Transformer architecture

Assembling the full block: attention, residual connections, layer norm, and the feed-forward MLP, stacked N times between an embedding layer and an output head. Why residuals make deep stacks trainable, and how the decoder-only variant became the template for essentially every modern LLM.