24 Pull requests merged by 9 people

• build(deps): bump Lightning-AI/utilities from 0.14.3 to 0.15.0

  #21010 merged Aug 4, 2025

• fix ci: progress bar console clearing for latest Rich release

  #21016 merged Aug 4, 2025

• build(deps): update torchmetrics requirement from <1.8.0,>=0.10.0 to >=0.10.0,<1.9.0 in /requirements

  #21006 merged Aug 4, 2025

• build(deps): update lightning-utilities requirement from <0.15.0,>=0.11.1 to >=0.11.1,<0.16.0 in /requirements

  #21005 merged Aug 4, 2025

• build(deps): update docutils requirement from <=0.19,>=0.18.1 to >=0.18.1,<=0.22 in /requirements

  #21027 merged Aug 4, 2025

• build(deps): update awscli requirement from <1.42.0,>=1.30.0 to >=1.30.0,<1.43.0 in /requirements

  #21026 merged Aug 4, 2025

• build(deps): bump mypy from 1.17.0 to 1.17.1 in /requirements

  #21025 merged Aug 4, 2025

• build(deps): bump coverage from 7.9.2 to 7.10.2 in /requirements

  #21024 merged Aug 4, 2025

• fix: rich progress bar error when resume training

  #21000 merged Jul 31, 2025

• fix broken links to studios

  #21014 merged Jul 28, 2025

• fix CI: awscli docutils version conflict

  #20997 merged Jul 26, 2025

• docs: updating flaking links

  #20980 merged Jul 23, 2025

• build(deps): update tensorboard requirement from <2.20.0,>=2.9.1 to >=2.9.1,<2.21.0 in /requirements

  #20992 merged Jul 22, 2025

• build(deps): bump mypy from 1.16.1 to 1.17.0 in /requirements

  #20991 merged Jul 22, 2025

• build(deps): update fsspec[http] requirement from <2025.6.0,>=2022.5.0 to >=2022.5.0,<2025.8.0 in /requirements

  #20990 merged Jul 22, 2025

• Allow dataloader_idx_ in log names when add_dataloader_idx=False

  #20987 merged Jul 18, 2025

• fix: failing markdown link test in ci

  #20979 merged Jul 14, 2025

• docs: update ref to latest tutorials

  #20977 merged Jul 14, 2025

• Add support nvcr.io/nvidia/pytorch:25.06-py3

  #20971 merged Jul 11, 2025

• Model checkpointing save_on_train_epoch_end default behavior documentation

  #20931 merged Jul 9, 2025

• Fix: Allow trainer to accept CUDAAccelerator instance as accelerator with FSDP strategy

  #20964 merged Jul 9, 2025

• Add dev env setup guide

  #20961 merged Jul 9, 2025

• build(deps): bump coverage from 7.9.1 to 7.9.2 in /requirements

  #20966 merged Jul 7, 2025

• build(deps): update awscli requirement from <1.41.0,>=1.30.0 to >=1.30.0,<1.42.0 in /requirements

  #20965 merged Jul 7, 2025

10 Pull requests opened by 6 people

• [pre-commit.ci] pre-commit suggestions

  #20968 opened Jul 7, 2025

• fix: remove extra parameter in accelerator registry decorator

  #20975 opened Jul 11, 2025

• Fix MLFlowLogger.save_dir Windows file URI handling (Fixes #20972)

  #20988 opened Jul 19, 2025

• nitpick: add make command to quickly setup the project on `lightning studio`

  #20996 opened Jul 23, 2025

• docker: simplify the docker name with CUDA

  #21001 opened Jul 25, 2025

• add/debug Lit CI [wip]

  #21002 opened Jul 25, 2025

• docs: update mail to developer@lightning.ai

  #21003 opened Jul 26, 2025

• Allow `training_step` in manual optimization to return general mappings

  #21011 opened Jul 28, 2025

• Sync dist clarification and consistency

  #21012 opened Jul 28, 2025

• Fix fabric examples and load_checkpoint hparams ref

  #21013 opened Jul 28, 2025

15 Issues closed by 4 people

• bugs too many

  #20875 closed Aug 4, 2025

• Documentation or main page is not loading [not available in your region]

