ComfyUI Native INT8 ConvRot Fix (No more KeyError: 'int8_rowwise')
If you tried using several INT8 ConvRot versions of this model available on others platforms in ComfyUI (v0.27.0+), you probably encountered a nasty crash: KeyError: 'int8_rowwise'.
Why did that happen? Early experimental quantization scripts generated files using row-wise scaling. However, ComfyUI's native INT8 implementation only supports tensor-wise scaling out of the box (int8_tensorwise). Even custom loaders like KJNodes bridge back to Comfy's native core, causing an instant pipeline crash.
What this version fixes: I re-quantized the official BF16 model from scratch using the official parameters, strictly forcing --scaling_mode tensor.
How to use it:
Drop this
.safetensorsfile into yourComfyUI/models/diffusion_models/folder.Load it natively using the standard Load Diffusion Model.
No custom nodes required! Enjoy massive VRAM savings and FP16-like image quality natively on your GPU.
Works with Sage Attention too for more speed!
Description
FAQ
Comments (15)
Using the Load Diffusion Model directly causes errors
Hello, try with the Load Diffusion Model from KJ.
@KinkReaper It didn't work.
@woaiyitiaofeng What is the errors in the terminal exactly?
According to the descriptions, this version should run faster than the regular FP versions. However, for me it runs twice as slow as the regular ones. With standard nodes.
What is your configuration?
@KinkReaper What do you mean? Comfy, standart nodes for Krea2. Standart wf.
@herkus_baronas631 I mean, what are the specs of your PC? What is your graphic card model? Are you on AMD?
@KinkReaper Nvidia.
Maybe for big ones only faster? For 5090 24-32 Gb?
My card is Nvidia 5070 Ti 16 Gb, but int8 works practically twice as slow as with fp8.
qwen 2512 fp8 8 steps 1600 x 896 took 16 sec,
krea 2 int8 8 steps 1600x896 took 28 sec.
krea 2 fp8 8 steps 1600x996 took 15 sec.
@herkus_baronas631 I have a 5080 with 16GB of VRAM, and INT8 cuts my generation time in half. I tested my checkpoint on another, less powerful computer and got the exact same result. There's something in your configuration that's blocking it.
@KinkReaper I used regular wf, used from Pixorama, another one... same results. Qwen work faster... where problems may to be? I dont know python or how to programing something. I'm simply user.
4070 TI SUPER 16GB. 32GB RAM. with default loader - Out of resources. Loading take endless.
100%|██████████| 8/8 [09:13<00:00, 69.21s/it]
Tried couple of common fp8, works fine. (13-15 sec. 1280x960)
With node "Load Diffusion Model INT8 (W8A8)" it works, but i can't normaly use loras. Or i do something wrong. But clean model spend 7-8 sec to same resolution, what's tempting
After updating comfy it works 2x faster on 3090. Thank you!