Machine Learning Inference Deactivation

Description

ML inference deactivation is a management and control procedure defined in 3GPP Release 19 (21.905) that disables the inference capability of an ML inference function. It is the counterpart to ML inference activation. When an ML inference function is deactivated, it ceases to execute inference operations, meaning it no longer processes input data to produce outputs. This state transition is typically initiated by management functions (e.g., OAM), policy decisions, or autonomous network logic. Deactivation may be triggered by factors such as reduced demand, resource constraints, energy-saving policies, or the need to reconfigure the inference function.

The deactivation process involves signaling to the entity hosting the inference function, instructing it to stop inference and release associated resources (e.g., compute, memory). It may also involve gracefully terminating ongoing inference tasks and notifying dependent functions that inferences will no longer be available. The state of the inference function changes from activated to deactivated. Importantly, deactivation does not necessarily delete the ML model or the inference function itself; it merely disables the inference capability. The function can be reactivated later. This is analogous to pausing a service rather than uninstalling it.

In 3GPP networks, ML inference deactivation is relevant wherever ML inference is deployed, including RAN (e.g., gNB) and core network (e.g., NWDAF, AMF). The procedure must be coordinated with other management aspects, such as performance management (to stop collecting inference-related metrics) and fault management (to avoid false alarms due to missing inferences). Deactivation can be graceful or immediate, depending on operational requirements. Graceful deactivation allows ongoing inference requests to complete, while immediate deactivation stops them abruptly. The choice may be specified by policy or management configuration.

The concept supports dynamic resource management and energy efficiency. By deactivating inference when not needed, operators can reduce power consumption and free up computational resources for other tasks. It also enables maintenance windows and software upgrades without disrupting active inference. The separation from ML model deactivation means that a model can remain active (or not) independently; deactivating inference does not deactivate the model. This flexibility is key for managing complex ML deployments.

Overall, ML inference deactivation is an essential control mechanism for lifecycle management of ML capabilities in 3GPP networks, ensuring that inference runs only when beneficial and that resources are used efficiently.

Purpose & Motivation

ML inference deactivation exists to provide a controlled way to stop ML inference when it is no longer needed or when resources must be conserved. Without such a mechanism, inference functions might run continuously, wasting energy and computational resources. The deactivation process allows network operators to dynamically manage ML workloads, aligning with broader goals of energy efficiency and cost reduction in 5G and beyond. It also supports operational flexibility, such as temporarily disabling inference during maintenance or when input data is unreliable.

The motivation for deactivation is similar to that for activation: the need for fine-grained control over ML execution. In early ML deployments, inference often ran as a background process without explicit management, leading to inefficiencies. 3GPP introduced deactivation as part of its AI/ML management framework to standardize how inference is turned off, ensuring interoperability and predictable behavior across vendors. It also enables scenarios where inference is only needed intermittently, such as during peak hours or for specific events.

Furthermore, deactivation is crucial for resource-constrained environments, such as IoT devices or edge nodes, where compute and power are limited. By allowing inference to be disabled, these devices can conserve resources when ML is not required. The concept also facilitates graceful shutdowns, preventing abrupt termination that could disrupt dependent functions. Overall, deactivation complements activation to form a complete lifecycle management solution for ML inference in 3GPP networks.