lightning-ai/pytorch-lightning is a deep learning framework for pre-training and fine-tuning PyTorch models. It automates repetitive engineering code while leaving model logic under the user's direct control, so training runs can scale from CPU environments up to multi-node GPUs. According to the GitHub releases, the latest version is 2.6.6.
The goal of this project is to structure the PyTorch training loop and distributed execution. The repository description states that it pre-trains and fine-tunes models of any size, from 1 GPU to more than 10,000 GPUs, without any code changes.
There are two core packages. PyTorch Lightning handles training and deployment at the scale you need, and Lightning Fabric provides finer-grained control. Both packages manage distributed strategies, checkpoints, precision settings, and iterative training runs. It is implemented in Python and uses the Apache-2.0 license. The repository was created in March 2019, and its GitHub star count is recorded as 31423.
However, because this record contains no release notes, the changes in 2.6.6 cannot be confirmed. No performance benchmarks or comparisons with other frameworks are presented either. The momentum score assigned by the research record is a low 2.6 out of 10. So rather than a version with new features added, it is better viewed as a subject for examining the structure of the training loop and distributed execution.