GitHub - Lightning-AI/pytorch-lightning: Pretrain, finetune ANY AI model of ANY size on multiple GPUs, TPUs with zero code changes. (original) (raw)

The deep learning framework to pretrain, finetune and deploy AI models.

NEW- Deploying models? Check out LitServe, the PyTorch Lightning for model serving

Quick start •Examples •PyTorch Lightning •Fabric •Lightning AI • Community •Docs

Lightning has 2 core packages

PyTorch Lightning: Train and deploy PyTorch at scale.
Lightning Fabric: Expert control.

Lightning gives you granular control over how much abstraction you want to add over PyTorch.

Quick start

Install Lightning:

Advanced install options

Install with optional dependencies

pip install lightning['extra']

Conda

conda install lightning -c conda-forge

Install stable version

Install future release from the source

pip install https://github.com/Lightning-AI/lightning/archive/refs/heads/release/stable.zip -U

Install bleeding-edge

Install nightly from the source (no guarantees)

pip install https://github.com/Lightning-AI/lightning/archive/refs/heads/master.zip -U

or from testing PyPI

pip install -iU https://test.pypi.org/simple/ pytorch-lightning

PyTorch Lightning example

Define the training workflow. Here's a toy example (explore real examples):

main.py

! pip install torchvision

import torch, torch.nn as nn, torch.utils.data as data, torchvision as tv, torch.nn.functional as F import lightning as L

--------------------------------

Step 1: Define a LightningModule

--------------------------------

A LightningModule (nn.Module subclass) defines a full system

(ie: an LLM, diffusion model, autoencoder, or simple image classifier).

class LitAutoEncoder(L.LightningModule): def init(self): super().init() self.encoder = nn.Sequential(nn.Linear(28 * 28, 128), nn.ReLU(), nn.Linear(128, 3)) self.decoder = nn.Sequential(nn.Linear(3, 128), nn.ReLU(), nn.Linear(128, 28 * 28))

def forward(self, x):
    # in lightning, forward defines the prediction/inference actions
    embedding = self.encoder(x)
    return embedding

def training_step(self, batch, batch_idx):
    # training_step defines the train loop. It is independent of forward
    x, _ = batch
    x = x.view(x.size(0), -1)
    z = self.encoder(x)
    x_hat = self.decoder(z)
    loss = F.mse_loss(x_hat, x)
    self.log("train_loss", loss)
    return loss

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

-------------------

Step 2: Define data

-------------------

dataset = tv.datasets.MNIST(".", download=True, transform=tv.transforms.ToTensor()) train, val = data.random_split(dataset, [55000, 5000])

-------------------

Step 3: Train

-------------------

autoencoder = LitAutoEncoder() trainer = L.Trainer() trainer.fit(autoencoder, data.DataLoader(train), data.DataLoader(val))

Run the model on your terminal

pip install torchvision python main.py

Why PyTorch Lightning?

PyTorch Lightning is just organized PyTorch - Lightning disentangles PyTorch code to decouple the science from the engineering.

Examples

Explore various types of training possible with PyTorch Lightning. Pretrain and finetune ANY kind of model to perform ANY task like classification, segmentation, summarization and more:

Task	Description	Run
Hello world	Pretrain - Hello world example
Image classification	Finetune - ResNet-34 model to classify images of cars
Image segmentation	Finetune - ResNet-50 model to segment images
Object detection	Finetune - Faster R-CNN model to detect objects
Text classification	Finetune - text classifier (BERT model)
Text summarization	Finetune - text summarization (Hugging Face transformer model)
Audio generation	Finetune - audio generator (transformer model)
LLM finetuning	Finetune - LLM (Meta Llama 3.1 8B)
Image generation	Pretrain - Image generator (diffusion model)
Recommendation system	Train - recommendation system (factorization and embedding)
Time-series forecasting	Train - Time-series forecasting with LSTM

Advanced features

Lightning has over 40+ advanced features designed for professional AI research at scale.

Here are some examples:

Train on 1000s of GPUs without code changes

8 GPUs

no code changes needed

trainer = Trainer(accelerator="gpu", devices=8)

256 GPUs

trainer = Trainer(accelerator="gpu", devices=8, num_nodes=32)

Train on other accelerators like TPUs without code changes

no code changes needed

trainer = Trainer(accelerator="tpu", devices=8)

16-bit precision

no code changes needed

trainer = Trainer(precision=16)

Experiment managers

from lightning import loggers

tensorboard

trainer = Trainer(logger=TensorBoardLogger("logs/"))

weights and biases

trainer = Trainer(logger=loggers.WandbLogger())

comet

trainer = Trainer(logger=loggers.CometLogger())

mlflow

trainer = Trainer(logger=loggers.MLFlowLogger())

neptune

trainer = Trainer(logger=loggers.NeptuneLogger())

