distributed-llm-training/train_toy.py at main · aaravM123/distributed-llm-training · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
import argparse
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import torchvision
import torchvision.transforms as transforms
import wandb
import torch.distributed as dist
from torch.utils.data.distributed import DistributedSampler
from torch.cuda.amp import autocast, GradScaler


class SimpleCNN(nn.Module):
    def __init__(self, num_classes=10):
        super(SimpleCNN, self).__init__()
        self.conv1 = nn.Conv2d(3,32,3,padding=1)
        self.conv2 = nn.Conv2d(32,64,3,padding=1)
        self.pool = nn.MaxPool2d(2,2)
        self.fc1 = nn.Linear(64*8*8,128)
        self.fc2 = nn.Linear(128,10)

    def forward(self,x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        x = x.view(-1,64*8*8)
        x = F.relu(self.fc1(x))
        x = self.fc2(x)
        return x

def train(args):
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    print("Using device:", device)

    wandb.init(project="ddp-toy")


    transform = transforms.Compose([transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,))])])


    trainset = torchvision.datasets.CIFAR10(root='./data', train=True, download=True, transform=transform)

    if args.mode == "ddp":
        dist.init_process_group(backend = args.backend)

    if args.mode == "ddp":
        train_sampler = DistributedSampler(trainset)
        trainloader = torch.utils.data.DataLoader(
            trainset, batch_size=args.batch_size, sampler=train_sampler, num_workers=2
        )
    else:
        trainloader = torch.utils.data.DataLoader(
            trainset, batch_size=args.batch_size, shuffle=True, num_workers=2
        )


    model = SimpleCNN().to(device)

    if args.mode == "dp":
        print("Running in DataParallel mode...")
        model = nn.DataParallel(model)
    elif args.mode == "ddp":
        print("Running in Distributed DataParallel mode...")
        model = nn.parallel.DistributedDataParallel(
            model, device_ids=[args.local_rank]
        )


    criterion = nn.CrossEntropyLoss()
    optimizer = optim.Adam(model.parameters(), lr=args.learning_rate)
    scaler = GradScaler(enabled=args.use_amp)

    for epoch in range(args.epochs):
        running_loss = 0.0
        for i, (inputs, labels) in enumerate(trainloader, 0):
            inputs, labels = inputs.to(device), labels.to(device)

            with autocast(enabled=args.use_amp):
                outputs = model(inputs)
                loss = criterion(outputs, labels)
                loss = loss / args.grad_accum_steps

            scaler.scale(loss).backward()

            if (i+1) % args.grad_accum_steps == 0:
                scaler.step(optimizer)
                scaler.update()
                optimizer.zero_grad()

            wandb.log({"loss": loss.item()})

            running_loss += loss.item()
            if i % 100 == 99:
                msg = f"[Epoch {epoch+1}, Step {i+1}] loss: {running_loss/100:.3f}"
                print(msg)
                with open("train.log", "a") as f:
                    f.write(msg + "\n")
                running_loss = 0.0

    print("Finished Training")

    if args.mode == "ddp":
        dist.destroy_process_group()

def main():
    import torch.distributed as dist

    if dist.is_initialized():
        rank = dist.get_rank()
        world_size = dist.get_world_size()
        print(f"Hello from rank {rank} out of {world_size}")
        dist.barrier()
        print(f"Rank {rank} passed barrier")

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--epochs", type=int, default=2)
    parser.add_argument("--batch_size", type=int, default=64)
    parser.add_argument("--learning_rate", type=float, default=0.001)
    parser.add_argument("--mode", choices=["single", "dp", "ddp"], default="single")
    parser.add_argument("--backend", type=str, default="nccl")
    parser.add_argument("--world_size", type=int, default=1)
    parser.add_argument("--local_rank", type=int, default=0)
    parser.add_argument("--grad_accum_steps", type = int, default=1, help="Number of steps to accumulate gradients before updating.")
    parser.add_argument("--use_amp", action="store_true", help="Enable mixed precision training")

    args = parser.parse_args()
    main()
    train(args)