epoch % self.save_every == 0:
+ if self.gpu_id == 0 and epoch % self.save_every == 0:
  self._save_checkpoint(epoch)

Warning

Collective calls are functions that run on all the distributed processes, and they are used to gather certain states or values to a specific process. Collective calls require all ranks to run the collective code. In this example, _save_checkpoint should not have any collective calls because it is only run on the rank:0 process. If you need to make any collective calls, it should be before the if self.gpu_id == 0 check.

Running the distributed training job

• Include new arguments rank (replacing device) and world_size.

• rank is auto-allocated by DDP when calling mp.spawn.

• world_size is the number of processes across the training job. For GPU training, this corresponds to the number of GPUs in use, and each process works on a dedicated GPU.

- def main(device, total_epochs, save_every):
+ def main(rank, world_size, total_epochs, save_every):
+  ddp_setup(rank, world_size)
   dataset, model, optimizer = load_train_objs()
   train_data = prepare_dataloader(dataset, batch_size=32)
-  trainer = Trainer(model, train_data, optimizer, device, save_every)
+  trainer = Trainer(model, train_data, optimizer, rank, save_every)
   trainer.train(total_epochs)
+  destroy_process_group()

if __name__ == "__main__":
   import sys
   total_epochs = int(sys.argv[1])
   save_every = int(sys.argv[2])
-  device = 0      # shorthand for cuda:0
-  main(device, total_epochs, save_every)
+  world_size = torch.cuda.device_count()
+  mp.spawn(main, args=(world_size, total_epochs, save_every,), nprocs=world_size)

Further Reading

• Fault Tolerant distributed training (next tutorial in this series)

• Intro to DDP (previous tutorial in this series)

• Getting Started with DDP

• Process Group initialization