All pull requests

Pull requests

• fix(examples): preserve tau2 user simulator tool history

  #1759·MaybeIcanShow opened Sep 27, 2026·

  ·
  1

• fix(proxy): assign terminal rewards to completed interactions

  #1755·Le8r0nJames opened Sep 23, 2026Collaborator·

  ·

• refactor(awex-colocate): remove patch awex and peft colocate flow

  #1754·PrometheusComing opened Sep 23, 2026Collaborator·

  ·

• feat(models): support Qwen Flash Next vision and chunked training

  safe-to-test
  #1751·dingzhiqiang opened Sep 22, 2026Collaborator·

  ·
  19

• feat: add prefix replay on-policy distillation

  #1745·fishcrap opened Sep 22, 2026Collaborator·

  ·

• fix(inference): manage PRM scorer lifecycle in the v1 proxy

  #1744·harooos opened Sep 21, 2026Contributor·

  ·

• feat(engine): add distributed Muon optimizer offload

  safe-to-test
  #1743·yulangz opened Sep 21, 2026Collaborator·

  ·
  1

• fix(trainer): preserve M2PO selection across microbatches

  #1732·gss10282025 opened Sep 18, 2026·

  ·
  1

• Fix(distillation): Fixed teacher log-probability alignment with prediction positions

  #1729·zahrayousefijamarani opened Sep 17, 2026Contributor·

  ·
  1

• feat(infra): add opt-in sample-level rollout refill

  safe-to-test
  #1722·dingzhiqiang opened Sep 17, 2026Collaborator·

  ·
  1

• fix(v2): propagate configured training RPC timeout to data proxy

  #1712·sitabulaixizawaluduo opened Sep 15, 2026Collaborator·

  ·
  1

• fix(utils): zero single-element leave-one-out normalization at batch level

  #1706·KunyangZhang opened Sep 13, 2026·

  ·
  1

  1

• feat(dpo): support SimPO reference-free preference optimization with target margin

  #1699·hsusul opened Sep 10, 2026Contributor·

  ·
  1

• feat(reward): support composable reasoning rewards with format verification and length regulation

  #1698·hsusul opened Sep 10, 2026Contributor·

  ·
  3

• feat: add training throughput, FLOPs and MoE balance metrics

  safe-to-test
  #1693·yulangz opened Sep 10, 2026Collaborator·

  ·
  19

• feat(openai): support zero-argument tools and XML-escaped parameter parsing in ToolCallParser

  #1691·hsusul opened Sep 9, 2026Contributor·

  ·
  2

• feat(trainer): support score reward clipping and teacher weight normalization in MOPD loss

  stale
  #1690·hsusul opened Sep 9, 2026Contributor·

  ·
  2

• feat(archon): add NPU support for varlen attention

  stale
  #1686·262913 opened Sep 8, 2026·

  ·
  1

• refactor: unify AWEX converters and adapters across v1 and v2

  safe-to-test
  #1684·sitabulaixizawaluduo opened Sep 7, 2026Collaborator·

  ·
  4

• feat(workflow): support turn-level rewards and raw reward preservation in MultiTurnWorkflow

  #1681·hsusul opened Sep 5, 2026Contributor·

  ·
  2

• feat(trainer): support configurable entropy regularization in PPO and GRPO actor

  #1680·hsusul opened Sep 5, 2026Contributor·

  ·
  2

• feat: support single-GPU GSM8K smoke runs

  #1678·CharlesXu-HQ opened Sep 5, 2026Contributor·

  ·
  1

  1

• perf(engine): encode each distinct image once per vision-tower forward

  #1676·MrlixiangWE opened Sep 4, 2026·

  ·
  3

• fix(openai): preserve multi-turn step rewards and raw baselines in group normalization

  #1674·hsusul opened Sep 3, 2026Contributor·

  ·
  4

• fix(trainer): prevent reward and advantage explosion on remainder group normalization

  #1673·hsusul opened Sep 3, 2026Contributor·

  ·
  1