●Change8

2.5.2

📦 pytorch-lightningView on GitHub →
✨ 1 features🐛 8 fixes🔧 7 symbols

Summary

This release introduces the `toggled_optimizer` context manager to LightningModule and resolves several bugs related to CLI integration, DDP synchronization, and checkpointing. Users are advised to update `fsspec` for cross-device checkpointing.

Migration Steps

  1. For cross-device local checkpoints, install `fsspec>=2025.5.0` if unavailable.

✨ New Features

  • Add `toggled_optimizer(optimizer)` method to the LightningModule, which is a context manager version of `toggle_optimize` and `untoggle_optimizer`.

🐛 Bug Fixes

  • Fixed `save_hyperparameters` not working correctly with `LightningCLI` when there are parsing links applied on instantiation.
  • Fixed `logger_connector` has an edge case where step can be a float.
  • Fixed Synchronize SIGTERM Handling in DDP to Prevent Deadlocks.
  • Fixed case-sensitive model name.
  • CLI: resolve jsonargparse deprecation warning.
  • Fix: move `check_inputs` to the target device if available during `to_torchscript`.
  • Fixed progress bar display to correctly handle iterable dataset and `max_steps` during training.
  • Fixed problem for silently supporting `jsonnet`.

🔧 Affected Symbols

LightningModulesave_hyperparametersLightningCLItoggle_optimizeuntoggle_optimizercheck_inputsto_torchscript