All pull requests

Pull requests

• Fix control images not bucketed for raw-control models with batch size > 1

  #1101·lucbrd8 opened Nov 3, 2026·

  ·

• Fix gradient_accumulation_steps discarding all but the last micro-batch

  #1100·joeuser12 opened Nov 3, 2026·

  ·

• Fix Qwen-Image 2.1 prompt-cache VRAM spikes by bypassing lm_head

  #1099·Epervios opened Nov 1, 2026·

  ·

• Automagic4: fix the runaway learning rate (#1084, #1091)

  #1098·speedyrulz opened Nov 1, 2026·

  ·
  1

• MiniMax-H3: honor quantize: false and a .safetensors name_or_path for the DiT

  #1094·jamin84 opened Oct 29, 2026·

  ·

• Use the sd1 arch name for SD 1.5 in the UI (#1092)

  #1093·ngo-binh0108 opened Oct 27, 2026·

  ·
  1

  1

• Diffusion separate sampling model.

  #1090·Lu-Gru opened Oct 26, 2026·

  ·
  1

• Feat/moss signed assistant lora

  #1089·a-l-e-x-d-s-9 opened Oct 26, 2026·

  ·

• 基于 FSDP2，支持 OPD

  #1083·ultism opened Oct 20, 2026·

  ·

• 支持多卡 FSDP2 训练，已验证通过

  #1082·ultism opened Oct 20, 2026·

  ·

• Add AMD ROCm support on Windows and Linux

  #1081·0xDELUXA opened Oct 20, 2026·

  ·
  4

• Don't let a missing fla break every extension job

  #1080·CobaltWoof opened Oct 19, 2026·

  ·

• Fix uv constraint paths in checkouts containing spaces

  #1073·wernjien opened Oct 5, 2026·

  ·

• Fix sample prompt caching size, Windows UTF-8 jobs, and two MiniMax-H3 bugs

  #1072·SibbiSibbi opened Oct 4, 2026·

  ·
  1

• YuE2: flush the CUDA cache every train step on Windows (fixes step-time collapse, #1052)

  #1069·bczreasons opened Sep 29, 2026Contributor·

  ·

• YuE2: put @torch.no_grad() back on merge_nar_lora (regression from #1055)

  #1068·bczreasons opened Sep 29, 2026Contributor·

  ·

• fix(v2): build real weights on CPU in aitk_from_pretrained to avoid meta tensor crash (Qwen-Image-2.1)

  #1067·limitcool opened Sep 26, 2026·

  ·

• Skip timers with no completed runs in Timer.print

  #1065·MohammadHijjawi97 opened Sep 25, 2026·

  ·
  1

• [bug fix] don't emit an empty widget key when push to hub

  #1062·linoytsaban opened Sep 24, 2026·

  ·

• Fix crash when lr_scheduler is set to "step"

  #1061·wcmille opened Sep 23, 2026·

  ·
  1

• Pass sample size when pre-encoding sample prompts (fixes Qwen Image 2.1 sampling with cached text embeddings)

  #1060·KJeon10 opened Sep 23, 2026·

  ·
  1

• Fix MiniMax H3 LoRA training producing zero gradients (fp16 adaLN overflow on text rows)

  #1046·HE1CO opened Sep 16, 2026·

  ·

• Added the use of tiled VAE in the case where the low VRAM mode is use…

  #1024·wingbrother opened Aug 29, 2026·

  ·

• Fix sample image grid ordering for 10+ prompts

  #1018·rossnot opened Aug 25, 2026·

  ·
  1

  2

• Add torchaudio to requirements

  #1000·matlowai opened Aug 7, 2026·

  ·
  2