Changes in 2.6.4

PyTorch Lightning

Removed

• Removed support for Neptune logger (#21572).

Note: We usually don't remove features in a patch release, however in this case it's an exception since even without removing it the integration would be broken due to Neptune being acquired and sunsetting public service

Changed

• Raised minimum LitLogger version to 2026-03-17 (#21591)

Fixed

• Fixed val_check_interval raising ValueError when limit_val_batches=0 and interval exceeds training batches (#21560)
• Fixed FSDP mixed precision (bf16-mixed, 16-mixed) initializing model parameters in half precision instead of fp32 (#21586)
• Fixed device_mesh type hint in FSDPStrategy to accept a 2-element tuple via the CLI (#21581)
• Fixed RichModelSummary model size display formatting (#21467)
• Fixed floating-point precision in SimpleProfiler duration aggregation by using math.fsum (#21525)

Lightning Fabric

Fixed

• Fixed FSDP mixed precision (bf16-mixed, 16-mixed) initializing model parameters in half precision instead of fp32 (#21586)
• Fixed device_mesh type hint in FSDPStrategy to accept a 2-element tuple via the CLI (#21581)

Full commit list: 2.6.1 -> 2.6.4

Contributors

New Contributors

• @AsherJingkongChen made their first contribution in #21467
• @ManasVardhan made their first contribution in #21560

We thank all folks who submitted issues, features, fixes and doc changes. It's the only way we can collectively make Lightning ⚡ better for everyone, nice job!

In particular, we would like to thank the authors and co-authors of the pull-requests above:

@AsherJingkongChen, @Borda, @bhimrazy, @deependujha, @justusschock, @littlebullGit, @ManasVardhan, @taha-yassine

Changes in 2.6.1

PyTorch Lightning

Added

• Added method chaining support to LightningModule.freeze() and LightningModule.unfreeze() by returning self (#21469)
• Added litlogger integration(#21430)

Deprecated

• Deprecated to_torchscript method due to deprecation of TorchScript in PyTorch (#21397)

Removed

• Removed support for Python 3.9 due to end-of-life status (#21398)

Fixed

• Fixed save_hyperparameters(ignore=...) behavior so subclass ignore rules override base class rules (#21490)
• Fixed LightningDataModule.load_from_checkpoint to restore the datamodule subclass and hyperparameters (#21478)
• Fixed ModelParallelStrategy single-file checkpointing when torch.compile wraps the model so optimizer states no longer raise KeyError during save (#21357)
• Sanitize profiler filenames when saving to avoid crashes due to invalid characters (#21395)
• Fixed StochasticWeightAveraging with infinite epochs (#21396)
• Fixed _generate_seed_sequence_sampling function not producing unique seeds (#21399)
• Fixed ThroughputMonitor callback emitting warnings too frequently (#21453)

Lightning Fabric

Added

• Exposed weights_only argument for loading checkpoints in Fabric.load() and Fabric.load_raw() (#21470)

Fixed

• Fixed DistributedSamplerWrapper not forwarding set_epoch to the underlying sampler (#21454)
• Fixed DDP notebook CUDA fork check to allow passive initialization when CUDA is not actively used (#21402)

Full commit list: 2.6.0 -> 2.6.1

Contributors

New Contributors

• @arrdel made their first contribution in #21402
• @CodeVishal-17 made their first contribution in #21470
• @aditya0by0 made their first contribution in #21478

We thank all folks who submitted issues, features, fixes and doc changes. It's the only way we can collectively make Lightning ⚡ better for everyone, nice job!

In particular, we would like to thank the authors of the pull-requests above

Changes in 2.6.0

PyTorch Lightning

Added

• Added WeightAveraging callback that wraps the PyTorch AveragedModel class (#20545)
• Added Torch-Tensorrt integration with LightningModule (#20808)
• Added time-based validation support though val_check_interval (#21071)
• Added attributes to access stopping reason in EarlyStopping callback (#21188)
• Added support for variable batch size in ThroughputMonitor (#20236)
• Added EMAWeightAveraging callback that wraps Lightning's WeightAveraging class (#21260)

