Updated Optimum SDXL Usage (markdown)

missionfloyd 2023-09-08 21:27:06 -06:00
parent 09b40bc3ea
commit 5bb3238a12

@ -2,20 +2,19 @@ Here's a quick listing of things to tune for your setup:
## Commandline arguments:
- (nvidia) 8gb `--medvram-sdxl --xformers`
- (nvidia) 12gb+ `--xformers`
- (nvidia) 4gb `--lowvram --xformers`
- Nvidia (12gb+) `--xformers`
- Nvidia (8gb) `--medvram-sdxl --xformers`
- Nvidia (4gb) `--lowvram --xformers`
## System:
- downgrade nvidia drivers to 531 or lower prevent extreme slowdowns for largest pictures
- add a pagefile to prevent failure to load weights due to low cpu ram
- (Linux) install tcmalloc - greatly reducing ram usage: `sudo apt install --no-install-recommends google-perftools` [#10117](https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/10117)
- use an SSD, for faster loadtime, especially if a pagefile is required
- converting `.safetensors` to `.ckpt` for reduced ram usage [#12086](https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/12086#issuecomment-1691154698)
- Downgrade [Nvidia drivers](https://www.nvidia.com/en-us/geforce/drivers/) to 531 or lower prevent extreme slowdowns for largest pictures.
- Add a pagefile to prevent failure loading weights due to low RAM.
- (Linux) install `tcmalloc`, greatly reducing RAM usage: `sudo apt install --no-install-recommends google-perftools` ([#10117](https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/10117)).
- Use an SSD for faster load time, especially if a pagefile is required.
- Convert `.safetensors` to `.ckpt` for reduced RAM usage ([#12086](https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/12086#issuecomment-1691154698)).
## Model weights:
- use vae that will not need to run in fp32 for increased speed and less vram usage: [sdxl_vae.safetensors](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors)
- use [TAESD](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#taesd)
- use fp16 (~7gb) weights for less cpu ram usage
- Use a VAE that will not need to run in fp32 for increased speed and less VRAM usage: [sdxl_vae.safetensors](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors).
- Use [TAESD](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#taesd).
- Use fp16 weights (~7gb) for less RAM usage.