This work unifies 150+ studies on early-exit and dynamic inference models, analyzing their architectures, training strategies, exit policies, application domain, and open challenges and future directions.