Using CleanDIFT leads to memory leak

[Bavesh-B]

I'm integrating cleanDIFT with some other modules for my dense prediction task. However, when I train the entire network (as in other modules of the network while the diffusion model is frozen), I am getting memory leaks (my GPU memory slowly keeps increasing subsequently leading to CUDA OOM error). Is there any reason for this phenomenon as a lot of people report memory leaks with different stable diffusion models but no real solution is available anywhere. Any help would be really appreciated!!

[KoljaBauer]

Hi, sorry to hear that, we never experienced any memory leaks during the training of our models. Are you sure that the OOM error is not caused by the other modules you are referring to, or that you simply don't have enough GPU memory for all modules combined? Is it already failing in the first iteration or after multiple iterations?