All pull requests

Pull requests

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

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

  ·

• feat(inf-train parity): add Qwen3-MoE train-inference parity plugin, part 1/2 — vLLM integration

  need review
  #504·cboss6 opened Sep 23, 2026Contributor·

  ·

• perf(minimax-h3): batch replay timesteps

  wip
  #500·leviking98z-rgb opened Sep 22, 2026Collaborator·

  ·

• refactor: rename residual DiffRL identifiers to UniRL

  need review
  #498·CjhHa1 opened Sep 21, 2026Collaborator·

  ·

• fix(reward): fail fast when selected prompt is absent

  need review
  #497·seshakolli opened Sep 21, 2026Contributor·

  ·

• refactor(train): share per-micro anchor preparation between train stacks

  need review
  #495·chi030303 opened Sep 20, 2026Contributor·

  ·
  1

• fix(vLLM): harden native IPC synchronization

  need review
  #494·CjhHa1 opened Sep 20, 2026Collaborator·

  ·

• feat(reward): add guarded CUDA MPS runtime

  wip
  #492·Zcchill opened Sep 19, 2026Collaborator·

  ·

• feat(experimental): add the offline replay trace contract and converter

  need review
  #487·0z5a opened Sep 18, 2026·

  ·
  1

• feat(fsdp): add a measured resident/streamed block placement planner

  need review
  #486·0z5a opened Sep 18, 2026·

  ·

• fix(rollout): validate aligned prompt-group chunk boundaries

  need review
  #485·0z5a opened Sep 18, 2026·

  ·
  2

• fix(hi3): support batched trainside AR with dynamic KV cache

  need review
  #479·PushpakAg opened Sep 17, 2026·

  ·
  1

  1

• feat(dyref): add multi-reference SFT and RL training

  wip
  #478·Ideny42 opened Sep 17, 2026Collaborator·

  ·

• feat(qwen3-omni): add Thinker offline DPO on the supervised trainer

  need review
  #476·leviking98z-rgb opened Sep 16, 2026Collaborator·

  ·

• feat(ltx2): add CLAP-only T2AV reward recipe

  need review
  #475·leviking98z-rgb opened Sep 16, 2026Collaborator·

  ·
  1

• perf(minimax-h3): shard the video VAE decode across the SP group

  wip
  #469·leviking98z-rgb opened Sep 15, 2026Collaborator·

  ·
  1

• fix(sglang): preserve CUDA compatibility across scheduler spawn

  need review
  #468·CjhHa1 opened Sep 15, 2026Collaborator·

  ·

• feat(flowgrpo): add bounded-memory chunked replay backward

  need review
  #465·Yuanjia-Lee opened Sep 14, 2026Contributor·

  ·

• fix(eval): include ragged prompt tails in PE and unified evaluation

  need review
  #464·Yuanjia-Lee opened Sep 14, 2026Contributor·

  ·

• fix(reward): reject ImageBind audio_video as a silent training reward

  need review
  #461·chi030303 opened Sep 14, 2026Contributor·

  ·
  1

• feat(opd): introduce diffusion teacher-provider contract and full-model teacher

  need review
  #451·BeatSeat opened Sep 14, 2026·

  ·

• feat(rollout): integrate vllm-omni prompt-embed cache and sleep lifecycle (Slice 1, #443)

  need review
  #449·BeatSeat opened Sep 13, 2026·

  ·

• fix(algorithms): apply GRPO's masked token-loss reduction to DRPO/CPPO/DPPO

  need review
  #448·GoldWaterFall opened Sep 13, 2026·

  ·

• feat(reward): micro-batch the diffusion rollout so a remote reward overlaps generation

  need review
  #445·celve opened Sep 12, 2026Collaborator·

  ·
  1

• perf(train): keep eager AdamW step on CPU across offload/onload

  need review
  #441·Yuanjia-Lee opened Sep 12, 2026Contributor·

  ·