All pull requests

Pull requests

• fix(wan): drop hardcoded device='cuda' in WanToDance music encoder construction

  #1708·li-lizhe opened Sep 23, 2026·

  ·
  1

• fix(loader): raise when no local file matches origin_file_pattern

  #1705·Linxiushen opened Sep 21, 2026·

  ·

• fix(qwen-image-2.1): avoid OOM when building block-causal mask at hig…

  #1703·mi804 opened Sep 21, 2026Collaborator·

  ·
  1

• feat(loader): optionally stage ZeRO-3 checkpoints on CPU

  #1700·iMountTai opened Sep 21, 2026·

  ·

• fix(checkpoint): preserve keyword arguments with DeepSpeed

  #1699·iMountTai opened Sep 21, 2026·

  ·

• perf(minimax-h3): unpack QKV with unbind

  #1698·iMountTai opened Sep 21, 2026·

  ·

• [perf] MiniMax-H3 LoRA training on a B200 node: 3.74 s to 2.47 s per step at 8 GPUs, mostly from compiling the DiT block body

  #1680·TarzanZhao opened Sep 9, 2026·

  ·
  4

• Lora

  #1647·mi804 opened Aug 27, 2026Collaborator·

  ·

• wan: add dual-GPU model-parallel path for Wan 2.x LoRA training (depends on #1435)

  #1436·genno-whittlery opened May 11, 2026·

  ·
  8

• flux2: dual-GPU model-parallel + transformers 5.8 compat + Mistral-on-CPU data_process

  #1434·genno-whittlery opened May 11, 2026·

  ·
  4

• Fix/wan animate frame alignment

  #1335·zwplus opened Mar 7, 2026·

  ·
  10

• feat(wan): Add prior-based diffusion step skip for ~70% fewer inference steps

  #1325·Efrat-Taig opened Mar 4, 2026·

  ·
  6

• [New Feasture]: Add a FLOPs collection interface

  #1302·mahaocong90 opened Feb 19, 2026·

  ·
  12

• [New Feasture]: Support SP + DP parallal on Wan training

  #1223·mahaocong90 opened Jan 21, 2026·

  ·
  19

• Batch size

  #1205·ahmeddawy opened Jan 15, 2026·

  ·
  6

• feat: Add multi-GPU support with data/model/tensor parallelism

  #1153·longkeyy opened Dec 21, 2025·

  ·
  9

• fix: Add MPS and CPU device support

  #1152·longkeyy opened Dec 21, 2025·

  ·
  4

• [New Feasture]: Support for Flow-GRPO/Dance-GRPO

  #1141·Jayce-Ping opened Dec 17, 2025·

  ·
  10

• 支持int8-quanto量化微调Qwen-Image-Edit-2509,48G显存即可微调🎉,训练脚本添加result_image_field_name参数.

  #1101·Deng-Xian-Sheng opened Dec 7, 2025·

  ·
  8

• HunyuanImage-2.1

  #1044·yjy415 opened Nov 7, 2025Collaborator·

  ·
  8

• Adding tea cache wan2.2s2v

  #1017·aviveise opened Oct 29, 2025·

  ·
  7

• support Wan2.2-Fun-5B-Control[-Camera] inference

  #930·lifeiteng opened Sep 19, 2025·

  ·
  9

• feat: Add LoRA-based knowledge distillation for Wan-Video

  #922·MaulanaYusufIkhsanRobbani opened Sep 16, 2025·

  ·
  10

• I have created the demonstration script `examples/wanvideo/model_infe…

  #867·MaulanaYusufIkhsanRobbani opened Aug 28, 2025·

  ·
  10