All pull requests

Pull requests

• fix(verl): match rollout alpha compositing for RGBA images

  #701·Chao Wang (chaoworks) opened Nov 6, 2026·

  ·
  1

• fix(server): preserve numeric zero data_id in terminal projection

  #700·KVR Vishnu (VISHNU0906) opened Nov 6, 2026·

  ·
  1

  1

• fix(server): accept case-insensitive Bearer auth scheme

  #699·KVR Vishnu (VISHNU0906) opened Nov 6, 2026·

  ·
  1

  1

• fix(swe-smith): stop command descendants on timeout

  #698·Zhengyang Liu (liuzhengyang699) opened Nov 6, 2026·

  ·
  1

  1

• fix(multimodal): preserve distinct image prompts

  #696·Zhengyang Liu (liuzhengyang699) opened Nov 6, 2026·

  ·
  1

  2

• fix(verl): stop admitting batches at the epoch limit

  #694·Zhengyang Liu (liuzhengyang699) opened Nov 6, 2026·

  ·
  1

  2

• fix(search-r1): score the selected answer action

  #692·Zhengyang Liu (liuzhengyang699) opened Nov 6, 2026·

  ·
  1

  1

• fix(swe-smith): preserve parametrized test IDs when grading

  #690·Zhengyang Liu (liuzhengyang699) opened Nov 6, 2026·

  ·
  1

  1

• fix(verl): preserve routed expert IDs above 255

  #688·Zhengyang Liu (liuzhengyang699) opened Nov 6, 2026·

  ·
  1

  1

• fix(science-world): preserve action delimiters in stopped replies

  #686·Zhengyang Liu (liuzhengyang699) opened Nov 6, 2026·

  ·
  1

  1

• fix(swe-smith): support native tool calls and thinking

  #684·Zhengyang Liu (liuzhengyang699) opened Nov 6, 2026·

  ·
  3

• feat(verl): add sealed process credit for on-policy training

  #683·Cookie0103 opened Nov 5, 2026·

  ·
  5

• fix(verl): retain outcome rewards on truncated trajectories

  #682·Chao Wang (chaoworks) opened Nov 4, 2026·

  ·
  2

• fix(verl): extract scalar rewards from tensor and array values in rollout manager

  #681·Hardik K (hardikkaurani) opened Nov 3, 2026·

  ·
  2

• fix(verl): preserve singleton group advantages and metrics in GRPO training

  #680·Hardik K (hardikkaurani) opened Nov 3, 2026·

  ·
  2

• fix(client): fail fast on non-retryable 4xx responses in post_with_retry

  #674·Chen Yufeiyang (feiiiiii5) opened Nov 1, 2026Contributor·

  ·
  1

  3

• fix(examples): keep completion tokenizer in sync with model overrides

  #670·Tianyuan Ding (likecake) opened Oct 31, 2026·

  ·
  1

  4

• fix(server): record invalid upstream responses as gateway errors

  #658·JianhaoYu (JianhaoYu-AI) opened Oct 26, 2026·

  ·
  1

  4

• fix(server): drop all superseded routed experts across multi-turn trajectories

  #653·Hardik K (hardikkaurani) opened Oct 24, 2026·

  ·
  8

• fix(verl): prevent stale/failed rollouts from breaking batch completion

  #649·Anvarbek Ziyodov (Mr-Perfectuz) opened Oct 17, 2026·1 of 2 tasks·

  ·
  5

• fix(server): preserve completion cursors when deleting rollouts

  #648·ningyixi opened Oct 17, 2026Contributor·

  ·
  3

• Fix dynamic rollout hook module registration

  #644·Newman Gao (bozarnr) opened Oct 15, 2026·

  ·
  1

• fix(verl): avoid replaying completion hooks across async steps

  #643·JiahaoTanXX (JiahaoTanXX) opened Oct 14, 2026Contributor·

  ·
  1

• feat(calc_x): add --val-only to score a checkpoint without training

  #642·Abdelhadi Salmaoui (abdelhadi703) opened Oct 13, 2026·

  ·
  1

  5

• test(verl): cover CISPO clipping and gradient semantics

  #614·Zhenhuan Wang (NaCr05) opened Oct 2, 2026Contributor·

  ·
  2