sd-webui-controlnet icon indicating copy to clipboard operation
sd-webui-controlnet copied to clipboard

Published 20 hours ago • verified publisher icon Mikubill

OOM error does not release GPU VRAM

Open entmike opened this issue 1 year ago • 5 comments

If an API call is done and goes CUDA OOM, a lot of memory is left allocated when looking innvidia-smi and does not seem to release it without restarting A1111.

entmike avatar Feb 25 '23 05:02 entmike

Try using a lower model cache size may help

Mikubill avatar Feb 25 '23 05:02 Mikubill

Try using a lower model cache size may help

Sorry I should have given more details. I am only using 1 model loaded ever. RTX 3090 in Linux.

entmike avatar Feb 25 '23 05:02 entmike

I think this fixed this: https://github.com/Mikubill/sd-webui-controlnet/issues/347#issuecomment-1444647723 - Closing unless I can cause it to happen again.

entmike avatar Feb 26 '23 22:02 entmike

Reopening actually. It still will not release any memory after an OOM error, after all. Looking for a hand here, please let me know if I can provide additional details!

entmike avatar Feb 27 '23 05:02 entmike

Yes, the video memory will not be released at all. Even if I choose 0 cache models, it cannot be released.

kkiqq22 avatar Mar 04 '23 12:03 kkiqq22