Quantize Stable Diffusion examples

[LukeLIN-web]

I am using A6000.
I tried python quantize_StableDiffusion.py --batch_size=1 --torch_dtype="fp16" , it can work well.
But python quantize_StableDiffusion.py --batch_size=1 --unet_qtype="fp8" , it is very slow. Why?

Because readme.md we have installed latest

git clone https://github.com/huggingface/quanto
cd quanto
pip install -e .

Problem

1. What is different between this repo Optimum Quanto and https://github.com/huggingface/quanto
2. what is different between torch_dtype and unet dtype

[dacorvo]

1. quanto has just been renamed to optimum-quanto
2. the dtype is the type used in non-quantized operations (basically everything except Linear layers), and the qtype is the weight quantization for Linear in unet