Offloading ML Processing to Storage Devices with NGD Systems