How would you implement speculative decoding to accelerate inference speeds in a production environment at Relace?
