ParentFull threadjarbus·I think quantization for large language models already do something like this - they compress the parameters in memory and then decompress when performing the forward passesView on HN