Video — Deep dive — Better Attention layers for Transformer models

The self-attention mechanism is at the core of transformer models. As amazing as it is, it requires a significant amount of computing and memory bandwidth, leading to scalability issues as models get more complex and context length increases.

In this video, we’ll quickly review the computation involved in the self-attention mechanism and its multi-head variant. Then, we’ll discuss newer attention implementations focused on compute and memory optimizations, namely Multi-Query Attention, Group-Query Attention, Sliding Window Attention, Flash Attention v1 and v2, and Paged Attention.