All pull requests

Pull requests

• feat(ar): add multi-teacher OPD task-cycle accumulation

  wip
  #624·ruiling-smartbear opened Oct 23, 2026Contributor·1 of 2 tasks·

  ·

• refactor(train): read loss-normalization units from StageAlgorithm.loss_units

  need review
  #622·MinhaoLi0318 opened Oct 21, 2026Contributor·2 tasks done·

  ·

• fix(reward): validate capabilities before model loading (#423)

  need review
  #621·GBX-Max1220 opened Oct 20, 2026·2 tasks done·

  ·
  1

• feat(async): expose rollout manager depth, lag, and activity

  need review
  #620·coding-famer opened Oct 19, 2026Contributor·2 tasks done·

  ·
  1

• feat(rollout): support Qwen-Image RL with vLLM-Omni native IPC

  need review
  #617·Acerak01-fy opened Oct 19, 2026·1 of 2 tasks·

  ·
  10

• fix(train): keep empty AR rows in FSDP replay and backward

  wip
  #598·ruiling-smartbear opened Oct 13, 2026Contributor·2 tasks done·

  ·
  1

• feat(trainer): share bounded rollout and eval sample observation

  wip
  #597·ruiling-smartbear opened Oct 13, 2026Contributor·1 of 2 tasks·

  ·

• feat(minimax-h3): train original block weights with frozen peripherals

  wip
  #593·ruiling-smartbear opened Oct 12, 2026Contributor·2 tasks done·

  ·

• feat(diffusion): support per-prompt SDE step rollout and Wan replay

  need review
  #589·ruiling-smartbear opened Oct 12, 2026Contributor·2 tasks done·

  ·

• feat(minimax-h3): carry FL2VA keyframe condition rows through rollout, replay and NFT

  need review
  #584·MinhaoLi0318 opened Oct 11, 2026Contributor·2 tasks done·

  ·
  2

• feat(benchmarks): add unified smoke result schema

  need review
  #576·MciG-ggg opened Oct 8, 2026·2 tasks done·

  ·
  1

  1

• feat(ar): monitor full-vocabulary policy entropy

  need review
  #563·ruiling-smartbear opened Oct 5, 2026Contributor·2 tasks done·

  ·
  1

• feat(ar): add bounded dynamic sampling for synchronous GRPO

  need review
  #562·ruiling-smartbear opened Oct 5, 2026Contributor·2 tasks done·

  ·
  2

• feat(agentic): add bounded async turn-boundary rollout

  need review
  #558·ruiling-smartbear opened Oct 5, 2026Contributor·1 of 2 tasks·

  ·

• feat(grpo): add opt-in token rollout correction for replay anchors

  need reviewwip
  #557·ruiling-smartbear opened Oct 5, 2026Contributor·2 tasks done·

  ·
  1

• feat(tempflowgrpo): add TempFlow-GRPO with per-step branching

  need review
  #552·ruiling-smartbear opened Oct 4, 2026Contributor·2 tasks done·

  ·
  1

• feat(flashgrpo): rework Flash-GRPO onto current main

  wip
  #546·ruiling-smartbear opened Oct 4, 2026Contributor·2 tasks done·

  ·

• feat(sde): MixGRPO-Flash post-window DPM-Solver++ tail for SD3 trainside

  need review
  #545·ruiling-smartbear opened Oct 4, 2026Contributor·2 tasks done·

  ·
  1

• fix(hi3): repair checkpoint tokenizer batchify at the root

  need review
  #544·ruiling-smartbear opened Oct 3, 2026Contributor·2 tasks done·

  ·
  1

• fix(async): record phase timings and full step duration

  need review
  #539·coding-famer opened Sep 30, 2026Contributor·2 tasks done·

  ·
  7

• feat(algorithms): AROPD — sampled-token on-policy distillation for AR

  need review
  #534·daii-0818 opened Sep 28, 2026Contributor·2 tasks done·

  ·
  12

• refactor(opd): harden frozen teacher adapter state restoration

  need review
  #519·WenzheWang opened Sep 27, 2026·2 tasks done·

  ·
  3

• feat(qwen_image21): add Qwen-Image-2.1 text-to-image trainside support

  need review
  #518·CjhHa1 opened Sep 26, 2026Collaborator·2 tasks done·

  ·

• feat(trainer): opt-in row-level reward dispatch for diffusion

  need review
  #517·CjhHa1 opened Sep 26, 2026Collaborator·2 tasks done·

  ·
  1

• feat(inf-train parity): add Qwen3-MoE train-inference parity plugin, part 2/2 — training integration

  need review
  #505·cboss6 opened Sep 23, 2026Contributor·4 tasks done·

  ·