●Change8

2.6.0

📦 pytorch-lightningView on GitHub →
✨ 7 features🐛 14 fixes🔧 17 symbols

Summary

Version 2.6.0 introduces several new features like WeightAveraging callbacks and Torch-Tensorrt integration, alongside numerous bug fixes across PyTorch Lightning and Fabric components.

✨ New Features

  • Added WeightAveraging callback that wraps the PyTorch AveragedModel class.
  • Added Torch-Tensorrt integration with LightningModule.
  • Added time-based validation support though val_check_interval.
  • Added attributes to access stopping reason in EarlyStopping callback.
  • Added support for variable batch size in ThroughputMonitor.
  • Added EMAWeightAveraging callback that wraps Lightning's WeightAveraging class.
  • Added kwargs-filtering for Fabric.call to support different callback method signatures.

🐛 Bug Fixes

  • Fixed edgecase when max_trials is reached in Tuner.scale_batch_size.
  • Fixed case where LightningCLI could not be initialized with trainer_default containing callbacks.
  • Fixed missing reset when ModelPruning is applied with lottery ticket hypothesis.
  • Fixed preventing recursive symlink creation iwhen save_last='link' and save_top_k=-1.
  • Fixed last.ckpt being created and not linked to another checkpoint.
  • Fixed bug that prevented BackboneFinetuning from being used together with LearningRateFinder.
  • Fixed ModelPruning sparsity logging bug that caused incorrect sparsity percentages.
  • Fixed LightningCLI loading of hyperparameters from ckpt_path failing for subclass model mode.
  • Fixed check the init args only when the given frames are in __init__ method.
  • Fixed how ThroughputMonitor calculated training time.
  • Fixed synchronization of gradients in manual optimization with DDPStrategy(static_graph=True).
  • Fixed FSDP mixed precision semantics and added user warning (in both PyTorch Lightning and Lightning Fabric).
  • Fixed issue in detecting MPIEnvironment with partial mpi4py installation.
  • Learning rate scheduler is stepped at the end of epoch when on_train_batch_start returns -1.

🔧 Affected Symbols

WeightAveragingLightningModuleval_check_intervalEarlyStoppingThroughputMonitorEMAWeightAveragingTrainer.{fit,validate,test,predict}RichProgressBarModelSummaryTQDMProgressBarModelPruningLightningCLIBackboneFinetuningLearningRateFinderDDPStrategyFabric.call_DeviceDtypeModuleMixin._device