Closed zackees closed 2 months ago
Update: I don't see any GPU usage at all going on with this package. However, I get a 2x increase when I simply partitioned the source image directory into two and run rembg p
on both.
Yeah, the one runtime is not running in gpu mode
In the following function setting the CUDA_VISIBLE_DEVICES does not work. All work goes to GPU 1 (no matter what combination of settings I use), which is my smallest graphics card of the two that I have.
It's notable that running parallel rembg on partitioned folders is way faster than running all of them through just one, despite the fact it's going to just one GPU. It appears that there's a lot of computation that is being done on the CPU that is delaying the content being sent to the GPU so that when two jobs run, there is greater utilization of the CPU.
I'm not sure what is necessary to make CUDA_VISIBLE_DEVICES work with rembg. But I hope that this can be specified so that at least I can run my faster graphics card (3060) instead of 1060.