How would you optimize a Transformer model for inference on a device with limited resources in Cisco edge or on-prem environments?
