All pull requests

Pull requests

• fix: drop unused "gear" dependency that resolves to an unrelated PyPI package

  #96·quinncheong opened Oct 6, 2026·

  ·
  1

• Fix uninitialized first-frame output in video-only local attention

  #94·Sung-Jae-Seong opened Sep 11, 2026·

  ·

• Mix rank and worker id into the sharded loader's within-shard shuffle seed

  #93·dineshmane opened Sep 10, 2026·

  ·
  1

• Exclude frozen parameters from DeepSpeed checkpoints when training LoRA (95 GB → 3.6 GB per save)

  #92·dineshmane opened Sep 10, 2026·

  ·
  1

• Skip short trailing windows so batch > 1 collate sees uniform shapes

  #87·jeqcho opened Aug 7, 2026·

  ·
  1

  1

• Unpack DiT block output tuple in the non-checkpointed training branch

  #86·jeqcho opened Aug 7, 2026·

  ·
  1

• Fix load_lora to rebuild post-trained LoRA checkpoints on their recorded pretrained base

  #83·dineshmane opened Jul 24, 2026·

  ·
  1

  1

• docs: complete data flow performance model for DreamZero training

  #81·zhangqin200182 opened Jul 21, 2026·

  ·

• Fix prompt padding mask per sample

  #71·seasideGargantua opened May 20, 2026·

  ·

• Fix action_loss_per_sample shape for has_real_action

  #45·Loki-Lu opened Mar 24, 2026·

  ·
  1

  2

• add annotation

  #40·flowerinheart opened Mar 23, 2026·

  ·

• Fix: inject LoRA adapters before loading LoRA weights in from_pretrained

  #28·haiyuanwan opened Mar 10, 2026·2 tasks·

  ·