All pull requests

Pull requests

• feat(flux3-image): add FLUX.3 Image SFT through an encode-only server

  #345·zhihengy opened Oct 28, 2026Collaborator·3/3·

  ·

• feat(sft): encode SFT cache misses through an sglang-diffusion encode-only server

  #344·zhihengy opened Oct 28, 2026Collaborator·2/3·

  ·

• [DMD2 9/9] feat(dmd2): add the MiniMax H3 4-step DMD2 recipe with eval media and DMD2 metrics

  #343·zhihengy opened Oct 26, 2026Collaborator·5 of 6 tasks·

  ·

• [DMD2 8/9] perf(ray): sleep each trainer group right after its train

  #342·zhihengy opened Oct 26, 2026Collaborator·5 of 6 tasks·

  ·

• [DMD2 4/9] feat(fsdp): add a train-input hook that preprocesses train pairs before micro-batching

  #338·zhihengy opened Oct 24, 2026Collaborator·4/7·6 tasks done·

  ·

• [DMD2 1/9] refactor(ray): build each trainer's args from a trainer config list

  run-ci-basic
  #335·zhihengy opened Oct 21, 2026Collaborator·1/7·6 tasks done·

  ·

• [DMD2 7/9] feat(dmd2): train the student with the distribution matching gradient from the critic group's scores

  #334·zhihengy opened Oct 19, 2026Collaborator·7/7·5 of 6 tasks·

  ·

• [DMD2 6/9] feat(dmd2): train a fake-score critic group from the teacher and score the student's samples

  #333·zhihengy opened Oct 19, 2026Collaborator·6/7·5 of 6 tasks·

  ·
  1

• [DMD2 5/9] feat(rollout): stop the rollout at a random step for DMD2 backward simulation

  run-ci-basic
  #331·zhihengy opened Oct 19, 2026Collaborator·5/7·6 tasks done·

  ·

• [DMD2 3/9] feat(ray): hand earlier trainer groups' outputs to later groups as external_data

  run-ci-basic
  #330·zhihengy opened Oct 19, 2026Collaborator·3/7·6 tasks done·

  ·

• [DMD2 2/9] feat(ray): train several colocated trainer groups in the sync driver

  run-ci-basic
  #329·zhihengy opened Oct 19, 2026Collaborator·2/7·6 tasks done·

  ·

• fix(rollout): stop dynamic sampling from leaking generation past the rollout

  #324·OnePunchMonk opened Oct 17, 2026Contributor·6 tasks done·

  ·
  3

• feat(flux3-image): add FLUX 3 Image Klein RL support

  #316·Rockdu opened Oct 16, 2026Collaborator·1/3·

  ·
  1

• perf: startup timing probes and env-gated startup optimizations

  #310·OnePunchMonk opened Oct 11, 2026Contributor·

  ·
  6

• fix(rollout): wrap data source across epochs and cancel leftover rollout tasks

  #301·OnePunchMonk opened Oct 7, 2026Contributor·

  ·
  1

  1

• feat(fsdp): score centering for Flow-GRPO under train/rollout mismatch

  #298·HJSang opened Oct 6, 2026·

  ·

• feat(reward): add DOVER video quality scoring

  #260·JingwenGu0829 opened Sep 24, 2026Contributor·

  ·

• fix(fsdp): preserve FA3 autograd and deterministic dispatch

  #259·Odysseusailoon opened Sep 24, 2026·

  ·

• fix(rollout): stop OcrScorer crashing the batch on malformed prompts

  #255·OnePunchMonk opened Sep 20, 2026Contributor·

  ·
  1

  4

• fix(rollout): stop dropping every single-sample group in check_reward_nonzero_std

  #253·OnePunchMonk opened Sep 20, 2026Contributor·

  ·
  1

• fix(http): stop leaking detached Ray actors for distributed POST

  #251·OnePunchMonk opened Sep 17, 2026Contributor·

  ·
  2

• feat(diffusion): support Wan2.2 TI2V Flow-GRPO training

  #228·wwwsctvcom opened Sep 3, 2026·

  ·
  1

• perf(startup): overlap engine boot with train init + mmap loading

  #215·Rockdu opened Aug 29, 2026Collaborator·5 of 6 tasks·

  ·

• feat(sft): single-card equivalent of the 8-GPU H3 SFT recipe

  #210·Rockdu opened Aug 26, 2026Collaborator·4 of 6 tasks·

  ·

• feat(fsdp): add --fsdp-reshard-after-forward for ZeRO-2 style training

  #196·Rockdu opened Aug 21, 2026Collaborator·3 of 6 tasks·

  ·