After installing the ultralytics library, GPU usage decreased by half and Ksample time doubled about comfyui

[ZhenshengWu]

I have solved this problem by commenting on this line of code. Although the original Ksampler took about 0.85 milliseconds, it now takes about 0.93ms. However, it has greatly improved my system

[ltdrdata]

In my experience, the ultralytics module does something in the global scope. Instead of importing ultralytics initially, when I tried dynamically importing ultralytics inside a function, I encountered a phenomenon where it would malfunction from the second use onwards, producing unreliable results.

Perhaps this phenomenon could be due to ultralytics preemptively occupying some of torch's resources.

[bakatimeproject]

Use the package linked on here: https://github.com/comfyanonymous/ComfyUI/releases/tag/v0.0.4 it has pytorch 2.3.1

[ZhenshengWu]

使用此处链接的包：https://github.com/comfyanonymous/ComfyUI/releases/tag/v0.0.4它有 pytorch 2.3.1

Thank you for comment. However, have replaced all versions of PyTorch, including 2.1.2 to 2.4. It does not work

[bakatimeproject]

使用此处链接的包：https://github.com/comfyanonymous/ComfyUI/releases/tag/v0.0.4它有 pytorch 2.3.1

Thank you for comment. However, have replaced all versions of PyTorch, including 2.1.2 to 2.4. It does not work

Try a clean installation, back up the folders: custom nodes, models, output, input.
delete all comfyui.
Unzip, place the backup folders and run the .bat, it was the only thing that restored my speed. don't update anything. Only then did everything return to normal.