text encoder of choice in prompt upsampler
#11
by squallelf - opened
I am not really familiar with LLM quantization. but for image generation, 4 bits model seems too bad to the quality.
But in the world of LLM, i seems commonly encounter Q4 model. Is using Q4 model in LLM still good to go?
Actually, what i would like to ask is, if i use the W4A8 prompt upsampler and text encoder, will it still a good choice without much quality loss?
And also i would like to know if the vision tower in W4A8 CLIP also 4 bits quantized? or fully preserved? since we almost always have image input, so the vision of CLIP is somehow quite important.