search

comfyanonymous / ComfyUI

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface.
https://www.comfy.org/
GNU General Public License v3.0
55.8k stars 5.89k forks source link

Higher GPU usage, with no performance gains #5172

Open JorgeR81 opened 1 month ago

JorgeR81 commented 1 month ago

Expected Behavior

Don't have extra GPU usage, if it does not improve performance.

Actual Behavior

I have a GTX 1070. My GPU usage was always low in the KSampler stage ( less than 5 % ).

GPU usage was also low with Flux, when it was first implemented.  But after some updates, the GPU usage got higher ( about 60 % ).

At first thought this was related to the Flux speed improvements. Because, currently I have faster both performance and higher GPU usage, than when Flux was first implemented.

But I remember using some past Comfy UI versions ( maybe a month ago ) which had low GPU usage ( less than 5 % ) and the same fast performance.   I don't know the exact versions that gave fast performance, with low GPU usage, because I use GGUF models more regularly.  And currently, with GGUF models, I always have fast performance and low GPU usage. I only noticed this issue recently, when I tried a Flux FP8 fine tune. 

Steps to Reproduce

.

Debug Logs

.

Other

Windows 10 

Total VRAM 8192 MB, total RAM 32727 MB
pytorch version: 2.3.1+cu121
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce GTX 1070 : cudaMallocAsync
JorgeR81 commented 1 month ago

But I remember using some past Comfy UI versions ( maybe a month ago ) which had low GPU usage ( less than 5 % ) and the same fast performance.

Maybe I'm getting confused and FP8 always required high GPU usage for faster performance. 

But in this case, can't this be optimized, like with GGUF models. Because I also had the same high GPU usage ( about 60% ) with GGUF models, but this is now improved ( less than 5% ).