Apparatus, method and computer program

Abstract

A request is received to offload of at least one inference step of a machine learning (ML) model for an application from a user equipment (UE) 900 to at least one of a plurality of host nodes 901, 902 of a network. Information is then acquired, and used to determine at least one deployment option (federated inference orchestration) for offloading the ML inference step. The information comprises first information relating to the application (such as a quality of service (QoS) indicator, time sensitivity information, hardware requirements or data privacy requirements), second information relating to the machine learning model (such as input or output data size, computing complexity or inference step parameter size), and third information relating to at least one of the plurality of host nodes. This third information includes an indication of the energy efficiency of the host node(s), in the form of the carbon emission footprint associated with the host node(s). An indication of the determined deployment option(s) is then provided, so that the ML inference step(s) can be offloaded from the UE to the host node(s) in an environmentally-aware manner.

Images