All pull requests

Pull requests

• fix(critic): report clipfrac 0 when value_clip is not set

  #513·flyingdoubleG opened Oct 20, 2026·

  ·

• fix(agentic): normalize GiGPO step rewards with the configured mean/std types

  #512·flyingdoubleG opened Oct 20, 2026·

  ·

• fix(reward): keep the grouped shape in reward_norm when only std is per group

  #511·flyingdoubleG opened Oct 20, 2026·

  ·

• fix(advantage): mask critic values before GAE

  #510·flyingdoubleG opened Oct 20, 2026·

  ·

• fix(loss): apply loss_mask in agg_loss token-mean

  #509·flyingdoubleG opened Oct 20, 2026·

  ·
  1

• fix(dpo): honor objective configuration in training and validation

  #508·JiahaoTanXX opened Oct 16, 2026·

  ·
  1

• fix(rewards): keep ifeval reward loop control flow well-defined

  #503·AmirF194 opened Sep 28, 2026Contributor·

  ·
  1

• feat: add experimental Qwen3.8-Flash-Next post-training support

  #502·ruijieguo opened Sep 27, 2026·

  ·

• fix(rlvr): raise on unsupported tag in GeneralValRuleRewardWorker

  #501·Linxiushen opened Sep 21, 2026·

  ·

• Apply the model's own hf_to_vllm_mapper during weight sync

  #500·TracebaK opened Sep 15, 2026·

  ·

• Fix Qwen3.5 hybrid attention and add torch GDN fallback

  #497·ruijieguo opened Sep 11, 2026·

  ·

• fix(dynamic batching): preserve static mini-batch updates

  #494·JimmyWang0417 opened Aug 30, 2026·

  ·
  1

  1

• [WIP]feat(npu): add Ascend FP8/MXFP8 RLVR support and Qwen3.5 MoE reload

  #492·UsernameFull opened Aug 27, 2026Contributor·5 of 7 tasks·

  ·

• Add opt-in CPU-staging transport for colocated weight sync (ROLL_WEIGHT_SYNC_USE_CPU)

  #486·Donec-x opened Aug 21, 2026·

  ·
  1

  1

• fix(mcore_adapter): support fused MTP expert weights in qwen3_5 template

  #480·returnL opened Aug 12, 2026·

  ·
  1

• fix: is_offload_optimizer_states_in_train_step not passed to batch.meta_info in RLVR and agentic pipelines

  #478·sanmuf opened Aug 4, 2026Contributor·

  ·
  1

• fix: pin click<8.2 to unbreak ray CLIPatch 1

  #477·XiangyiWang2 opened Aug 1, 2026·

  ·
  3

• Add Mooncake DataProto transfer backend

  #469·zxpdemonio opened Jun 30, 2026·

  ·
  1

• [feat] Support ascend mindspeed&megatron and add CI tests

  #456·shun001 opened Jun 4, 2026Contributor·

  ·
  2

• Fix VLM position_ids packing in Megatron strategy with sequence packing enabled

  #452·sanmuf opened May 26, 2026Contributor·

  ·

• Add Trackio rollout trace logging

  #450·abidlabs opened May 22, 2026·

  ·
  1

• Fix GAE masked critic values bootstrapping valid advantages

  #447·haoyang9804 opened May 20, 2026·

  ·

• fix: mask token-mean loss aggregation

  #444·haoyang9804 opened May 18, 2026·

  ·
  1

• fix: account for multi-turn chunks in agentic LR scheduler budget (#407)

  #439·dashitongzhi opened May 8, 2026·

  ·
  1

  5

• fix: resolve distributed stability issues, LoRA weight sync, and recovery crash

  #438·Wangxiaoxiaoa opened May 8, 2026·2 tasks done·

  ·
  1

  1