This allows offloading model parameters to the CPU.