PyTorch Lightning 2.6.1

Breaking Changes

• Removed Python 3.9 support — end-of-life status

Deprecations

• to_torchscript() deprecated due to PyTorch TorchScript deprecation

Features

• LightningModule.freeze() and unfreeze() now return self for method chaining
• Added litlogger integration
• Exposed weights_only argument in Fabric.load() and Fabric.load_raw()

Fixes

• save_hyperparameters(ignore=...) — subclass rules now properly override base class rules
• LightningDataModule.load_from_checkpoint() now correctly restores subclass type and hyperparameters
• ModelParallelStrategy checkpointing with torch.compile — fixed optimizer state KeyError on save
• Profiler filename sanitization prevents crashes from invalid characters
• StochasticWeightAveraging works with infinite epochs
• _generate_seed_sequence_sampling() now produces unique seeds
• ThroughputMonitor callback reduced excessive warning frequency
• DistributedSamplerWrapper properly forwards set_epoch() to underlying sampler
• DDP notebook CUDA fork check allows passive initialization when CUDA unused