Lightning AI

Technical Getting Started

Installation

pip install lightning

Define a Training Module (PyTorch Lightning)

import lightning as L
import torch
from torch import nn

class LitModel(L.LightningModule):
    def __init__(self):
        super().__init__()
        self.layer = nn.Linear(28 * 28, 10)

    def training_step(self, batch, batch_idx):
        x, y = batch
        loss = nn.functional.cross_entropy(self.layer(x.view(x.size(0), -1)), y)
        return loss

    def configure_optimizers(self):
        return torch.optim.Adam(self.parameters(), lr=1e-3)

Run Training on Cloud GPUs

• Sign up at lightning.ai and create a Studio
• Connect your GitHub repo or upload code directly
• Select GPU hardware (A100, H100, etc.) and launch training
• Monitor runs via built-in experiment tracking

Key Concepts

• LightningModule: Encapsulates model, training loop, optimizer, and validation logic
• Trainer: Handles all engineering complexity — multi-GPU, mixed precision, checkpointing
• Lightning Studios: Cloud-based development environments with pre-configured compute
• Lightning Apps: Composable workflows linking data, training, and serving components

Distributed Training

trainer = L.Trainer(
    accelerator="gpu",
    devices=8,
    num_nodes=4,
    strategy="ddp"
)
trainer.fit(model, datamodule)