Pre-LN Architecture

An architecture design that applies LayerNorm before each Transformer sublayer, keeping the residual path as an identity mapping. This makes training more stable and scalable to deeper networks, and it is the mainstream choice in current large models.