Example of Starting PyTorch DDP Training Based on a Training Job

Creating a Training Job

• Use PyTorch preset images and run the mp.spawn command.

  Set Boot Mode to Preset image, select a PyTorch preset image below, set Code Directory to the OBS path where the code folder is stored, and choose the main.py file as the boot file.

  Figure 1 Creating a training job
  

  If you select a resource flavor with multiple cards on one node, specify the hyperparameters world_size and rank when creating a training job. If you select a resource flavor with multiple nodes (more than one compute node in a training job), you do not need to set these hyperparameters. world_size and rank are automatically injected by ModelArts.

  Figure 2 Hyperparameters
  
• Use the custom image and run the torch.distributed.launch command.
  Set Boot Mode to Custom image, select a custom image, and set Code Directory to the OBS path where the code folder is stored. The boot command is as follows:

  bash ${MA_JOB_DIR}/code/torchlaunch.sh

  Figure 3 Creating a training job
  
• Use the custom image and run the torch.distributed.run command.
  Set Boot Mode to Custom image, select a custom image, and set Code Directory to the OBS path where the code folder is stored. The boot command is as follows:

  bash ${MA_JOB_DIR}/code/torchrun.sh

  Figure 4 Creating a training job
  

Code Examples

Upload the following files to an OBS bucket:

code                             # Root directory of the code
 └─torch_ddp.py                # Code file for PyTorch DDP training
 └─main.py                     # Boot file for starting training using the PyTorch preset image and the mp.spawn command
 └─torchlaunch.sh              # Boot file for starting training using the custom image and the torch.distributed.launch command
 └─torchrun.sh                 # Boot file for starting training using the custom image and the torch.distributed.run command

torch_ddp.py

import os
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

# Start training by running mp.spawn.
def init_from_arg(local_rank, base_rank, world_size, init_method):
    rank = base_rank + local_rank
    dist.init_process_group("nccl", rank=rank, init_method=init_method, world_size=world_size)
    ddp_train(local_rank)

# Start training by running torch.distributed.launch or torch.distributed.run.
def init_from_env():
    dist.init_process_group(backend='nccl', init_method='env://')
    local_rank=int(os.environ["LOCAL_RANK"])
    ddp_train(local_rank)

def cleanup():
    dist.destroy_process_group()

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 ddp_train(device_id):
    # create model and move it to GPU with id rank
    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()
    cleanup()

if __name__ == "__main__":
    init_from_env()

main.py

import argparse
import torch
import torch.multiprocessing as mp

parser = argparse.ArgumentParser(description='ddp demo args')
parser.add_argument('--world_size', type=int, required=True)
parser.add_argument('--rank', type=int, required=True)
parser.add_argument('--init_method', type=str, required=True)
args, unknown = parser.parse_known_args()

if __name__ == "__main__":
    n_gpus = torch.cuda.device_count()
    world_size = n_gpus * args.world_size
    base_rank = n_gpus * args.rank
    # Call the start function in the DDP sample code.
    from torch_ddp import init_from_arg
    mp.spawn(init_from_arg,
        args=(base_rank, world_size, args.init_method),
        nprocs=n_gpus,
        join=True)

#!/bin/bash
# Default system environment variables. Do not modify them.
MASTER_HOST="$VC_WORKER_HOSTS"
MASTER_ADDR="${VC_WORKER_HOSTS%%,*}"
MASTER_PORT="6060"
JOB_ID="1234"
NNODES="$MA_NUM_HOSTS"
NODE_RANK="$VC_TASK_INDEX"
NGPUS_PER_NODE="$MA_NUM_GPUS"

# Custom environment variables to specify the Python script and parameters.
PYTHON_SCRIPT=${MA_JOB_DIR}/code/torch_ddp.py
PYTHON_ARGS=""

CMD="python -m torch.distributed.launch \
    --nnodes=$NNODES \
    --node_rank=$NODE_RANK \
    --nproc_per_node=$NGPUS_PER_NODE \
    --master_addr $MASTER_ADDR \
    --master_port=$MASTER_PORT \
    --use_env \
    $PYTHON_SCRIPT \
    $PYTHON_ARGS
"
echo $CMD
$CMD

torchrun.sh

In PyTorch 2.1, you must set rdzv_backend to static: --rdzv_backend=static.

#!/bin/bash
# Default system environment variables. Do not modify them.
MASTER_HOST="$VC_WORKER_HOSTS"
MASTER_ADDR="${VC_WORKER_HOSTS%%,*}"
MASTER_PORT="6060"
JOB_ID="1234"
NNODES="$MA_NUM_HOSTS"
NODE_RANK="$VC_TASK_INDEX"
NGPUS_PER_NODE="$MA_NUM_GPUS"

# Custom environment variables to specify the Python script and parameters.
PYTHON_SCRIPT=${MA_JOB_DIR}/code/torch_ddp.py
PYTHON_ARGS=""

if [[ $NODE_RANK == 0 ]]; then
    EXT_ARGS="--rdzv_conf=is_host=1"
else
    EXT_ARGS=""
fi

CMD="python -m torch.distributed.run \
    --nnodes=$NNODES \
    --node_rank=$NODE_RANK \
    $EXT_ARGS \
    --nproc_per_node=$NGPUS_PER_NODE \
    --rdzv_id=$JOB_ID \
    --rdzv_backend=c10d \
    --rdzv_endpoint=$MASTER_ADDR:$MASTER_PORT \
    $PYTHON_SCRIPT \
    $PYTHON_ARGS
    "
echo $CMD
$CMD