[bug]: SDXL generation extremely slow since macOS Sonoma update

[LuLeRoemer]

Is there an existing issue for this?

• I have searched the existing issues

OS

macOS

GPU

mps

VRAM

No response

What version did you experience this issue on?

3.4.0post2

What happened?

Since updating to macOS Sonoma on my M1 Pro (16GB), image generation is unusably slow (> 10 minutes for 512x512). I tried to fix it with a complete reinstall but was not successful. Even using image sizes around 512x512 takes forever. Before running the update, I could generate an image 960x704 in ~ 1-2 minutes. Any help is highly appreciated.

Screenshots

No response

Additional context

No response

Contact Details

No response

[Adreitz]

I'm not seeing this on my M2 Max MBP, 64GB. Please verify the following things in Activity Monitor when you're experiencing the problem: how much memory Python is using and amounts of Cached and Swap memory (Memory tab), Python's % GPU usage (CPU tab), data written per second (Disk tab). You want to determine whether or not Invoke is still trying to use your GPU and whether memory use has increased for some reason causing data to swap between RAM and SSD, which will greatly slow your generation.

[Millu]

Can you check the size of your VRAM cache in the invoke.yaml file? If it's too high, generation speed will slow down significantly.

As a test, try setting the cache to 0.25, restarting the server and then trying a generation.

[LuLeRoemer]

I'm not seeing this on my M2 Max MBP, 64GB. Please verify the following things in Activity Monitor when you're experiencing the problem: how much memory Python is using and amounts of Cached and Swap memory (Memory tab), Python's % GPU usage (CPU tab), data written per second (Disk tab). You want to determine whether or not Invoke is still trying to use your GPU and whether memory use has increased for some reason causing data to swap between RAM and SSD, which will greatly slow your generation.

Thanks for your help!

• Python process uses around ~15 GB memory when generating images
• Swap: ~4GB
• cache: 700MB
• GPU usage: 70%
• up to 300MB/s written on disk

[LuLeRoemer]

Can you check the size of your VRAM cache in the invoke.yaml file? If it's too high, generation speed will slow down significantly.

As a test, try setting the cache to 0.25, restarting the server and then trying a generation.

Tried it but unfortunately didn't help. Default was set to 0.0.