●Change8

2.5.3

📦 pytorch-lightningView on GitHub →
✨ 5 features🐛 13 fixes🔧 11 symbols

Summary

This release brings numerous bug fixes across PyTorch Lightning and Lightning Fabric, including improvements to checkpointing, logging, profiling, and progress bar rendering. New features include more flexibility in ModelCheckpoint options and handling of training_step returns.

Migration Steps

  1. If using Lightning Fabric CLI arguments for device/accelerator configuration, 'auto' is now supported.

✨ New Features

  • Added 'save_on_exception' option to ModelCheckpoint Callback.
  • Allow 'dataloader_idx_' in log names when 'add_dataloader_idx=False'.
  • Allow returning ONNXProgram when calling to_onnx(dynamo=True).
  • Extended support for general mappings being returned from training_step when using manual optimization.
  • Enable "auto" for 'devices' and 'accelerator' as CLI arguments in Lightning Fabric.

🐛 Bug Fixes

  • Fixed allowing trainer to accept CUDAAccelerator instance as accelerator with FSDP strategy.
  • Fixed progress bar console clearing for Rich 14.1+.
  • Fixed AdvancedProfiler to handle nested profiling actions for Python 3.12+.
  • Fixed rich progress bar error when resume training.
  • Fixed double iteration bug when resumed from a checkpoint.
  • Fixed support for more dtypes in ModelSummary.
  • Fixed metrics in RichProgressBar being updated according to user provided 'refresh_rate'.
  • Fixed 'save_last' behavior in the absence of validation.
  • Fixed integration between LearningRateFinder and EarlyStopping.
  • Fixed gradient calculation in lr_finder for mode="exponential".
  • Fixed save_hyperparameters crashing with dataclasses using init=False fields.
  • Fixed remove extra 'name' parameter in accelerator registry decorator in Lightning Fabric.
  • Fixed XLA strategy to add support for 'global_ordinal', 'local_ordinal', 'world_size' which came instead of deprecated methods in Lightning Fabric.

🔧 Affected Symbols

ModelCheckpointtraining_stepCUDAAcceleratorFSDPRichProgressBarAdvancedProfilerModelSummaryLearningRateFinderEarlyStoppingsave_hyperparametersaccelerator registry decorator