"Understanding Transformers and Attention Mechanisms" - a concise mathematical introduction to the attention mechanism, one of the key ideas in modern language models.

This document explains tokenization and embeddings, queries, keys, and values, attention scores and weights, multi-head attention, self-attention, causal attention and masking, cross-attention, and the basic structure of the Transformer architecture.

It also presents important techniques that make the attention mechanism in modern LLMs more efficient: KV-caching, grouped query attention (GQA), multi-query attention (MQA), and latent attention.

https://arxiv.org/pdf/2604.00965