Changelog¶

Version 2.4.0 (2024-10-02)¶

New Features

Added strategy="fsdp_native" for PyTorch 2.0+ native FSDP with activation checkpointing.
New StochasticWeightAveraging callback.
Added Trainer(barebones=True) mode — disables all non-essential features for maximum throughput.
LightningModule.log_dict() now supports nested dict keys (e.g., "metrics/train/loss").
New NeptuneLogger for Neptune.ai integration.
Trainer(compile=True) shorthand to compile the model with torch.compile().
Async checkpointing: checkpoints are now saved in a background thread by default.

Breaking Changes

Dropped Python 3.8 support. Minimum is now Python 3.9.
Trainer(gpus=...) and Trainer(tpu_cores=...) removed (deprecated since 2.0). Use devices= instead.
LightningModule.on_post_move_to_device() removed.

Deprecations

Trainer(auto_lr_find=True) deprecated; use from lightning.pytorch.tuner import Tuner instead.

New Features

New Lightning CLI — train models from YAML config files without writing Python:

lightning fit --config config.yaml
lightning validate --config config.yaml --ckpt_path best
lightning predict --config config.yaml --ckpt_path last

Bug Fixes

Fixed gradient accumulation with accumulate_grad_batches > 1 and strategy="fsdp".
Fixed ModelCheckpoint.best_model_path returning None after trainer.test().
Fixed memory leak in RichProgressBar when leave=False.

New Features

LightningModule now supports configure_model() hook for model parallel setups.
Added Trainer(reload_dataloaders_every_n_epochs=N) parameter.
New lightning.pytorch.utilities.grad_norm utility function.

Major Release

Complete rewrite of the Trainer internals.
Minimum PyTorch bumped to 2.0.
Introduced lightning.fabric.Fabric — a lightweight alternative to Trainer for fine-grained control.
New plugin system for custom accelerators and strategies.
DeepSpeed strategy updated to DeepSpeed 0.12.
Unified lightning package combining pytorch-lightning and lightning-fabric.