[BUG] - Hyperparameter Tuning with Ray Tune doesn't work for multi GPU

[f2010126]

Add Link

The tutorial given here using DP Pytorch and Ray doesn't work when gpus_per_trial>1. Link: https://pytorch.org/tutorials/beginner/hyperparameter_tuning_tutorial.html

Describe the bug

Using the code from the tutorial without any changes

from functools import partial
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import random_split
import torchvision
import torchvision.transforms as transforms
from ray import tune
from ray.air import Checkpoint, session
from ray.tune.schedulers import ASHAScheduler

def load_data(data_dir="./data"):
    transform = transforms.Compose(
        [transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]
    )

    trainset = torchvision.datasets.CIFAR10(
        root=data_dir, train=True, download=True, transform=transform
    )

    testset = torchvision.datasets.CIFAR10(
        root=data_dir, train=False, download=True, transform=transform
    )

    return trainset, testset

class Net(nn.Module):
    def __init__(self, l1=120, l2=84):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(3, 6, 5)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(6, 16, 5)
        self.fc1 = nn.Linear(16 * 5 * 5, l1)
        self.fc2 = nn.Linear(l1, l2)
        self.fc3 = nn.Linear(l2, 10)

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        x = torch.flatten(x, 1)  # flatten all dimensions except batch
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x

def train_cifar(config, data_dir=None):
    net = Net(config["l1"], config["l2"])

    device = "cpu"
    if torch.cuda.is_available():
        device = "cuda:0"
        if torch.cuda.device_count() > 1:
            net = nn.DataParallel(net)
    net.to(device)

    criterion = nn.CrossEntropyLoss()
    optimizer = optim.SGD(net.parameters(), lr=config["lr"], momentum=0.9)

    checkpoint = session.get_checkpoint()

    if checkpoint:
        checkpoint_state = checkpoint.to_dict()
        start_epoch = checkpoint_state["epoch"]
        net.load_state_dict(checkpoint_state["net_state_dict"])
        optimizer.load_state_dict(checkpoint_state["optimizer_state_dict"])
    else:
        start_epoch = 0

    trainset, testset = load_data(data_dir)

    test_abs = int(len(trainset) * 0.8)
    train_subset, val_subset = random_split(
        trainset, [test_abs, len(trainset) - test_abs]
    )

    trainloader = torch.utils.data.DataLoader(
        train_subset, batch_size=int(config["batch_size"]), shuffle=True, num_workers=8
    )
    valloader = torch.utils.data.DataLoader(
        val_subset, batch_size=int(config["batch_size"]), shuffle=True, num_workers=8
    )
    print('Starting training')
    for epoch in range(start_epoch, 1):  # loop over the dataset multiple times
        running_loss = 0.0
        epoch_steps = 0
        for i, data in enumerate(trainloader, 0):
            # get the inputs; data is a list of [inputs, labels]
            inputs, labels = data
            inputs, labels = inputs.to(device), labels.to(device)

            # zero the parameter gradients
            optimizer.zero_grad()

            # forward + backward + optimize
            outputs = net(inputs)
            loss = criterion(outputs, labels)
            loss.backward()
            optimizer.step()
            print(f'Stepped')

            # print statistics
            running_loss += loss.item()
            epoch_steps += 1
            if i % 2000 == 1999:  # print every 2000 mini-batches
                print(
                    "[%d, %5d] loss: %.3f"
                    % (epoch + 1, i + 1, running_loss / epoch_steps)
                )
                running_loss = 0.0

        # Validation loss
        val_loss = 0.0
        val_steps = 0
        total = 0
        correct = 0
        for i, data in enumerate(valloader, 0):
            with torch.no_grad():
                inputs, labels = data
                inputs, labels = inputs.to(device), labels.to(device)

                outputs = net(inputs)
                _, predicted = torch.max(outputs.data, 1)
                total += labels.size(0)
                correct += (predicted == labels).sum().item()

                loss = criterion(outputs, labels)
                val_loss += loss.cpu().numpy()
                val_steps += 1

        checkpoint_data = {
            "epoch": epoch,
            "net_state_dict": net.state_dict(),
            "optimizer_state_dict": optimizer.state_dict(),
        }
        checkpoint = Checkpoint.from_dict(checkpoint_data)

        session.report(
            {"loss": val_loss / val_steps, "accuracy": correct / total},
            checkpoint=checkpoint,
        )
    print("Finished Training")

def test_accuracy(net, device="cpu"):
    trainset, testset = load_data()

    testloader = torch.utils.data.DataLoader(
        testset, batch_size=4, shuffle=False, num_workers=2
    )

    correct = 0
    total = 0
    with torch.no_grad():
        for data in testloader:
            images, labels = data
            images, labels = images.to(device), labels.to(device)
            outputs = net(images)
            _, predicted = torch.max(outputs.data, 1)
            total += labels.size(0)
            correct += (predicted == labels).sum().item()

    return correct / total

def main(num_samples=10, max_num_epochs=10, gpus_per_trial=2):
    data_dir = os.path.abspath("./data")
    load_data(data_dir)
    config = {
        "l1": tune.choice([2**i for i in range(9)]),
        "l2": tune.choice([2**i for i in range(9)]),
        "lr": tune.loguniform(1e-4, 1e-1),
        "batch_size": tune.choice([2, 4, 8, 16]),
    }
    scheduler = ASHAScheduler(
        metric="loss",
        mode="min",
        max_t=max_num_epochs,
        grace_period=1,
        reduction_factor=2,
    )
    result = tune.run(
        partial(train_cifar, data_dir=data_dir),
        resources_per_trial={"cpu": 2, "gpu": gpus_per_trial},
        config=config,
        num_samples=num_samples,
        scheduler=scheduler,
    )

    best_trial = result.get_best_trial("loss", "min", "last")
    print(f"Best trial config: {best_trial.config}")
    print(f"Best trial final validation loss: {best_trial.last_result['loss']}")
    print(f"Best trial final validation accuracy: {best_trial.last_result['accuracy']}")

    best_trained_model = Net(best_trial.config["l1"], best_trial.config["l2"])
    device = "cpu"
    if torch.cuda.is_available():
        device = "cuda:0"
        if gpus_per_trial > 1:
            best_trained_model = nn.DataParallel(best_trained_model)
    best_trained_model.to(device)

    best_checkpoint = best_trial.checkpoint.to_air_checkpoint()
    best_checkpoint_data = best_checkpoint.to_dict()

    best_trained_model.load_state_dict(best_checkpoint_data["net_state_dict"])

    test_acc = test_accuracy(best_trained_model, device)
    print("Best trial test set accuracy: {}".format(test_acc))

if __name__ == "__main__":
    # You can change the number of GPUs per trial here:
    main(num_samples=3, max_num_epochs=10, gpus_per_trial=2)

Expected The run works using 2 GPUs per trial.

Observed The code starts, the data loaders are created. But when creating the distributed process group, everything hangs.

Describe your environment

Env: OS: Ubuntu 22.04 on a SLURM Cluster Python: 3.10 torch==2.0.1 ray ==2.6.3

NVCC -V output:

Cuda compilation tools, release 11.7, V11.7.99
Build cuda_11.7.r11.7/compiler.31442593_0

Nvidia-smi output

NVIDIA-SMI 530.30.02              
Driver Version: 530.30.02    
CUDA Version: 12.1

cc @sekyondaMeta @svekars @carljparker @NicolasHug @kit1980 @subramen

[soma2000-lang]

/assigntome

[svekars]

This issue has been unassigned due to inactivity. If you are working on this issue, assign it to yourself and send a PR ASAP.

[soma2000-lang]

/assigntome