Comment by PoignardAzur
Comment by PoignardAzur 4 days ago
Does Gemma use any specific scheme to compress embeddings? Which have you considered?
For instance, it's well-known that transformer embeddings tend to form clusters. Have you considered splitting the embedding table into "cluster centroid" and "offset from centroid" tables, where the later would presumably have a smaller range and precision?