Changed

• Expose weights_only argument for Trainer.{fit,validate,test,predict} and let torch handle default value (#21072)
• Default to RichProgressBar and RichModelSummary if the rich package is available. Fallback to TQDMProgressBar and ModelSummary otherwise (#20896)
• Add MPS accelerator support for mixed precision (#21209)

Fixed

• Fixed edgecase when max_trials is reached in Tuner.scale_batch_size (#21187)
• Fixed case where LightningCLI could not be initialized with trainer_default containing callbacks (#21192)
• Fixed missing reset when ModelPruning is applied with lottery ticket hypothesis (#21191)
• Fixed preventing recursive symlink creation iwhen save_last='link' and save_top_k=-1 (#21186)
• Fixed last.ckpt being created and not linked to another checkpoint (#21244)
• Fixed bug that prevented BackboneFinetuning from being used together with LearningRateFinder (#21224)
• Fixed ModelPruning sparsity logging bug that caused incorrect sparsity percentages (#21223)
• Fixed LightningCLI loading of hyperparameters from ckpt_path failing for subclass model mode (#21246)
• Fixed check the init args only when the given frames are in __init__ method (#21227)
• Fixed how ThroughputMonitor calculated training time (#21291)
• Fixed synchronization of gradients in manual optimization with DDPStrategy(static_graph=True) (#21251)
• Fixed FSDP mixed precision semantics and added user warning (#21361)

Lightning Fabric

Changed

• Expose weights_only argument for Trainer.{fit,validate,test,predict} and let torch handle default value (#21072)
• Set _DeviceDtypeModuleMixin._device from torch's default device function (#21164)
• Added kwargs-filtering for Fabric.call to support different callback method signatures (#21258)

Fixed

• Fixed issue in detecting MPIEnvironment with partial mpi4py installation (#21353)
• Learning rate scheduler is stepped at the end of epoch when on_train_batch_start returns -1 (#21296).
• Fixed FSDP mixed precision semantics and added user warning (#21361)

Full commit list: 2.5.0 -> 2.6.0

Contributors

We thank all folks who submitted issues, features, fixes and doc changes. It's the only way we can collectively make Lightning ⚡ better for everyone, nice job!

In particular, we would like to thank the authors of the pull-requests above

Changes in 2.5.6

<a name="changelog-pytorch"></a>

PyTorch Lightning

<details open><summary>Changed</summary>

• Add name() function to accelerator interface ((#21325))

</details>

<details open><summary>Removed</summary>

• Remove support for deprecated and archived lightning-habana package (#21327)

</details>

Changes in 2.5.5

<a name="changelog-pytorch"></a>

PyTorch Lightning

<details open><summary>Changed</summary>

• Include exclude_frozen_parameters to DeepSpeedStrategy (#21060)
• Include PossibleUserWarning that is raised if modules are in eval mode when training starts (#21146)

</details>

<details open><summary>Fixed</summary>

• Fixed LightningCLI not using ckpt_path hyperparameters to instantiate classes (#21116)
• Fixed callbacks by defer step/time-triggered ModelCheckpoint saves until validation metrics are available (#21106)
• Fixed with adding a missing device id for pytorch 2.8 (#21105)
• Fixed TQDMProgressBar not resetting correctly when using both a finite and iterable dataloader (#21147)
• Fixed cleanup of temporary files from Tuner on crashes (#21162)

</details>

<a name="changelog-fabric"></a>

Lightning Fabric

<details open><summary>Changed</summary>

• Include exclude_frozen_parameters to DeepSpeedStrategy (#21060)
• Let _get_default_process_group_backend_for_device support more hardware platforms ( #21057, #21093)

</details>

<details open><summary>Fixed</summary>

• Fixed with adding a missing device id for pytorch 2.8 (#21105)
• Respecting verbose=False in seed_everything when no seed is provided (#21161)

</details>

</br>

Full commit list: 2.5.4 -> 2.5.5

<a name="contributors"></a>

Contributors

We thank all folks who submitted issues, features, fixes and doc changes. It's the only way we can collectively make Lightning :zap: better for everyone, nice job!

In particular, we would like to thank the authors of the pull-requests above, in no particular order:

@Borda, @KAVYANSHTYAGI, @littlebullGit, @mauvilsa, @SkafteNicki, @taozhiwei

Thank you :heart: and we hope you'll keep them coming!

