import torchimport torch.nn.functional as Ffrom torch.utils.data import Dataset, DataLoaderfrom sklearn.datasets import load_winefrom sklearn.model_selection import train_test_splitfrom sklearn.preprocessing import StandardScalerimport numpy as np
class WineDataset(Dataset): def __init__(self, data, targets):        self.data = data        self.targets = targets
 def __len__(self): return len(self.data)
 def __getitem__(self, idx): return torch.tensor(self.data[idx], dtype=torch.float), torch.tensor(self.targets[idx], dtype=torch.long)
class SimpleNN(torch.nn.Module): def __init__(self): super(SimpleNN, self).__init__()        self.fc1 = torch.nn.Linear(13, 64)        self.fc2 = torch.nn.Linear(64, 3)
 def forward(self, x):        x = F.relu(self.fc1(x))        x = self.fc2(x) return x
class Trainer(): def __init__(self, model, train_data, optimizer, gpu_id, save_every):        self.model = model        self.train_data = train_data        self.optimizer = optimizer        self.gpu_id = gpu_id        self.save_every = save_every        self.losses = []
 def _run_batch(self, source, targets):        self.optimizer.zero_grad()        output = self.model(source)        loss = F.cross_entropy(output, targets)        loss.backward()        self.optimizer.step() return loss.item()
 def _run_epoch(self, epoch):        total_loss = 0.0        num_batches = len(self.train_data) for source, targets in self.train_data:            source = source.to(self.gpu_id)            targets = targets.to(self.gpu_id)            loss = self._run_batch(source, targets)            total_loss += loss
        avg_loss = total_loss / num_batches        self.losses.append(avg_loss) print(f"Epoch {epoch}, Loss: {avg_loss:.4f}")
 def _save_checkpoint(self, epoch):        checkpoint = self.model.state_dict()        PATH = f"model_{epoch}.pt"        torch.save(checkpoint, PATH) print(f"Epoch {epoch} | Model saved to {PATH}")
 def train(self, max_epochs):        self.model.train() for epoch in range(max_epochs):            self._run_epoch(epoch) if epoch % self.save_every == 0:                self._save_checkpoint(epoch)
def load_train_objs():    wine_data = load_wine()    X = wine_data.data    y = wine_data.target
 # Normalize and split    X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)    scaler = StandardScaler().fit(X_train)    X_train = scaler.transform(X_train)    X_test = scaler.transform(X_test)
    train_set = WineDataset(X_train, y_train)    test_set = WineDataset(X_test, y_test)
 print("Sample from dataset:")    sample_data, sample_target = train_set[0] print(f"Data: {sample_data}") print(f"Target: {sample_target}")
    model = SimpleNN()    optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
 return train_set, model, optimizer
def prepare_dataloader(dataset, batch_size): return DataLoader(dataset, batch_size=batch_size, pin_memory=True, shuffle=True)
def main(device, total_epochs, save_every, batch_size):    dataset, model, optimizer = load_train_objs()    train_data = prepare_dataloader(dataset, batch_size)    trainer = Trainer(model, train_data, optimizer, device, save_every)    trainer.train(total_epochs)
main(device=torch.device("cuda:0" if torch.cuda.is_available() else "cpu"), total_epochs=100, save_every=50, batch_size=32)

import torch.multiprocessing as mpfrom torch.utils.data.distributed import DistributedSamplerfrom torch.nn.parallel import DistributedDataParallel as DDPfrom torch.distributed import init_process_group, destroy_process_groupimport os

def ddp_setup(rank, world_size): """    Set up the distributed environment.     Args:        rank: The rank of the current process. Unique identifier for each process in the distributed training.        world_size: Total number of processes participating in the distributed training.    """  # Address of the main node. Since we are doing single-node training, it's set to localhost.    os.environ["MASTER_ADDR"] = "localhost"  # Port on which the master node is expected to listen for communications from workers.    os.environ["MASTER_PORT"] = "12355"  # Initialize the process group.  # 'backend' specifies the communication backend to be used, "nccl" is optimized for GPU training.    init_process_group(backend="nccl", rank=rank, world_size=world_size)  # Set the current CUDA device to the specified device (identified by rank). # This ensures that each process uses a different GPU in a multi-GPU setup.    torch.cuda.set_device(rank)

