Issues

• GradientAccumulationScheduler rejects its own previous assignment on a second fit

  Lightning-AI/pytorch-lightning#22102 · Filter by author anishpentyala opened on Nov 1, 2026

• Repeated compiled-model Trainer fits leave Dynamo wrappers active after to_uncompiled

  Lightning-AI/pytorch-lightning#22101 · Filter by author anishpentyala opened on Nov 1, 2026

• WeightAveraging validation exceptions reverse live/average weights in exception checkpoints

  Lightning-AI/pytorch-lightning#22100 · Filter by author anishpentyala opened on Nov 1, 2026

• Reused WeightAveraging skips fresh-fit updates and restores initial weights

  Lightning-AI/pytorch-lightning#22099 · Filter by author anishpentyala opened on Nov 1, 2026

• ModelPruning lottery-ticket reset aliases the saved initial weights on CPU

  Lightning-AI/pytorch-lightning#22098 · Filter by author anishpentyala opened on Nov 1, 2026

• ThroughputMonitor retains the previous model FLOPs capability when reused

  Lightning-AI/pytorch-lightning#22097 · Filter by author anishpentyala opened on Nov 1, 2026

• LR finder leaves temporary callbacks and search limits after a model exception

  Lightning-AI/pytorch-lightning#22096 · Filter by author anishpentyala opened on Nov 1, 2026

• AsyncCheckpointIO skips the wrapped backend teardown hook

  Lightning-AI/pytorch-lightning#22095 · Filter by author anishpentyala opened on Nov 1, 2026

• AsyncCheckpointIO retains live NumPy arrays while snapshotting model tensors

  Lightning-AI/pytorch-lightning#22086 · Filter by author salma2vec opened on Oct 31, 2026

• Resuming from a checkpoint saved at a max_steps stop on an epoch boundary replays the next epoch's first batch and drops its last

  Lightning-AI/pytorch-lightning#22077 · Filter by author anishpentyala opened on Oct 29, 2026
  2

• Docs version pinning replaces working stable links with unavailable versions

  Lightning-AI/pytorch-lightning#22073 · Filter by author littlebullGit opened on Oct 29, 2026
  3

• Multi GPU training fails with ddp-notebook

  bugver: 2.6.xwaiting on author
  Lightning-AI/pytorch-lightning#22064 · Filter by author lf28 opened on Oct 26, 2026
  2

• Prediction writers receive stale batch indices for IterableDataset

  ver: 2.7.x
  Lightning-AI/pytorch-lightning#22059 · Filter by author AryanRG opened on Oct 25, 2026

• Trainer.estimated_stepping_batches ignores GradientAccumulationScheduler, so schedules sized with it end early

  Lightning-AI/pytorch-lightning#22048 · Filter by author mottopanikeiku opened on Oct 22, 2026

• Fabric experimental API warnings link to a 404 versioning-policy page

  Lightning-AI/pytorch-lightning#22046 · Filter by author littlebullGit opened on Oct 22, 2026

• save_last=True no longer saves the latest state when save_top_k is used (behavior change in 2.6.0, released as a fix)

  bug
  Lightning-AI/pytorch-lightning#22043 · Filter by author potipot opened on Oct 22, 2026
  1

• LearningRateMonitor(log_weight_decay=True) crashes with KeyError for optimizers without weight decay

  ver: 2.7.x
  Lightning-AI/pytorch-lightning#22035 · Filter by author RaphaelFakhri opened on Oct 20, 2026

• ModelCheckpoint overwrites the checkpoint it resumed from when using gradient accumulation

  ver: 2.5.xver: 2.6.xver: 2.7.x
  Lightning-AI/pytorch-lightning#22033 · Filter by author MissieKahsay opened on Oct 19, 2026
  2

• Quick survey: your experience as a PyTorch Ecosystem maintainer

  Lightning-AI/pytorch-lightning#22027 · Filter by author michellearoth opened on Oct 16, 2026

• MixedPrecision globally patched no_grad clears autocast cache in unrelated threads

  ver: 2.7.x
  Lightning-AI/pytorch-lightning#22023 · Filter by author ternaus opened on Oct 12, 2026

• Timer.time_elapsed becomes negative during repeated validation runs

  ver: 2.7.x
  Lightning-AI/pytorch-lightning#22011 · Filter by author aalquwayfili opened on Oct 8, 2026

• make standalone returns success when its test runner fails

  bugver: 2.7.x
  Lightning-AI/pytorch-lightning#22003 · Filter by author mukeshr06 opened on Oct 3, 2026
  3

• EMAWeightAveraging updates the average model on steps skipped by the AMP GradScaler

  Lightning-AI/pytorch-lightning#21990 · Filter by author SeverinVisionary opened on Oct 1, 2026
  1

• Add overloads to Fabric.setup, setup_optimizers, and setup_dataloaders for precise return types

  Lightning-AI/pytorch-lightning#21988 · Filter by author Buddha7771 opened on Oct 1, 2026

• Mixed precision uses torch.autocast("cuda") for custom accelerators and Accelerator instances, silently running fp32

  ver: 2.1.xver: 2.2.xver: 2.3.xver: 2.4.xver: 2.5.xver: 2.6.x
  Lightning-AI/pytorch-lightning#21982 · Filter by author UniversalBasicInc opened on Sep 29, 2026