All pull requests

Pull requests

• fix(compat): only stub flash_attn/bitsandbytes if actually installed

  #641·SputnikKaputtnik opened Sep 18, 2026·

  ·
  1

  1

• fix(cli): normalize image loading by actual bit depth, not hardcoded /255 (#513)

  #637·Delcado19 opened Sep 4, 2026·6 tasks done·

  ·
  1

• fix(color): avoid CUDA int32 sort overflow in LAB color correction (#522)

  #636·Delcado19 opened Sep 4, 2026·7 tasks done·

  ·
  1

• fix(seed): support full 64-bit seed range (#385)

  #635·Delcado19 opened Sep 4, 2026·6 tasks done·

  ·
  1

• ROCm VAE decode optimization: causal Conv3D → 3× Conv2D decomposition (MIOpen gfx1100 im2col workaround)

  #634·deadivan opened Sep 4, 2026·

  ·

• Add comfy/comfy_kitchen attention modes and int8_convrot checkpoint support

  #632·Deaquay opened Sep 2, 2026·

  ·

• fix(vae): guard MPS causal large-tensor ops

  #629·KepanWang opened Aug 27, 2026·

  ·

• feat(model-path)!: adapt to Comfy Desktop shared model registry; enable local_models_only to use shared models

  #625·Yewisp opened Aug 22, 2026·

  ·

• Add native INT8 ConvRot and NVFP4 SeedVR2 support

  #624·entropicnoise opened Aug 20, 2026·

  ·
  2

• Fix Windows crash in BF16 capability probe

  #620·the0cp opened Aug 16, 2026·

  ·
  1

• Fix: prepended warm-up frames are never stripped on the single-GPU path

  #616·geoffhtaylor3d opened Aug 5, 2026·

  ·

• Support for INT8, MXFP8 and NVFP4 models

  #613·Joly0 opened Jul 25, 2026·

  ·
  1

  18

• Prevent bfloat16 startup probe from crashing ComfyUI

  #611·lokitsar opened Jul 18, 2026·

  ·
  1

• Fix flash_attn_2_cuda import compatibility

  #608·ivellioscolin opened Jul 14, 2026·

  ·

• sageattn_1: SageAttention 1 (Turing/SM75+, recommended for RTX 20xx / GTX 16xx)

  #606·naxci1 opened Jul 5, 2026Contributor·

  ·

• Add experimental Intel XPU backend support

  #590·a86582751 opened Jun 7, 2026·

  ·
  5

• Fix CompatibleDiT len() crash with cache_model + torch.compile (#502)

  #588·Deaquay opened Jun 1, 2026·

  ·
  2

• Small fix in inference cli: Do final multiplication in Pytorch instead

  #578·sognetic opened May 3, 2026·

  ·

• ComfyUI Dynamic VRAM compatibility

  #572·Pizzawookiee opened Apr 13, 2026·

  ·
  15

• Allow loading ComfyUI-SeedVR2_VideoUpscaler nodes when running ComfyUI with --cpu

  #565·yungYEEZY opened Apr 5, 2026·

  ·
  3

• Harden shared cache lifecycle and add optional DiT tiling

  #551·xmarre opened Mar 15, 2026·

  ·
  6

• fixed bf16 support probe function

  #495·hum-ma opened Jan 17, 2026·

  ·
  1

• Copilot/integrate spargeattn optimization again

  #491·naxci1 opened Jan 14, 2026Contributor·

  ·
  1

• Add NVFP4 (4-bit floating point) quantization support for Blackwell GPUs

  #486·naxci1 opened Jan 12, 2026Contributor·11 tasks done·

  ·
  27

• ROCm repeat interleave HIP error

  #480·logicalor opened Jan 6, 2026·

  ·