class Trainer(): def __init__(self, model, train_data, optimizer, gpu_id, save_every):        self.model = model.to(gpu_id)        self.train_data = train_data        self.optimizer = optimizer        self.gpu_id = gpu_id        self.save_every = save_every        self.losses = []  # This changes        self.model = DDP(self.model, device_ids=[gpu_id])

def prepare_dataloader(dataset: Dataset, batch_size: int): return DataLoader(        dataset,        batch_size=batch_size,        pin_memory=True,        shuffle=False,        sampler=DistributedSampler(dataset) )

def main(rank: int, world_size: int, save_every: int, total_epochs: int, batch_size: int): """    Main training function for distributed data parallel (DDP) setup.     Args:        rank (int): The rank of the current process (0 <= rank < world_size). Each process is assigned a unique rank.        world_size (int): Total number of processes involved in the distributed training.        save_every (int): Frequency of model checkpoint saving, in terms of epochs.        total_epochs (int): Total number of epochs for training.        batch_size (int): Number of samples processed in one iteration (forward and backward pass).    """  # Set up the distributed environment, including setting the master address, port, and backend.    ddp_setup(rank, world_size)  # Load the necessary training objects - dataset, model, and optimizer.    dataset, model, optimizer = load_train_objs()  # Prepare the data loader for distributed training. It partitions the dataset across the processes and handles shuffling.    train_data = prepare_dataloader(dataset, batch_size)  # Initialize the trainer instance with the loaded model, data, and other configurations.    trainer = Trainer(model, train_data, optimizer, rank, save_every)  # Train the model for the specified number of epochs.    trainer.train(total_epochs)  # Cleanup the distributed environment after training is complete.    destroy_process_group()

if __name__ == "__main__": import argparse    parser = argparse.ArgumentParser(description='simple distributed training job')    parser.add_argument('total_epochs', type=int, help='Total epochs to train the model')    parser.add_argument('save_every', type=int, help='How often to save a snapshot')    parser.add_argument('--batch_size', default=32, type=int, help='Input batch size on each device (default: 32)')    args = parser.parse_args()     world_size = torch.cuda.device_count()    mp.spawn(main, args=(world_size, args.save_every, args.total_epochs, args.batch_size), nprocs=world_size)

def ddp_setup(local_rank, world_size_per_node, node_rank):    os.environ["MASTER_ADDR"] = "MASTER_NODE_IP" # <-- Replace with your master node IP    os.environ["MASTER_PORT"] = "12355"     global_rank = node_rank * world_size_per_node + local_rank    init_process_group(backend="nccl", rank=global_rank, world_size=world_size_per_node*torch.cuda.device_count())    torch.cuda.set_device(local_rank)

def main(local_rank: int, world_size_per_node: int, save_every: int, total_epochs: int, batch_size: int, node_rank: int):    ddp_setup(local_rank, world_size_per_node, node_rank) # ... (rest of the main function)

if __name__ == "__main__": import argparse    parser = argparse.ArgumentParser(description='simple distributed training job')    parser.add_argument('total_epochs', type=int, help='Total epochs to train the model')    parser.add_argument('save_every', type=int, help='How often to save a snapshot')    parser.add_argument('--batch_size', default=32, type=int, help='Input batch size on each device (default: 32)')    parser.add_argument('--node_rank', default=0, type=int, help='The rank of the node in multi-node training')    args = parser.parse_args()
    world_size_per_node = torch.cuda.device_count()    mp.spawn(main, args=(world_size_per_node, args.save_every, args.total_epochs, args.batch_size, args.node_rank), nprocs=world_size_per_node)

#!/bin/bash#SBATCH --job-name=DDPTraining       # Name of the job#SBATCH --nodes=$1                   # Number of nodes specified by the user#SBATCH --ntasks-per-node=1          # Ensure only one task runs per node#SBATCH --cpus-per-task=1            # Number of CPU cores per task#SBATCH --gres=gpu:1                 # Number of GPUs per node#SBATCH --time=01:00:00              # Time limit hrs:min:sec (1 hour in this example)#SBATCH --mem=4GB                    # Memory limit per GPU#SBATCH --output=training_%j.log     # Output and error log name (%j expands to jobId)#SBATCH --partition=gpu              # Specify the partition or queue
srun python3 your_python_script.py --total_epochs 10 --save_every 2 --batch_size 32 --node_rank $SLURM_NODEID

sbatch train_net.sh 2 # for using 2 nodes