Martin Genzel

New paper! We compress LLM parameters down to effective lengths of 2 bits/parameter while retaining 8 bit performance. The core idea is to first quantize with an entropy regularizer and then use lossless compression in the latent space.

I’m in Vancouver! Reach out if you want to grab a coffee and talk learning theory, model compression, Markov processes, inverse problems… We’ll also present a poster at Efficient Systems for Foundation Models Workshop!