Change8

2.5.4

📦 pytorch-lightningView on GitHub →
1 features🐛 5 fixes🔧 5 symbols

Summary

This patch release for PyTorch Lightning focuses on bug fixes across checkpointing, callbacks, and strategy integrations. Lightning Fabric also added support for NVIDIA H200 GPUs.

✨ New Features

  • Added support for NVIDIA H200 GPUs in `get_available_flops`.

🐛 Bug Fixes

  • Fixed `AsyncCheckpointIO` snapshots tensors to avoid race with parameter mutation.
  • Fixed `AsyncCheckpointIO` threadpool exception if calling fit or validate more than one.
  • Fixed learning rate not being correctly set after using `LearningRateFinder` callback.
  • Fixed misalignment column while using rich model summary in `DeepSpeedstrategy`.
  • Fixed `RichProgressBar` crashing when sanity checking using val dataloader with 0 len.

🔧 Affected Symbols

AsyncCheckpointIOLearningRateFinderDeepSpeedstrategyRichProgressBarget_available_flops