low gpu-util

[SongYxing]

I run in the A100 80g bur the gpu-util is keeping lower than 20%. It makes the bad inference speed (60~80 second / frame).

[williamyang1991]

I suppose that's the ebsynth part (not fully GPU-based, mostly CPU-based). Ebsynth is not implemented by us. If you are interested, you can improve the Ebsynth part to make a high gpu-util.

Our diffusion part does not have low gpu-util.

[SongYxing]

I am sorry. But (60~80 second / frame) infers to the ddim phase speed.

[williamyang1991]

But in my case, on V100 32G, the speed is much faster than your case. (512x512 videos, take about 14.23s per frame) Did you use very large video resolution?

[SongYxing]

I also fixed the 512 x 512 resolution. And varm cost about 10g. In my case, we spend about 200s to load the model and inference per frame cost 60s. I first though it work in the CPU, but i check it work in CUDA:4 (I set it through torch.set.cuda). And it is difficulty to set gpu id in weiui.py. CUDA_VISIA... and os cmd are not working. It only work when i set it in front of the per process_x.

[williamyang1991]

I remember that I can set the GPU by CUDA_VISIBLE_DEVICES=0 python webUI.py. The slow speed might also caused by WebUI? You can try to use rerender.py instead to test the speed. If rerender.py is also slow, then I think there may be something wrong with your hardware or software.

[SongYxing]

rerender.py also performs bad. I test another rep (fresco)

[williamyang1991]

https://github.com/williamyang1991/Rerender_A_Video/blob/main/inference_playground.ipynb You can see the recorded speed in my inference_playground.ipynb

[SongYxing]

thanks for your patient resp

[SongYxing]

this image shows the speed

[williamyang1991]

I cannot help you with it. I think it is not the problem of my code. Your hardware or software or setting or environment may have some problem that slow down the code.