Offloading
8-bit models can offload weights between the CPU and GPU to support fitting very large models into memory.