PyTorch Lightning Spells

This package contains some useful plugins for PyTorch Lightning.

Disclaimer: This package is a third-party extension for PyTorch Lightning. We are not affiliated with the PyTorch Lightning project or the company behind it.

Documentation

Installation

PIP

pip install pytorch-lightning-spells

or the latest code in the main branch:

pip install https://github.com/veritable-tech/pytorch-lightning-spells/archive/main.zip

UV

uv add pytorch-lightning-spells

or the latest code in the main branch:

uv add git+https://github.com/veritable-tech/pytorch-lightning-spells.git@main

Release Notes

0.2.1 (2025-10-08)

A small patch focused on clarity, type-safety, and minor robustness fixes. No public API changes; runtime behavior is preserved.

• Metrics
  • Rename internal state target → targets for clarity (internal only).
  • Add guidance/warnings for AUC and FBeta specialized behavior; refactor compute paths with safer typing.
  • Minor cleanup in SpearmanCorrelation.
• Callbacks
  • TelegramCallback: use event loop run_until_complete, re-enable on_exception, skip sanity validation, read from trainer.logged_metrics, and handle TimedOut gracefully.
  • Lookahead callbacks: assert/operate only on Lookahead optimizers to prevent misuse.
  • RandomAugmentationChoiceCallback: fix typing for p (Sequence[float]) and use random.choices.
• Loggers
  • ScreenLogger: sanitize and pretty-print hyperparams/metrics to ensure JSON-serializable output.
• LR Schedulers
  • Migrate base and multistage schedulers to torch.optim.lr_scheduler.LRScheduler; ensure initial LR setup via _initial_step.
• CI
  • Fix uv sync command and minor workflow hygiene.

Compatibility notes

• No breaking changes expected. Only the internal metric state name changed; user code depending on internal buffers should update from target to targets.

0.2.0 (2025-09-03)

• This release modernizes the project's tooling, dependency management, and testing infrastructure.
  • The build system has been migrated from setup.py to the standard pyproject.toml, and the CI/CD pipeline has been overhauled to use uv for significantly faster and more reliable dependency management.
  • It also introduces comprehensive tests for the callback modules, improving code reliability. It also includes several bug fixes and enhancements to existing callbacks, most notably making LookaheadModelCheckpoint more flexible and robust.
• TelegramCallback: Update the code to work with the latest version of the python-telegram-bot package.
• The GlobalMetric class has been updated to align with recent changes in torchmetrics, removing the deprecated compute_on_step parameter.