  #20989 closed Aug 1, 2025

• Rich progress_bar_id is None if restore training state from a step checkpoint

  #21015 closed Jul 31, 2025

• Missing jsonargparse as dependency

  #21018 closed Jul 31, 2025

• Rich progress bar error when resume training

  #20976 closed Jul 31, 2025

• Spend a lot of time to load large ckpt

  #21017 closed Jul 31, 2025

• on_validation_batch_end is not called when Loss is NaN

  #20999 closed Jul 30, 2025

• Allow user to use `dataloader_idx` in log name in `LightningModule.log`

  #20485 closed Jul 18, 2025

• `validation_epoch_end` is still mentionned in the documentation with version >= 2.0.0 while it has been removed from the code

  #20559 closed Jul 17, 2025

• `load_from_checkpoint` returns `None`

  #20607 closed Jul 17, 2025

• Lightning is requiring packaging < 25.0

  #20772 closed Jul 14, 2025

• `ModelCheckpoint`'s argument `save_on_train_epoch_end`'s documentation unclear when value is `None`

  #20781 closed Jul 9, 2025

• Strategy `fsdp` requires a GPU accelerator, but got CUDAAccelerator

  #20957 closed Jul 9, 2025

• Recommend dev setup / support uv

  #20954 closed Jul 9, 2025

• lightning throws an exception on MacOS when the pytorch default device is set

  #20696 closed Jul 7, 2025

16 Issues opened by 15 people

• DDP Strategy Does Not Automatically Shard Batch Sizes Despite Documentation Claims

  #21023 opened Aug 4, 2025

• Trainer parameter limit-train-batches was meant to be per-worker

  #21022 opened Aug 3, 2025

• The difference of Trainer.test with ddp strategy

  #21004 opened Jul 27, 2025

• Remove an unnecessary TODO in `src/lightning/pytorch/loops/fit_loop.py`

  #20998 opened Jul 24, 2025

• Changing `on_step` in `self.log` causes `batch_to_device` to change

  #20995 opened Jul 23, 2025

• Support BatchSizeFinder in DDP

  #20994 opened Jul 22, 2025

• Accept `TensorDict` (or more generally, dict-like's) as a `training_step` return type

  #20993 opened Jul 22, 2025

• When the model is in an eval state before calling trainer.fit it should be moved to train state

  #20986 opened Jul 18, 2025

• Doing full validation on step 0

  #20985 opened Jul 17, 2025

• uv for CI

  #20984 opened Jul 16, 2025

• MoE (mixture of experts) support for expert parallel

  #20982 opened Jul 15, 2025

• Accelerator registry decorator usage fails with TypeError due to incorrect function signature

  #20973 opened Jul 11, 2025

• MLFlowLogger.save_dir mishandles absolute file: URIs on Windows

  #20972 opened Jul 10, 2025

• Proper way to use mixed precision with manual optimization

  #20970 opened Jul 9, 2025

• Recommend uv commands for development scripts

  #20969 opened Jul 9, 2025

• Improve Fault Tolerance via TorchFT

  #20967 opened Jul 7, 2025

72 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• Fix double iteration bug when resumed from a checkpoint.

  #20775 commented on Aug 4, 2025 • 6 new comments

• Generic weight averaging callback that supports EMA

  #20545 commented on Jul 17, 2025 • 1 new comment

• Added warmup parameter to early stopping cb

  #20778 commented on Jul 19, 2025 • 0 new comments

• [Not finished] Allow customized parameter grouping for automatic optimzier configuration.

  #20742 commented on Jul 19, 2025 • 0 new comments

• fix: `overfit_batches` uses same batch for train and val

  #20731 commented on Jul 19, 2025 • 0 new comments

• Cpu memory accumulation bug

  #20730 commented on Jul 19, 2025 • 0 new comments

• docs(LightningModule): update docs for `.training` mode in loops

  #20716 commented on Jul 19, 2025 • 0 new comments

• FabricModule: wrap forward methods instead of monkeypatch-based redirect

  #20711 commented on Aug 4, 2025 • 0 new comments

• Added support for flushing Comet experiment data to the Comet after saving a checkpoint.

