Difference between pytorch lightning and lightning · Lightning

Hi @sudarshan85 👋 let's see if I can help clarify this. Also, thank you (and @surak) for being persistent in this thread.

Note

we are going to address import conventions in issue 18327

I started using the Lightning ecosystem last year before joining the team a couple of months ago. I've experienced the changes as a community member, and now as a developer advocate. Before I get into the weeds, here is a high level explanation.

Diving In

Note

this is my framing and not an official position

Lightning is a mono package that contains 1) PyTorch Lightning 2) Lightning Fabric and 3) Lightning Apps. Using pip install lightning installs all three and TorchMetrics. I'll circle back to this after briefly covering Flash, Bolts, and TorchMetrics.

I think of Flash and Bolts as the scikit-learn of the Lightning AI ecosystem in that each is high-level, and contains model implementations. Whereas without Flash or Bolts we would create our own models "from-scratch" with PyTorch, and then train with PyTorch Lightning or Lightning Fabric – Lightning Fabric is a lower level trainer solution that gives more control back to experienced practitioners.

On Flash. Flash is high level and is for getting started with Deep Learning. Flash offers common deep learning tasks you can use out-of-the-box in a few lines of code, and it is built on top of PyTorch Lightning.

On Bolts. Bolts is also high level and built on top of PyTorch Lightning. We would want to use bolts for its in-built models, to finetune on our own data, and to extend the in-built models for additional research.

With respect to Flash and Bolts, this means the leveling looks something like:

TorchMetrics is a collection of 100+ metrics for several domains, some of which are: audio, image, text, classification, and regression. TorchMetrics can be thought of as the equivalent of sklearn.metrics. However, TorchMetrics implementations are capable of being used on CPU or GPU, and in distributed training environments.

On Lightning and PyTorch Lightning

Last year the team rolled out Lightning Apps and with that came a decision to unify PyTorch Lightning and Lightning Apps into a single repo and framework – Lightning. Around that time Lighting Fabric – a lower level trainer – was also created and placed into the Lightning repo. Meaning, all of (PyTorch Lightning, Lightning Fabric, and Lightning Apps) exist in Lightning.

We can still install PyTorch Lightning as a standalone framework for backwards compatibility with Lightning 2.0.

With respect to PyTorch Lightning and Lightning Fabric, the levels are:

And the flow for deciding between a PyTorch Lightning and custom Lightning Fabric Trainer is something like:

Where it is implied that either a PyTorch Lighting Trainer or a custom Trainer built with Lightning Fabric enjoys out of the box distributed training capabilities by setting common arguments such as devices, num_nodes, and strategy, and training on either CPU or GPU by setting the accelerator argument.