Run a script on a GPU with one line of code

GPUs are powerful but only have an impact when they’re easy to use

GPUs can provide impressive performance boosts for certain workflows like training ML models, computer vision, analytics, and more. However GPU hardware can also be difficult to access and set up properly.

Here we show how to run a Python script on a GPU-enabled cloud machine with one line of code from your laptop.

$ coiled run --gpu python train.py

Example: Train a GPU-accelerated PyTorch model

Let’s train a GPU-accelerated PyTorch neural network on the CIFAR10 dataset. To start we’ll install the packages we need:

conda

$ conda create -n pytorch -c conda-forge python=3.11 pytorch torchvision coiled
$ conda activate pytorch

pip

$ pip install torch torchvision coiled

Next, we’ll save the PyTorch code below in a train.py Python script.

# train.py

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torchvision
import torchvision.transforms as transforms

transform = transforms.Compose(
    [
        transforms.ToTensor(),
        transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2470, 0.2435, 0.2616)),
    ]
)

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(3, 2_500, 5)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(2_500, 16, 5)
        self.fc1 = nn.Linear(16 * 5 * 5, 120)
        self.fc2 = nn.Linear(120, 84)
        self.fc3 = nn.Linear(84, 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, 16 * 5 * 5)
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x

# Select what hardware to use
if torch.cuda.is_available():
    device = torch.device("cuda:0")      # NVIDIA GPU
elif torch.backends.mps.is_available():
    device = torch.device("mps")         # Apple silicon GPU
else:
    device = torch.device("cpu")         # CPU
print(f"{device = }")

net = Net()
net = net.to(device)

trainset = torchvision.datasets.CIFAR10(
    root="./data",
    train=True,
    download=True,
    transform=transform,
)
trainloader = torch.utils.data.DataLoader(
    trainset,
    batch_size=400,
    shuffle=True,
)
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)

# Train model on 10 passes over the data
for epoch in range(10):
    print(f"Epoch {epoch}")
    for i, data in enumerate(trainloader, 0):
        inputs = data[0].to(device)
        labels = data[1].to(device)
        optimizer.zero_grad()
        outputs = net(inputs)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()

Note that PyTorch makes it straightforward to use different kinds of accelerated hardware. This portion of train.py:

# Select local hardware device to use
if torch.cuda.is_available():
    device = torch.device("cuda:0")      # NVIDIA GPU
elif torch.backends.mps.is_available():
    device = torch.device("mps")         # Apple silicon GPU
else:
    device = torch.device("cpu")         # CPU

tells PyTorch to run on an NVIDIA GPU, Apple silicon GPU, or traditional CPU, in that order, depending on what’s available locally where the code run.

When run on a Macbook Pro CPU, this script takes ~6.8 hours to run. When run on an Apple M1 GPU that time is reduced to ~1.4 hours (factor of ~4.9x speedup). From the same Macbook Pro we can run the training script on an NVIDIA GPU on a cloud VM using the following Coiled Run command:

$ coiled run --gpu python train.py

Under the hood, coiled run handles:

• Provisioning a cloud VM with GPU hardware. In this case a g4dn.xlarge (Tesla T4) instance on AWS.

• Setting up the appropriate NVIDIA drivers, CUDA runtime, etc.

• Automatically installing the same packages I have locally on the cloud VM.

• Making sure to install the CUDA-compiled version of PyTorch on the cloud VM.

• Running python train.py on the cloud VM.

Our coiled run command takes ~14.1 minutes to complete, costs $0.12, and we have good GPU hardware utilization.

GPU utilization and memory usage. Both are near their maximum capacity, meaning we’re utilizing the available hardware well.

We spend ~4.7 minutes requesting, provisioning, and setting up the cloud VM, and ~9.8 minutes running the script. That’s a ~6x and ~29x overall speedup (including VM startup time) over the Apple silicon and CPU cases, respectively. Model training is ~8.6x and ~42x faster on the NVIDIA T4 GPU than on Apple silicon and CPU, respectively.

Comparing duration for running the same PyTorch training script on a CPU, Apple silicon GPU, and NVIDIA T4 GPU.

As we can see, PyTorch and Coiled complement each other well here. PyTorch handles using hardware that’s available locally, and Coiled handles running code on advanced hardware on the cloud. What’s nice here is one doesn’t really need to think about cloud devops or GPU software environment management to train a model on the GPU of their choosing. One can develop locally and, when needed, easily scale out to a GPU on the cloud.

Extend as needed

coiled run works well out of the box for common use cases, but also accommodates more complex situations like:

• Larger GPU: Use --vm-type to run on any instance type available from your cloud provider

• Cloud-hosted data: Use --region to run in the same region as your data

• Complex software: Use --container to use a custom Docker container

• Warm start: Use --keepalive to keep VMs up for a given amount of time for quick reuse

• Interactive use: Use --interactive bash to be dropped into a terminal on your cloud VM

For example, running an LLM fine-tuning script might look something like:

$ coiled run \
    --vm-type p4d.24xlarge \                           # 8 A100 GPU node
    --region us-west-2 \                               # Same region as my data
    --container huggingface/transformers-pytorch-gpu \ # Use this Docker image
    python fine_tune.py  

Summary

We showed how to use coiled run to run a PyTorch script on a cloud GPU. This was:

• Easy: Run scripts on a GPU with one line of code

• Fast: Get benefits of GPU acceleration without any devops work

• Extensible: Extend coiled run to match your specific use case

We hope the approach discussed here helps reduce friction in GPU workloads and can be copied / adapted for your own use case.

Here are additional GPU examples where some of the same Coiled features discussed here are also used:

• How to Run Your Jupyter Notebook on a GPU in the Cloud

• How to Train a Neural Network on a GPU in the Cloud with Coiled Functions

previous

Machine Learning

next

XGBoost for Gradient Boosted Trees