  #20680 commented on Jul 7, 2025 • 0 new comments

• feat[logger] update mlflow limit for parameters length log

  #20636 commented on Jul 16, 2025 • 0 new comments

• fix(mlflow): Enabling multiple callbacks for checkpoint reporting

  #20585 commented on Jul 19, 2025 • 0 new comments

• updated `ModelCheckpoint` to add the facility of retaining periodic checkpoints

  #20547 commented on Jul 10, 2025 • 0 new comments

• Add Deepspeed Zero 3 MiCS support (Issues #20378)

  #20461 commented on Jul 19, 2025 • 0 new comments

• Add `best_k_metrics` parameter to the `ModelCheckpoint`

  #20457 commented on Jul 19, 2025 • 0 new comments

• Call configure_module before freeze_before_training

  #20428 commented on Jul 19, 2025 • 0 new comments

• [Backend]: Support device backend registration for a wide range of third-party hardware

  #20349 commented on Jul 24, 2025 • 0 new comments

• Add compile_fn parameter for Trainer

  #20269 commented on Jul 19, 2025 • 0 new comments

• Feat: support reusable instance of `ModelCheckpoint`

  #20202 commented on Jul 19, 2025 • 0 new comments

• Fix `save_last` behavior in absence of validation

  #20960 commented on Jul 7, 2025 • 0 new comments

• Make asyncio checkpointing work if validate/fit is called more than once

  #20952 commented on Aug 4, 2025 • 0 new comments

• docs(csv_logs): Clarify CSV and YAML logging distinction and improve examples

  #20951 commented on Jul 14, 2025 • 0 new comments

• update ModelSummary

  #20945 commented on Jul 26, 2025 • 0 new comments

• Fix wrong behavior of `DDPStrategy` option with simple GAN training using DDP

  #20936 commented on Jul 19, 2025 • 0 new comments

• Fix: `no_grad` with AMP bug

  #20921 commented on Jul 7, 2025 • 0 new comments

• Add `save_on_exception` option to `ModelCheckpoint`

  #20916 commented on Aug 2, 2025 • 0 new comments

• feat: Default to RichProgressBar and RichModelSummary if rich is avai…

  #20896 commented on Aug 4, 2025 • 0 new comments

• DOC: Clarify DeviceStatsMonitor logged metrics

  #20895 commented on Jul 19, 2025 • 0 new comments

• bugfix: add support for `global_ordinal`, `local_ordinal`, `world_size` in xla

  #20872 commented on Aug 4, 2025 • 0 new comments

• PR: Fix Duplicate Metric Logging in MLFlowLogger to Prevent MLflow Database Errors

  #20871 commented on Jul 23, 2025 • 0 new comments

• Add documentation warning: Don’t use torch.profiler.profile context manager around Trainer methods

  #20864 commented on Jul 19, 2025 • 0 new comments

• Fix: Respect `required=False` in `add_lightning_class_args` when `subclass_mode=False`

  #20856 commented on Jul 19, 2025 • 0 new comments

• Add Callback for Opacus integration

  #20853 commented on Jul 19, 2025 • 0 new comments

• to_onnx return ONNXProgram

  #20811 commented on Aug 4, 2025 • 0 new comments

• Fix advanced profiler for python >=3.12

  #20809 commented on Aug 4, 2025 • 0 new comments

• Torch-Tensorrt Integration with LightningModule

  #20808 commented on Jul 23, 2025 • 0 new comments

• Support `grad_clip_norm_()` for FSDP

  #20784 commented on Jul 19, 2025 • 0 new comments

• Restoring Trainer State with Early Stop fails

  #13225 commented on Jul 19, 2025 • 0 new comments

• ReduceLROnPlateu within configure_optimizers behave abnormally

  #20829 commented on Jul 19, 2025 • 0 new comments

• Support PyTorch/XLA 2.7

  #20852 commented on Jul 19, 2025 • 0 new comments

• The progress bar shows wrong length when using multiple dataloaders mixing dataset and iterable dataset

  #20695 commented on Jul 19, 2025 • 0 new comments

• Cannot call self.log in evaluation_hooks after using trainer.predict, even if using a new trainer object.

