All pull requests

Pull requests

• fix(reward): saturate length rewards beyond configured maxima

  #10500·YueshenZ opened Nov 3, 2026Contributor·6 of 8 tasks·

  ·
  1

• docs: unify Ascend NPU naming as Huawei Ascend NPU

  #10496·Devin-DC-Huang opened Nov 2, 2026·5 of 8 tasks·

  ·

• [v5] 修复自适应初始化把规则奖励当成奖励模型

  #10495·qychen2001 opened Nov 2, 2026Collaborator·1 of 4 tasks·

  ·

• [utils] handle numpy values in check_json_format

  #10493·anh-nguyenn opened Nov 2, 2026Contributor·5 of 8 tasks·

  ·
  1

• fix(dataset): keep message/tool columns intact when map writes run in spawned workers

  #10492·simpleqt opened Nov 2, 2026Contributor·5 of 8 tasks·

  ·
  1

• Add GPTQModel-backed AWQ export

  #10491·sfeng1996 opened Nov 2, 2026·2 of 8 tasks·

  ·

• fix(agent): honour the tool schema in the remaining unquoted-wire templates

  #10490·simpleqt opened Nov 1, 2026Contributor·2 of 8 tasks·

  ·
  1

• [megatron] Wrap DDP/Megatron-FSDP on the current stream in eager mode

  #10489·arun-elr opened Nov 1, 2026·5 of 8 tasks·

  ·
  1

• fix(grpo): check active teacher batches before rejecting loss options

  #10488·YueshenZ opened Nov 1, 2026Contributor·2 of 8 tasks·

  ·

• fix(trainers): avoid gradient accumulation scaling during evaluation

  #10487·YueshenZ opened Oct 31, 2026Contributor·1 of 8 tasks·

  ·

• perf(dataset): cap torch threads in dataset_num_proc workers

  #10484·kjgggggg opened Oct 30, 2026Contributor·6 of 8 tasks·

  ·

• fix: mask response prefixes for every generated assistant turn

  #10478·Liuchenbing-2026 opened Oct 30, 2026Contributor·3 of 8 tasks·

  ·
  1

• Micro batch 切分下沉到twinke( dev swift v5)

  #10476·meichangsu1 opened Oct 30, 2026Contributor·2 tasks done·

  ·

• fix(metrics): keep numpy scalars out of the training logs

  #10474·RerankerGuo opened Oct 29, 2026Contributor·5 of 8 tasks·

  ·
  1

• perf(ray): reuse prompt formatting within completion-log batches

  #10470·gusario opened Oct 29, 2026Contributor·6 of 8 tasks·

  ·

• Fix completion-based microbatch validation for Ray GRPO

  #10469·gusario opened Oct 29, 2026Contributor·6 of 8 tasks·

  ·

• fix(dataset): preserve RNG when decision augmentation is disabled

  #10468·ryznren opened Oct 29, 2026·6 of 8 tasks·

  ·
  1

• fix(eval): support EvalScope OPENAI_API constant

  #10467·ryznren opened Oct 29, 2026·6 of 8 tasks·

  ·
  1

• fix(decision): evaluate calibration on held-out folds

  #10463·YueshenZ opened Oct 29, 2026Contributor·3 of 8 tasks·

  ·

• fix(agent): keep string tool arguments for hy_v3 and minimax

  #10462·mengchen-atx opened Oct 28, 2026·5 of 8 tasks·

  ·
  1

  1

• Fix distinct-node placement for Ray Megatron resource pools

  pending
  #10460·gusario opened Oct 28, 2026Contributor·6 of 8 tasks·

  ·

• fix(ray): allocate exportable CUDA IPC weight-transfer buffers

  #10459·gusario opened Oct 28, 2026Contributor·6 of 8 tasks·

  ·
  1

• Fix naive colocated weight export and avoid redundant offload/reload

  #10457·gusario opened Oct 28, 2026Contributor·6 of 8 tasks·

  ·

• perf(ray): run rollout sleep/wake concurrently across replicas

  #10454·gusario opened Oct 28, 2026Contributor·6 of 8 tasks·

  ·

• fix(rlhf): avoid full logits copies for strided log-softmax inputs

  #10450·YueshenZ opened Oct 28, 2026Contributor·2 of 8 tasks·

  ·