Parameter-efficient transformer embeddings replace learned embeddings with hardware-aware polynomial expansions of token IDs. - View it on GitHub