Changes in 2.5.4

<a name="changelog-pytorch"></a>

PyTorch Lightning

<details open><summary>Fixed</summary>

• Fixed AsyncCheckpointIO snapshots tensors to avoid race with parameter mutation (#21079)
• Fixed AsyncCheckpointIO threadpool exception if calling fit or validate more than one (#20952)
• Fixed learning rate not being correctly set after using LearningRateFinder callback (#21068)
• Fixed misalignment column while using rich model summary in DeepSpeedstrategy (#21100)
• Fixed RichProgressBar crashing when sanity checking using val dataloader with 0 len (#21108)

</details>

<a name="changelog-fabric"></a>

Lightning Fabric

<details open><summary>Changed</summary>

• Added support for NVIDIA H200 GPUs in get_available_flops (#20913)

</details>

</br>

Full commit list: 2.5.3 -> 2.5.4

<a name="contributors"></a>

Contributors

We thank all folks who submitted issues, features, fixes and doc changes. It's the only way we can collectively make Lightning :zap: better for everyone, nice job!

In particular, we would like to thank the authors of the pull-requests above, in no particular order:

@fnhirwa, @GdoongMathew, @jjh42, @littlebullGit, @SkafteNicki

Thank you :heart: and we hope you'll keep them coming!

Notable changes in this release

<a name="changelog-pytorch"></a>

PyTorch Lightning

<details open><summary>Changed</summary>

• Added save_on_exception option to ModelCheckpoint Callback (#20916)
• Allow dataloader_idx_ in log names when add_dataloader_idx=False (#20987)
• Allow returning ONNXProgram when calling to_onnx(dynamo=True) (#20811)
• Extended support for general mappings being returned from training_step when using manual optimization (#21011)

</details>

<details open><summary>Fixed</summary>

• Fixed Allowing trainer to accept CUDAAccelerator instance as accelerator with FSDP strategy (#20964)
• Fixed progress bar console clearing for Rich 14.1+ (#21016)
• Fixed AdvancedProfiler to handle nested profiling actions for Python 3.12+ (#20809)
• Fixed rich progress bar error when resume training (#21000)
• Fixed double iteration bug when resumed from a checkpoint. (#20775)
• Fixed support for more dtypes in ModelSummary (#21034)
• Fixed metrics in RichProgressBar being updated according to user provided refresh_rate (#21032)
• Fixed save_last behavior in the absence of validation (#20960)
• Fixed integration between LearningRateFinder and EarlyStopping (#21056)
• Fixed gradient calculation in lr_finder for mode="exponential" (#21055)
• Fixed save_hyperparameters crashing with dataclasses using init=False fields (#21051)

</details>

<a name="changelog-fabric"></a>

Lightning Fabric

<details open><summary>Changed</summary>

• Enable "auto" for devices and accelerator as CLI arguments (#20913)
• Raise ValueError when seed is out-of-bounds or cannot be cast to int (#21029)

</details>

<details open><summary>Fixed</summary>

• Fixed remove extra name parameter in accelerator registry decorator (#20975)
• Fixed XLA strategy to add support for global_ordinal, local_ordinal, world_size which came instead of deprecated methods (#20852)

</details>

</br>

Full commit list: 2.5.2 -> 2.5.3

<a name="contributors"></a>

Contributors

We thank all folks who submitted issues, features, fixes and doc changes. It's the only way we can collectively make Lightning :zap: better for everyone, nice job!

In particular, we would like to thank the authors of the pull-requests above, in no particular order:

@baskrahmer, @bhimrazy, @deependujha, @fnhirwa, @GdoongMathew, @jonathanking, @relativityhd, @rittik9, @SkafteNicki, @sudiptob2, @vsey, @YgLK

Thank you :heart: and we hope you'll keep them coming!

