Subformer is a Transformer that combines sandwich-style parameter sharing, which overcomes naive cross-layer parameter sharing in generative models, and self-attentive embedding factorization (SAFE). In SAFE, a small self-attention layer is used to reduce embedding parameter count.