Hi, what are the memory requirements for backpropagation through zeroscope model (TextToVideoSDPipeline)? I am getting out of memory problems when backprop with a batch size >=2 (using 80GB A100 card)
Did you resolve this issue? I face similar memory constraints at the moment. setup: 24 frames, using manual mixed fp16 and fp32, 48 GB GPU. I would like to bring this further down.