Repository Issue Activity (beta)

#21895 Many time passes between epochs

#21915 Missing objects.inv in docs

#21913 PyPI release 2.6.5 still vulnerable to CVE-2026-58659 (RCE via load_from_checkpoint)

#21911 [mobile iOS Safari] Hyperlinks from the main docs page all give "Version not found"

#21908 mypy fails on unchanged code after `types-PyYAML` bump

#21900 Deprecation warnings in lightning.pytorch.cli with jsonargparse 4.49+

#21905 Unified storage_options support for FSDPStrategy, ModelParallelStrategy, and TorchCheckpointIO

#21903 scale_batch_size runs zero batches per trial when started after training has begun

#21901 BaseFinetuning re-freezes already unfrozen modules when resuming from a checkpoint

#21865 compatibility with rich 15.0.0

#21800 Checkpoints silently fail to save due to swallowed PermissionError in `_atomic_save`

#20456 [BUG] (`DataLoader`) sanity check fails due to `Input type (torch.FloatTensor) and weight type (torch.cuda.FloatTensor)`

#21893 `CombinedLoader(max_size_cycle)` stops immediately when one iterable is empty

#21808 `accumulate_grad_batches` is silently ignored when epoch length < `accumulate_grad_batches`, because Lightning always steps on the last batch

#21809 Add an end-to-end anomaly detection tutorial using PyTorch Lightning

#21842 ModelCheckpoint: manual-optimization pre-step snapshot is not applied when the save is deferred to on_validation_end

#21725 CUDAAccelerator.setup_device(device) initializes a different device (GPU 0)

#21812 Enable FSDP strategy on CPU (accelerator="cpu")

#21868 Optimize remote checkpoint loading with parallel multiprocess downloads and zero-copy mmap

#21813 `ModelCheckpoint` deletes *previous run's* checkpoint when remote filesystem

#21814 Device mesh initialization fails when running FSDP on CPU

#21815 Mixed precision autocast is disabled when running FSDP on CPU

#21825 Support remote/cloud storage (fsspec) for checkpoint consolidation

#21861 Add REFUTE scientific critique + calibration benchmark

#21838 sync_dist mean-reduced epoch metrics are inflated when ranks log unequal batch counts (integer batch-size accumulator floors during mean-sync)