Notable changes in this release

<a name="changelog-pytorch"></a>

PyTorch Lightning

<details open><summary>Changed</summary>

• Add toggled_optimizer(optimizer) method to the LightningModule, which is a context manager version of toggle_optimize and untoggle_optimizer (#20771)
• For cross-device local checkpoints, instruct users to install fsspec>=2025.5.0 if unavailable (#20780)
• Check param is of nn.Parameter type for pruning sanitization (#20783)

</details>

<details open><summary>Fixed</summary>

• Fixed save_hyperparameters not working correctly with LightningCLI when there are parsing links applied on instantiation (#20777)
• Fixed logger_connector has an edge case where step can be a float (#20692)
• Fixed Synchronize SIGTERM Handling in DDP to Prevent Deadlocks (#20825)
• Fixed case-sensitive model name (#20661)
• CLI: resolve jsonargparse deprecation warning (#20802)
• Fix: move check_inputs to the target device if available during to_torchscript (#20873)
• Fixed progress bar display to correctly handle iterable dataset and max_steps during training (#20869)
• Fixed problem for silently supporting jsonnet (#20899)

</details>

<a name="changelog-fabric"></a>

Lightning Fabric

<details open><summary>Changed</summary>

• Ensure correct device is used for autocast when mps is selected as Fabric accelerator (#20876)

</details>

<details open><summary>Removed</summary>

• Fix: TransformerEnginePrecision conversion for layers with bias=False (#20805)

</details>

</br>

Full commit list: 2.5.1 -> 2.5.2

<a name="contributors"></a>

Contributors

We thank all folks who submitted issues, features, fixes, and doc changes. It's the only way we can collectively make Lightning :zap: better for everyone, nice job!

In particular, we would like to thank the authors of the pull-requests above, in no particular order:

@adamjstewart, @Armannas, @bandpooja, @Borda, @chanokin, @duydl, @GdoongMathew, @KAVYANSHTYAGI, @mauvilsa, @muthissar, @rustamzh, @siemdejong

Thank you :heart: and we hope you'll keep them coming!

Full Changelog: https://github.com/Lightning-AI/pytorch-lightning/compare/2.5.1...2.5.1.post0

Changes

<a name="changelog-pytorch"></a>

PyTorch Lightning

<details open><summary>Changed</summary>

• Allow LightningCLI to use a customized argument parser class (#20596)
• Change wandb default x-axis to tensorboard's global_step when sync_tensorboard=True (#20611)
• Added a new checkpoint_path_prefix parameter to the MLflow logger which can control the path to where the MLflow artifacts for the model checkpoints are stored (#20538)
• CometML logger was updated to support the recent Comet SDK (#20275)
• bump: testing with latest torch 2.6 (#20509)

</details>

<details open><summary>Fixed</summary>

• Fixed CSVLogger logging hyperparameter at every write which increases latency (#20594)
• Fixed OverflowError when resuming from checkpoint with an iterable dataset (#20565)
• Fixed swapped _R_co and _P to prevent type error (#20508)
• Always call WandbLogger.experiment first in _call_setup_hook to ensure tensorboard logs can sync to wandb (#20610)
• Fixed TBPTT example (#20528)
• Fixed test compatibility as AdamW became a subclass of Adam (#20574)
• Fixed file extension of model checkpoints uploaded by NeptuneLogger (#20581)
• Reset trainer variable should_stop when fit is called (#19177)
• Fixed making WandbLogger upload models from all ModelCheckpoint callbacks, not just one (#20191)
• Error when logging to MLFlow deleted experiment (#20556)

</details>

<a name="changelog-fabric"></a>

Lightning Fabric

<details open><summary>Changed</summary>

• Added logging support for a list of dicts without collapsing to a single key (#19957)
• bump: testing with latest torch 2.6 (#20509)

</details>

<details open><summary>Removed</summary>

• Removed legacy support for lightning run model; use fabric run instead. (#20588)

</details>

</br>

Full commit list: 2.5.0 -> 2.5.1

<a name="contributors"></a>

Contributors

We thank all folks who submitted issues, features, fixes and doc changes. It's the only way we can collectively make Lightning :zap: better for everyone, nice job!

In particular, we would like to thank the authors of the pull-requests above, in no particular order:

@benglewis, @Borda, @cgebbe, @duydl, @haifeng-jin, @japdubengsub, @justusschock, @lantiga, @mauvilsa, @millskyle, @ringohoffman, @ryan597, @senarvi, @TresYap

Thank you :heart: and we hope you'll keep them coming!