upvote
The library I use for archiving compressed raw LLM weights ([1] , posted in my other comment) afaik at least reorders the weights to more easily compress them. They also have a paper linked, which I admittedly didn't read, that might have more background and tradeoffs. Likely they will have discussed methods to take advantage of the structure of weights.

[1] https://github.com/zipnn/zipnn#results

reply