All pull requests

Pull requests

• Fix scheduler stepping too fast with tensor/context parallelism

  #4497·akanyaani opened Nov 3, 2026·2 of 5 tasks·

  ·

• Handle unknown batch size in sharded dataloader metadata

  #4491·Aurantium-47 opened Nov 2, 2026·3 of 4 tasks·

  ·

• Fix GradientAccumulationPlugin.adjust_scheduler being silently ignored

  #4490·simpleqt opened Nov 1, 2026Contributor·5 tasks·

  ·
  2

• Fix gather_for_metrics dropping or duplicating samples from sized samplers

  #4489·simpleqt opened Nov 1, 2026Contributor·3 of 5 tasks·

  ·
  1

  6

• Fix saved boolean launch configuration precedence

  #4487·linjiaye929 opened Oct 31, 2026·2 of 5 tasks·

  ·

• Skip complete dispatched batches without changing grouping or tail padding

  #4486·anishpentyala opened Oct 31, 2026·3 of 4 tasks·

  ·

• Honor the source rank when broadcasting nested tensors on XLA

  #4485·Linxiushen opened Oct 31, 2026Contributor·2 of 4 tasks·

  ·

• Fix scalar tensor reconstruction in initialize_tensors

  #4484·DiadkoShmek opened Oct 31, 2026·3 tasks done·

  ·
  1

• Add RunyardTracker: log_with="runyard"

  #4483·cobanov opened Oct 31, 2026·

  ·
  1

• Fix stateful dataloader replaying the finished epoch after a boundary checkpoint

  #4482·anishpentyala opened Oct 31, 2026·

  ·
  1

  1

• restore dataloader epoch when resuming a stateful dataloader

  #4481·dundysm opened Oct 31, 2026·3 tasks done·

  ·
  1

  2

• Fix dead links to removed transformers examples in the docs

  #4479·sarankumar1325 opened Oct 30, 2026·1 task done·

  ·

• chore(deps): bump the actions group across 1 directory with 6 updates

  dependenciesgithub_actions
  #4478·dependabot[bot] opened Oct 30, 2026Bot·

  ·
  1

• FSDP2 cpu_ram_efficient_loading: honor ignored modules and load buffers

  #4475·eilamc14 opened Oct 30, 2026·3 of 5 tasks·

  ·
  1

  1

• Fix typos in docs, examples and comments

  #4473·sarankumar1325 opened Oct 29, 2026·1 task done·

  ·

• Preserve single-tensor inputs when clipping gradients

  #4471·Aurantium-47 opened Oct 29, 2026·2 of 5 tasks·

  ·
  1

• Fix reduce(..., "mean") on integer tensors in distributed runs

  #4470·MaiAditya opened Oct 29, 2026·4 of 5 tasks·

  ·
  1

  1

• Handle iterable loaders safely when joining uneven inputs

  #4468·Linxiushen opened Oct 27, 2026Contributor·

  ·
  1

• Clean up temporary checkpoint offloads when loading fails

  #4467·Linxiushen opened Oct 27, 2026Contributor·3 of 4 tasks·

  ·
  1

• Cap the CPU budget of get_max_memory at the cgroup memory limit

  #4466·prabhkodes opened Oct 27, 2026·4 of 5 tasks·

  ·
  1

  1

• Enforce execute_subprocess_async timeout and fail on signal kills

  #4458·SunMarc opened Oct 27, 2026Member·

  ·
  1

• Release first-process ordering barriers when the block raises

  #4457·Aurantium-47 opened Oct 27, 2026·2 of 3 tasks·

  ·
  1

• fix: support windows path separators when sorting automatic checkpoints (#4454)

  #4455·aakash1552005 opened Oct 25, 2026·

  ·
  1

  2

• Fix split-batch stateful dispatcher resume

  #4453·AryanRG opened Oct 25, 2026·3 of 5 tasks·

  ·
  1

  1

• Preserve no_grad when composing model hooks

  #4451·Ra1nhy opened Oct 24, 2026·3 of 4 tasks·

  ·
  1