CUDA errors or Crazy slow: pick your poison

[silverhammer751]

This is a local build problem, and I'm wondering if anyone has found any workarounds.

The issue is that I can either generate images really slowly (it took an hour to generate 4 images w/hires fix) by using the CUDA system fallback on Nvidia graphics cards, or I can disable CUDA system fallback and get a host of different CUDA errors (typically CUDA out of memory where it tries to do 50MiB beyond what my system has).

As I pointed out in the bugs, --lowvram is currently destroying image quality, so that's not an option.

I have an 8gb NVIDIA GeForce 2070 Super. It used to run SDXL totally fine, no CUDA errors. I've tried rolling back to SD v1.6 but that hasn't worked either.

Basically, the issue I'm running into is that my 8gb used to be enough to run SDXL. Now, I get yelled at any time I try to run it just on my video card. The only option where I don't get errors is if I use system ram in addition to VRAM, and that slows down the image generation by a ton.

Anyone else run into this issue? Found a way to fix it? An old hash? Old NVIDIA driver that doesn't screw up Stable Diffusion?

[Woisek]

I have the same graphics card, and normally I have no problems with it. I can generate 1024x720 images with SDXL in around a minute (without highres fix).

I can only add this:
Something I know that slows it down is currently using ReActor as it adds obviously heavy to the VRAM which I personally don't quite understand why this is, but it's investigated.
Then I experienced a slowdown lately, but I am not aware of what caused it, and it didn't happened again so far.
And I'm running a fresh install of A1111 v1.7.0
Maybe something else changed with your system.

[silverhammer751]

Yeah, that's what it used to be for me too. I left for the holidays, came back, and ran into tons of errors. I've updated and reinstalled v1.7.0 but I can't seem to be able to generate images when I limit python to only using VRAM (prefer no system fallback), and otherwise I'm getting 15.53s/it.

What NVIDIA driver are you using? What are your command line args?