Hi everybody :-)

[OleJ1964]

I’m trying to run SDAXL on AUTOMATIC 111 webui - and i bought a 4070 with 12 (Can't afford the ridiculous prices for the high end 4080/4090)

Now it takes 26 sec avg for one image 1024x1024 with 20 samples - BUT it's clear that it gets 80% of the way in 5 sec. and then it halts for a LONG time to purge/"garbage collect" for more memory. So 16 GB VRam would be lovely.

Perhaps someone have some tips, --MedRam --lowram does not make it better.

Anyone that have some good advice?- always willing to learn . Cheers

[ClashSAN]

Issue is caused wide variety of unknown combinations and system setup.

If you need to try the model immediately, use
https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_base_1.0_0.9vae.safetensors, downgrade and run with an older driver that won't fight with the other constructed memory management optimizations, setup a pagefile for model loading/swapping.

Check:

• if the model is properly adapted to fp16
• your newest nvidia drivers are using shared memory (upstream, no toggle settings yet)
• you have very little CPU ram, a large minimum 32gb is needed testing SDXL (further efforts were made to allow much lower ~24gb or less needed on dev, which will be the future release version)

[OleJ1964]

Thanks for the reply :-).
I'm running on 32GB Ram with a RTX 4070 12 GB VRAM
Using the XL Base Model on Automatic 1111 WEBUI
version: [v1.5.1]  •  python: 3.10.6  •  torch: 2.0.1+cu118  •  xformers: 0.0.20  •  gradio: 3.32.0
newest NVIDIA Drivers

Not really sure what I can do (if anything) about it. -