All pull requests

Pull requests

• fix(grpo): normalize Liger DAPO across gradient accumulation

  #10256·taking-lying-flat opened Sep 29, 2026Contributor·

  ·

• fix(agent_template): render tool call arguments as JSON in ReAct and GLM4 templates

  #10254·MohammadHijjawi97 opened Sep 28, 2026·

  ·

• fix(utils): treat a blank or non-numeric launch env var as unset

  discussion
  #10251·Lesereingrape opened Sep 27, 2026Contributor·

  ·

• [bugfix] drop dp_cp_group from content_metadata when saving mcore checkpoint

  discussionpending
  #10242·WhichPaths opened Sep 25, 2026·

  ·

• feat(megatron): add profiling support with pytorch/nsys profiler and memory history

  postpone
  #10237·HowardZorn opened Sep 23, 2026·

  ·
  5

• feat: add Moore Threads MUSA backend support via torchada

  #10224·caizhi-mt opened Sep 21, 2026·

  ·
  17

• fix(infer): align TransformersEngine with doc defaults by propagating…

  invalid
  #10222·DreamBinary opened Sep 21, 2026·

  ·
  1

• Fix/tool user followup root

  postpone
  #10201·tastelikefeet opened Sep 19, 2026Collaborator·

  ·

• fix(rlhf): percent-encode the IPv6 zone id in rollout base URLs

  invalidpending
  #10200·Redemption-ZTX opened Sep 19, 2026·

  ·

• fix(template): preserve native user follow-ups after tool results

  #10198·Excelius-Wang opened Sep 18, 2026Contributor·

  ·
  2

• [Feat] [FSDP2] [SFT] [ExpertParallel] [DeepSeekV4] fsdp2 dpsk v4 flash 0731 support on only 1Node with 8*H200

  postpone
  #10197·mugglewei97 opened Sep 18, 2026·

  ·

• Readme update

  #10195·indigoRM opened Sep 18, 2026·

  ·
  1

• Feature/twinkle kernel

  #10194·indigoRM opened Sep 18, 2026·

  ·
  1

• feat(sft): support logits selection with Ulysses sequence parallelism

  pending
  #10191·Excelius-Wang opened Sep 17, 2026Contributor·

  ·

• fix(sp): preserve packed boundaries when padding auxiliary fields

  #10180·Excelius-Wang opened Sep 17, 2026Contributor·

  ·

• fix(template): align multimodal fields and handle text-only inputs

  postpone
  #10176·taking-lying-flat opened Sep 17, 2026Contributor·

  ·
  1

• fix(gkd): promote low-precision logits before loss computation

  #10175·taking-lying-flat opened Sep 16, 2026Contributor·

  ·
  1

• fix(megatron): honor DPO loss type and normalize IPO logprobs

  #10167·taking-lying-flat opened Sep 16, 2026Contributor·

  ·

• fix(template): reject samples that lose all supervision during truncation

  #10166·Excelius-Wang opened Sep 16, 2026Contributor·

  ·

• fix(gkd): normalize HF loss across accumulation and DDP

  #10165·taking-lying-flat opened Sep 16, 2026Contributor·

  ·

• fix(train): serialize GaLore projector state in checkpoints

  #10161·MrCapricornLiu opened Sep 16, 2026Contributor·

  ·
  1

• fix(dataset): invalidate metadata cache on registration

  #10160·MrCapricornLiu opened Sep 16, 2026Contributor·

  ·
  1

• fix(train): distinguish activation views during CPU offload

  #10157·MrCapricornLiu opened Sep 16, 2026Contributor·

  ·
  1

• feat(megatron): add profiler support

  #10155·lixin-sxty opened Sep 16, 2026·

  ·

• fix(dataset): handle failed non-streaming packing workers

  #10152·Excelius-Wang opened Sep 15, 2026Contributor·

  ·