All pull requests

Pull requests

• Fix sampler handling in skip_first_batches

  #4363·Michael-RDev opened Oct 3, 2026·2 of 5 tasks·

  ·
  1

• Clean up the files of a previous save in Accelerator.save_model for safetensors

  #4362·SulimanAbdulrazzaq opened Oct 3, 2026·1 of 5 tasks·

  ·

• Warn when the checkpoint random states cannot be restored

  #4361·CJstate opened Oct 2, 2026·4 of 5 tasks·

  ·
  1

  1

• Fix debug_launcher on Windows

  #4360·CJstate opened Oct 2, 2026·3 of 5 tasks·

  ·
  1

• Fix gather_tensor_shape discarding zero-sized dimensions and crash in copy_tensor_to_devices

  #4357·Ryanakml opened Sep 30, 2026·

  ·
  1

• FSDP2: preserve tied-parameter aliasing in get_state_dict()

  #4340·UnitedSnakes opened Sep 26, 2026·4 of 5 tasks·

  ·
  1

  2

• Only disable the compiler on the offload path in AlignDevicesHook

  #4327·jiqing-feng opened Sep 24, 2026Contributor·

  ·
  2

• Fix disabled ZeRO-3 initialization contexts

  #4319·1fanwang opened Sep 23, 2026·4 of 5 tasks·

  ·

• Enforce subprocess execution timeouts and clean up interrupted launches

  #4304·Titus-von-Koeller opened Sep 21, 2026Contributor·

  ·
  4

• Fix next-token labels and loss normalization in the accumulation example

  #4303·Titus-von-Koeller opened Sep 21, 2026Contributor·

  ·
  2

• Raise on subprocesses terminated by signals

  #4301·Titus-von-Koeller opened Sep 21, 2026Contributor·

  ·
  3

• Test causal-LM training with DDP, gradient accumulation and mixed precision

  #4300·Titus-von-Koeller opened Sep 21, 2026Contributor·

  ·
  23

• Align Accelerator.clip_grad_norm_ with Transformers

  #4266·michaelbenayoun opened Sep 15, 2026Member·

  ·
  4

• Fix compile_regions breaking ModuleDict and ModuleList indexing

  #4260·hjinnkim opened Sep 14, 2026Contributor·

  ·
  2

• Fix shape metadata when copying tensors across devices

  #4245·MrCapricornLiu opened Sep 10, 2026·

  ·
  2

• Fix stateful dataloader checkpoint cache after loading

  #4244·MrCapricornLiu opened Sep 10, 2026·

  ·
  2

• Fix skipped-step detection for AMP-aware optimizers

  #4243·gss10282023 opened Sep 10, 2026·

  ·
  1

  3

• Add an FSDP2-native Ulysses sequence parallelism backend

  #4209·qgallouedec opened Sep 2, 2026Member·

  ·
  37

• Ask transformers whether a model supports context parallelism

  #4198·qgallouedec opened Aug 31, 2026Member·

  ·
  4

• Allow tensor parallelism next to replicated data parallelism

  #4193·qgallouedec opened Aug 30, 2026Member·5 tasks·

  ·
  2

• Fix stateful dataloader checkpointing across processes

  #4165·jrsmartin opened Aug 14, 2026·4 of 5 tasks·

  ·
  6