pytorch-lightning

Project Overview

PyTorch Lightning is a deep learning framework for pretraining and finetuning AI models. Training in plain PyTorch means writing repetitive, error-prone engineering code for backpropagation, mixed precision, multi-GPU, and distributed training. Lightning automates that infrastructure so you focus on model and data logic. The core code scales from CPU to multi-node GPU clusters without changes, and Lightning Fabric offers finer-grained expert control.

You write the science, Lightning handles the engineering, and training scales linearly with hardware.

Key Features

• Automated training loop: backpropagation, optimizer scheduling, validation, and logging are handled for you
• Hardware-agnostic scaling: switch the same code across CPUs, single GPUs, multi-GPU, and multi-node clusters
• Mixed precision: enable 16-bit training with one flag
• LightningModule: organize model, training step, and optimizers into a standard structure
• Fabric expert mode: insert fine-grained control where needed without rewriting your code
• Callbacks and checkpoints: resume from checkpoints, early stopping, and metric monitoring as pluggable callbacks

Target Audience

• Researchers tired of rewriting training loops
• Engineers running multi-GPU distributed training
• Teams maintaining many deep learning projects
• Students reproducing paper experiments

Quick Start

1. Install:

pip install lightning

2. Write your model as a LightningModule implementing training_step and configure_optimizers
3. Start training with the Trainer class, switching GPU count and precision via flags
4. Switch to Lightning Fabric when you need finer control; code changes are minimal

Core Advantages

• Scale without changes: from 1 GPU to 10,000+ with the same core model code
• Clean structure: training code organized by module, reproducible and easy to hand over
• Optional abstraction: from fully automatic Trainer to semi-automatic Fabric, you pick the level

System Requirements

Notes

• Prerequisite: PyTorch must be installed first; GPU training needs a matching CUDA build
• Two interfaces: PyTorch Lightning and Fabric differ in usage; pick one per project