PyTorch’s dynamic computation graph (“define-by-run”) makes debugging neural networks intuitive — you use standard Python control flow and print() statements. It became the dominant framework for ML research and, with TorchServe and torch.compile, is now widely used in production at Meta, Tesla, and OpenAI. Most modern LLMs (including GPT, LLaMA, and Mistral) are built and trained in PyTorch.

Quick start

pip install torch torchvision

import torch
import torch.nn as nn
import torch.optim as optim

# Define a simple feedforward network
class Net(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc1 = nn.Linear(784, 128)
        self.fc2 = nn.Linear(128, 10)
        self.relu = nn.ReLU()

    def forward(self, x):
        x = self.relu(self.fc1(x))
        return self.fc2(x)

model = Net()
optimizer = optim.Adam(model.parameters(), lr=1e-3)
criterion = nn.CrossEntropyLoss()

# Training loop
for epoch in range(10):
    for batch_x, batch_y in dataloader:
        optimizer.zero_grad()
        output = model(batch_x)
        loss = criterion(output, batch_y)
        loss.backward()
        optimizer.step()
    print(f"Epoch {epoch+1}, Loss: {loss.item():.4f}")

# Save and load
torch.save(model.state_dict(), 'model.pt')
model.load_state_dict(torch.load('model.pt'))

When to use

PyTorch is the first choice for deep learning research, fine-tuning pretrained models (via HuggingFace), and production ML systems. Its Pythonic API and dynamic graphs make experimentation fast. TensorFlow/Keras is still used in mobile deployment (TFLite) and some production pipelines, but PyTorch’s HuggingFace integration and torch.compile speedups have made it the default in most new ML projects. JAX is worth considering for large-scale TPU training.

// features

Dynamic computation graph — define-by-run, debug with standard Python tools

Autograd — automatic differentiation for any computation graph

CUDA acceleration — move tensors to GPU with `.to('cuda')`

`torch.nn` module system for building neural network layers

DataLoader and Dataset abstractions for efficient data pipelines

TorchScript for exporting models to production without Python

torch.compile (PyTorch 2.0) for significant speedups with no code changes

Vast ecosystem — HuggingFace Transformers, Lightning, torchvision, torchaudio

PyTorch

Quick start

When to use

// features

// installation

// tags