Most tensor backends try to optimize performance by using a lower precision for matmuls. However, some operations (such as kq) on some models are sensitive to this and require full precision.