... and dozens more

Early Stopping

es = EarlyStopping(monitor="val_loss") trainer = Trainer(callbacks=[es])

Checkpointing

checkpointing = ModelCheckpoint(monitor="val_loss") trainer = Trainer(callbacks=[checkpointing])

Export to torchscript (JIT) (production use)

torchscript

autoencoder = LitAutoEncoder() torch.jit.save(autoencoder.to_torchscript(), "model.pt")

Export to ONNX (production use)

onnx

with tempfile.NamedTemporaryFile(suffix=".onnx", delete=False) as tmpfile: autoencoder = LitAutoEncoder() input_sample = torch.randn((1, 64)) autoencoder.to_onnx(tmpfile.name, input_sample, export_params=True) os.path.isfile(tmpfile.name)

Advantages over unstructured PyTorch

• Models become hardware agnostic
• Code is clear to read because engineering code is abstracted away
• Easier to reproduce
• Make fewer mistakes because lightning handles the tricky engineering
• Keeps all the flexibility (LightningModules are still PyTorch modules), but removes a ton of boilerplate
• Lightning has dozens of integrations with popular machine learning tools.
• Tested rigorously with every new PR. We test every combination of PyTorch and Python supported versions, every OS, multi GPUs and even TPUs.
• Minimal running speed overhead (about 300 ms per epoch compared with pure PyTorch).

Lightning Fabric: Expert control

Run on any device at any scale with expert-level control over PyTorch training loop and scaling strategy. You can even write your own Trainer.

Fabric is designed for the most complex models like foundation model scaling, LLMs, diffusion, transformers, reinforcement learning, active learning. Of any size.

Key features

Easily switch from running on CPU to GPU (Apple Silicon, CUDA, …), TPU, multi-GPU or even multi-node training

Use your available hardware

no code changes needed

fabric = Fabric()

Run on GPUs (CUDA or MPS)

fabric = Fabric(accelerator="gpu")

8 GPUs

fabric = Fabric(accelerator="gpu", devices=8)

256 GPUs, multi-node

fabric = Fabric(accelerator="gpu", devices=8, num_nodes=32)

Run on TPUs

fabric = Fabric(accelerator="tpu")

Use state-of-the-art distributed training strategies (DDP, FSDP, DeepSpeed) and mixed precision out of the box

Use state-of-the-art distributed training techniques

fabric = Fabric(strategy="ddp") fabric = Fabric(strategy="deepspeed") fabric = Fabric(strategy="fsdp")

Switch the precision

fabric = Fabric(precision="16-mixed") fabric = Fabric(precision="64")

All the device logic boilerplate is handled for you

no more of this!

• model.to(device)
• batch.to(device)

Build your own custom Trainer using Fabric primitives for training checkpointing, logging, and more

import lightning as L

class MyCustomTrainer: def init(self, accelerator="auto", strategy="auto", devices="auto", precision="32-true"): self.fabric = L.Fabric(accelerator=accelerator, strategy=strategy, devices=devices, precision=precision)

def fit(self, model, optimizer, dataloader, max_epochs):
    self.fabric.launch()

    model, optimizer = self.fabric.setup(model, optimizer)
    dataloader = self.fabric.setup_dataloaders(dataloader)
    model.train()

    for epoch in range(max_epochs):
        for batch in dataloader:
            input, target = batch
            optimizer.zero_grad()
            output = model(input)
            loss = loss_fn(output, target)
            self.fabric.backward(loss)
            optimizer.step()

You can find a more extensive example in our examples

Examples

Self-supervised Learning

• CPC transforms
• Moco v2 transforms
• SimCLR transforms

Convolutional Architectures

• GPT-2
• UNet

Reinforcement Learning

• DQN Loss
• Double DQN Loss
• Per DQN Loss

GANs

• Basic GAN
• DCGAN

Classic ML

• Logistic Regression
• Linear Regression

Continuous Integration

Lightning is rigorously tested across multiple CPUs, GPUs and TPUs and against major Python and PyTorch versions.

*Codecov is > 90%+ but build delays may show less

Current build statuses

System / PyTorch ver.	1.13	2.0	2.1
Linux py3.9 [GPUs]
Linux (multiple Python versions)
OSX (multiple Python versions)
Windows (multiple Python versions)

Community

The lightning community is maintained by

• 10+ core contributors who are all a mix of professional engineers, Research Scientists, and Ph.D. students from top AI labs.
• 800+ community contributors.

Want to help us build Lightning and reduce boilerplate for thousands of researchers? Learn how to make your first contribution here

Lightning is also part of the PyTorch ecosystem which requires projects to have solid testing, documentation and support.

Asking for help

If you have any questions please:

1. Read the docs.
2. Search through existing Discussions, or add a new question
3. Join our discord.