Vram leak or bad managment on rx7900xtx

[Kademo15]

So i am new to this ai stuff and i tried nod-ai SHARK for stable diffusion first but then also tried automatic 1111 because of the many features it has.
When i start generating my vram goes up to 24/24 gb and stays there even when finished generating. Same on a 5700xt.
The problem is when i want to use hiresfx or i create an image bigger than like +-700*700 px I get an "not enough memory" crash because there is none because of the leak.
I heard that it is a directml-torch bug, is that true?
Is there anyway to fix this myself or a workaround.
Thanks for the help

[Sakura-Luna]

Why don't you look at other discussions? If your shark works fine, don't use auto1111.
#8344

[Kademo15]

ok well it finished and at the and gave me an RuntimeError: The parameter is incorrect. The image size was 2048x2048

[Kademo15]

With 1920x1080 it worked was just very slow 10s/it but that was to be expected but it gave me a weird image the prompt was1girl

[Kademo15]

Do you think this will help with the high usage and slow times here

[Sakura-Luna]

I don't answer usage related questions, try browsing related discussions.
--medvram trades speed for memory, whether to use it is up to you, the best choice is still to run WebUI on Linux. Don't worry too much about the size of the output, some models struggle to generate meaningful output at high resolutions. By the way, you should consider closing the issue submitted to direct-ml.

[Kademo15]

Ok thx