  #19101 commented on Jul 19, 2025 • 0 new comments

• Error when learning on tpu

  #20891 commented on Jul 19, 2025 • 0 new comments

• RuntimeError: Bad StatusOr access: UNKNOWN: TPU initialization failed: Invalid --2a886c8_slice_builder_worker_addresses specified. Expected 4 worker addresses, got 1.

  #20244 commented on Jul 19, 2025 • 0 new comments

• Warnings when learning on tpu

  #20890 commented on Jul 19, 2025 • 0 new comments

• Weird bug when setting `val_check_interval` dynamically in `setup()`

  #20894 commented on Jul 19, 2025 • 0 new comments

• Tqdm print multi lines with refresh

  #20909 commented on Jul 19, 2025 • 0 new comments

• Logging in `on_test_epoch_end` with multiple dataloaders

  #20885 commented on Jul 19, 2025 • 0 new comments

• Ignore Keyword Arguments Outside of Callback Signature During `Fabric.call`

  #20915 commented on Jul 19, 2025 • 0 new comments

• Global step reset when restoring checkpoints with trainer.validate

  #17127 commented on Jul 15, 2025 • 0 new comments

• `ModelCheckpoint` not saving best model

  #20657 commented on Jul 13, 2025 • 0 new comments

• MLFlow logger with remote tracking fails with CLI

  #16310 commented on Jul 11, 2025 • 0 new comments

• stateful dataloaders do not load their state_dict if self.trainer.estimated_stepping_batches called beforehand

  #20550 commented on Jul 9, 2025 • 0 new comments

• Inconcistency in loading from checkpoint in LightningCLI

  #20801 commented on Jul 9, 2025 • 0 new comments

• Metrics get mapped twice to the same epoch in MLflow logger

  #20902 commented on Jul 7, 2025 • 0 new comments

• docs: fixed the `init_module` and deepspeed

  #20175 commented on Jul 19, 2025 • 0 new comments

• Call `configure_model` from LightningCLI

  #19111 commented on Jul 19, 2025 • 0 new comments

• deprecation: Is `frequency` key necessary in `lr_scheduler_config`?

  #20714 commented on Aug 4, 2025 • 0 new comments

• PyTorchProfiler: not showing CPU memory used even with `profile_memory=True`

  #20339 commented on Aug 1, 2025 • 0 new comments

• Validation takes place every N time

  #13324 commented on Aug 1, 2025 • 0 new comments

• LearningRateMonitor broken on MPS backend with Apple silicon

  #20250 commented on Jul 31, 2025 • 0 new comments

• Resuming should allow to differentiate what to resume (steps/opti/weights)

  #5339 commented on Jul 29, 2025 • 0 new comments

• Model diverges or struggles to converge with complex-valued tensors in DDP

  #20480 commented on Jul 28, 2025 • 0 new comments

• Gradient accumulation calcluation may be incorrect

  #20350 commented on Jul 28, 2025 • 0 new comments

• Error using wandb when learning on tpu

  #20880 commented on Jul 27, 2025 • 0 new comments

• Light / dark mode for documentation

  #20396 commented on Jul 23, 2025 • 0 new comments

• Parameters and Gradient is not logged by WandB under FSDP strategy

  #17512 commented on Jul 21, 2025 • 0 new comments

• deepspeed strategy can't save checkpoint, TypeError: cannot pickle `torch._C._distributed_c10d.ProcessGroup` object

  #17369 commented on Jul 19, 2025 • 0 new comments

• diff-svc(winerror3 when the training starts)

  #20849 commented on Jul 19, 2025 • 0 new comments

• Fabric FSDP with bitsandbytes plugin is not supported

  #20855 commented on Jul 19, 2025 • 0 new comments

• `add_lightning_class_args` `required` argument ignored if not using subclass mode

  #20851 commented on Jul 19, 2025 • 0 new comments

• Mlflow logging LR duplicate key issue with PostgreSQL DB #190

  #20865 commented on Jul 19, 2025 • 0 new comments

• lightning.fabric.utilities.exceptions.MisconfigurationException: No supported gpu backend found! Maybe latest gpu compability issue..?

  #20626 commented on Jul 19, 2025 • 0 new comments