Torch cache not releasing

[modern-online]

Running on a local machine, torch cache fills up with every run. It would be good to implement cache cleaning in the code.

Also, the mesh-large model config allocates 16GB of VRAM for inference. Is there a way to reduce the batch size?