After upgrading the webui to the latest version with pytorch v2, the performance is great for the first image, then drops to really slow values

[honzajavorek]

I'm running the webui on a M1 MacBook with the following commands:

export COMMANDLINE_ARGS="--opt-split-attention-v1 --medvram --skip-torch-cuda-test --upcast-sampling --no-half-vae --use-cpu interrogate"

These ensured the best performance I was able to achieve. It was slow, but consistently slow, generating an image in a minute or a few. I was okay with it, because I just don't have better device 😄

Anyway. With the release of the v1.1.0 of the webui, I went to the project directory, git-pulled the latest commits, updated the dependencies. Now when I run the ./webui.sh, it:

• Generates the first txt2img image much, much faster! Wow! I get around: 3.19s/it
• But then for the second and following images, the performance drops to really slow values: 31.05s/it, but even hundreds.

With the same prompt, same model, same all settings, if I restart the ./webui.sh, it gives me the fast performance again. With the same prompt, same model, same all settings, the second image is slow. I don't know what to do now.

Obvious workaround is that I could restart the webui process after each image, but that's a really shitty workaround. Also I'm not entirely sure how to downgrade back in case I'd want my old slowish-but-not-so-bad-and-consistent performance back. Any ideas? Anyone else experience this? I searched through issues, but with no luck.

[honzajavorek]

Okay so right after writing all this, I went to experiment further and just commented out my changes to webui-user.sh effectively discarding custom COMMANDLINE_ARGS. And you know what? That really helps. With more and more images it looks like the speed still slows down, but it's still much better than with the custom settings.

[fanshu7]

me too........m2 MacBook Air

[honzajavorek]

See my own reply: #10096 (comment)