1 paper · 1 filter
Yang Liu, Zhaokai Luo, Huayi Jin +10
Multi-head latent attention (MLA) exposes many logical query heads through one packed latent KV stream. This representation is memory efficient, but it removes the physical per-hea…