2080ti22g，It takes more than ten minutes to generate an image.
#57
by
dwedwqe21w
- opened
After carefully reading the instructions, I understand that if the framework uses BF16, otherwise it will default to converting to FP32 precision for image generation, which results in it taking more than ten minutes to generate an image. I would like to ask when support for RTX 20-series graphics cards will be available, as these cards only support FP16.
This is not possible because the model will generate a lot of NAN when using fp16 datatype inference and will not get any valid output. fp16 is too small a dynamic range for it.