Larger than GPU memory tensors

Hi all,

Is there an elegant way to apply a network to a tensor that is larger than GPU memory?
Tensorflow has tensorflow-mesh, maybe there’s something similar for pytorch?

I am aware of a sliding window approach, but that can lead to artifacts at edges of outputs.

Thanks!

have you looked at something like https://pytorch.org/docs/stable/checkpoint.html

I think checkpoint_sequential will do the trick, thanks!