Comment by somnial

true, but no reason the predictor model couldn't use linear attention (i.e. mamba, GDN etc) to predict KV caches