All pull requests

Pull requests

• Support ROCm on Windows with ADLX VRAM readings

  #1799·0xDELUXA opened Oct 30, 2026·

  ·

• feat: add experimental sequence-parallel training for MiniMax-H3

  #1798·yeruoxiangchen opened Oct 30, 2026·

  ·

• Gather ZeRO-3 parameters before fusing LoRA weights

  #1797·yeruoxiangchen opened Oct 30, 2026·

  ·
  1

• LTX-2.5 IC-LoRA training: reference_video path, first-frame loss mask, LR from step one, VRAM guard

  #1795·DomBeckett opened Oct 30, 2026·

  ·
  1

• krea2: route attention through the shared attention_forward (cuDNN SDPA pin has no ROCm kernel)

  #1793·DomBeckett opened Oct 30, 2026·

  ·
  1

• docs: add Traditional Chinese (Taiwan) README

  #1792·monarchsofchaos004-gif opened Oct 30, 2026·

  ·

• Fix OPD step selection in full and LoRA training examples

  #1791·yeruoxiangchen opened Oct 29, 2026·

  ·

• fix moss image sampling latent precision

  #1787·mi804 opened Oct 25, 2026Collaborator·

  ·

• fix(vram): keep the trainable parameters of the wrapped module

  #1784·li-lizhe opened Oct 22, 2026Contributor·

  ·
  1

  1

• multi-controlnet

  #1778·Yuze-e20 opened Oct 21, 2026Contributor·

  ·
  1

• Fix LTX video loss on fixed conditioning tokens

  #1755·taking-lying-flat opened Oct 14, 2026·

  ·

• Fix SDXL clip skip selection for both text encoders

  #1754·taking-lying-flat opened Oct 14, 2026·

  ·

• fix: fail MiniMax-H3 training when audio latents were never encoded

  #1748·mvanhorn opened Oct 13, 2026·

  ·
  1

• fix(training): preserve H3 conditioning gradients in split training

  #1747·Shenrui-Ma opened Oct 13, 2026·

  ·

• feat: Minimax-H3 surpport NPU

  #1745·ceshust opened Oct 13, 2026·

  ·

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

  #1436·genno-whittlery opened May 11, 2026·4 of 5 tasks·

  ·
  8

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

  #1434·genno-whittlery opened May 11, 2026·4 of 6 tasks·

  ·
  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, 2025·3 of 5 tasks·

  ·
  9

• fix: Add MPS and CPU device support

  #1152·longkeyy opened Dec 21, 2025·3 tasks done·

  ·
  4

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

  #1141·Jayce-Ping opened Dec 17, 2025·2 of 10 tasks·

  ·
  10