What is multi-head latent attention?

Definition

Multi-head latent attention is an attention architecture that represents key and value information through a compressed latent space while retaining multiple attention heads. The design aims to reduce key-value cache size and memory bandwidth requirements.

Efficient execution depends on kernels tailored to the architecture and target hardware. A capable chip may underperform if its software stack lacks an implementation optimized for the model’s exact attention operations.

Acronyms and aliases

MLA acronym

Frequently asked questions

What does MLA stand for?

MLA stands for multi-head latent attention, an attention design that compresses key and value representations through latent variables.

Why does multi-head latent attention need specialized kernels?

Its data layout and operations must be mapped efficiently to hardware memory and compute units to deliver the intended serving benefits.

Videos explaining multi-head latent attention