Multi-head Latent Attention

An attention variant that greatly compresses KV cache usage by projecting keys and values into a low-dimensional latent vector, so that inference caches only the latent vector rather than the full K/V. It was popularized by the DeepSeek series of models.