Gguf-comfy

update highlights:

- sd3.5-large/medium supported
- gguf node supported
- bug fixed + code optimized
- sample workflows included

update highlights:

- sd3.5-medium support (for both safetensors and gguf)
- remove the kornia library conflict warning
- resolve attention abstraction in mmdit and tensor error message(s)
- SimpleTuner support for sd3 etc., for those details should refer to [comfyui](https://github.com/comfyanonymous/ComfyUI/releases) side; thanks

major update highlights:
- this version is leveraging on gguf-connector's generic tensor writer and reader; advantages see below
- gguf module is no longer needed since it doesn't support numpy >= 2.0.0 recently
- developer or/and expert user can git clone this repo for operating the gguf node with the latest version of comfyui
- backward support the old version(s) comfyui and gguf node still
- better code control and/or management as you don't need to wait for the major library/dependency update
- credit should always give/share to all contributors, i.e., comfyanonymous (comfyui), city96 (comfyui-gguf), ggerganov (llama.cpp), etc.

upgrade highlights:
- update standalone package to: python 3.12, torch 2.5, numpy 2.1.2 etc.
- support fp8 model for best quality
- support hardware up to: Nvidia 40 series/ada/H100 and above
- bug fixed for some library conflicts
- GGUF node supported enhancement for: flux, sd3

Stable Diffusion 3.5 Large model supported upgrade

SETUP:

sd3.5_large-q4_0.gguf (4.77GB) > ComfyUI/models/unet
clip_g.safetensors (1.39GB) > ComfyUI/models/clip
clip_l.safetensors (234MB) > ComfyUI/models/clip
t5xxl_fp8_e4m3fn.safetensors (4.55GB) > ComfyUI/models/clip
diffusion_pytorch_model.safetensors (168MB) > ComfyUI/models/vae

drag the workflow file to your browser:
gguf-workflow-sd3.5.json

You are good to GO! (now you can run sd3 with the cheapest Nvidia card or merely CPU) ENJOY!

bug fixed; polished