Anyone else have a sudden bump up in VRAM usage? Mine happened over the weekend...

[Tom2112]

So this is strange and I'm trying to wrap my head around what might have caused it.

TL;DNR version:
Over the weekend, my automatic1111 using the same settings I had been using for months, now completely bogs down, making a single image generation take 10-20 minutes (occasionally more!), where the same settings prior to the problem would take 1-1.5 min and not bog down the computer.

The long version:
I run automatic1111 on a Win10 box with 128gig or ram, a Core i9-9900K, and a RTX 4090. That gives me a lot of room to generate images. I generate PonyXL images at 720x1080 and use hires fix to upscale them 2x to 1440x2160. I also run them through Adetailer, but I don't think that's part of the problem. I use the SerendipityPDXL v2 model, and I'm not using any loras or embeddings. I've been happily generating images using this same setup for months without any issues. I do use --medvram --xformers launch options, because xformers gives great speed and the medvram allows me to use the computer "normally" for web surfing and other general tasks while automatic1111 is generating images in the background. The VRAM usage that shows in the bottom right of the webUI after an image is generated usually showed in the mid 90's %, call it 95% on average. Each image took approx. 1 to 1.5 minutes to generate. So I was running close to the max, but still under it.
But this weekend, that changed. All of a sudden my generations were taking forever, the system would bog down heavily and it could take 10 or even 20 minutes to generate a single image. This is the same behavior I would get if I upped the resolution of my image past 1440x2160, say to a 2.5 upscale.
Generally, I automatic1111 stays running generating images 24/7 for me. So I go long periods of not restarting automatic1111 or restarting my computer. Usually I only break this cycle when it misbehaves or there's a critical windows update that needs installed and the system rebooted. I can't remember when that last happened, but I think it was right before this started. So I suspect maybe the Windows update is part of the problem. But it could also have been an update to automatic1111, as it would autoupdate when it was restarted.

What's really crazy is that I have a separate system running Forge. It's got 64gig of system ram, a Core i7-14700K, and a RTX 3080. The 3080 only has 10gig of VRAM, and it is generating images with the same exact settings 720x1080, upscaled to 1440x2160, without any issues. Sure, it's slower - it takes about 1.5 minutes to generate 1 image, vs. the 4090 system taking about 1 minute (when it was working properly). So what gives with my 4090 system?!? I'm very confused how a card with less than half the VRAM is not getting bogged down when a 4090 is.

EDIT:
I forgot to mention: my 4090 system bogs down at the hires fix stage. The initial image generation is fine and works the way it always did before. But as soon as that initial image gen hits 99% complete - where it switches over to the hires fix part - the whole system bogs down, all my VRAM is allocated (100%), and the system runs like crap - very sluggish, even mouse movement is slow/delayed.
I tried removing all of my added extensions except for dynamic prompts and saved state. But that didn't seem to make any difference. I do know that I'm using the exact same settings as before, because I used the saved state extension, so I've loaded states that worked fine before.

[jsxiong95]

Having the same issue did you ever figure it out?

[Tom2112]

I did not figure it out. I ended up switching to Forge. Forge not only runs quicker, it doesn't bog down like A1111 was doing. The down-side is that Forge doesn't stay running when I lock my session or if the browser window isn't visible - which is really annoying, but at least it runs well.