Intelligence Report

Efficient KV-Cache Reduction via Multi-Head Latent Attention: Technical Alpha: Low-rank latent projections in MLA compress the KV cache drastically while preserving attention expressivity.