How to reduce the memory usage of DKM algorithm?

I am trying to use the DKM to palletize the weights of my Torch model during training.
However, on a somewhat large model ~500M parameters, this always leads to OOM.
Therefore, I would like to ask if there is a way to use DKM (or any type of training time palletization) with larger models?

Start by reviewing the DKM implementation and how it is used with Torch models during training. The issue names no files or tests, so identify the relevant entry points and establish whether reducing memory use or supporting larger models is feasible; done should include a documented approach or working training-time palletization for models around 500M parameters.

Written by the indexing model from the issue text.