Edge ML Service Scheduling for Distributed LLM Sub-Layers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Wireless communications systems face challenges in supporting large language models due to resource constraints at edge devices, leading to inefficiencies in memory and computational power usage, and underdeveloped system architecture and signaling aspects for deploying large language model sub-layers as a service.
Innovation Solution
Edge devices cooperate to provide partial large language model services by using downlink slots for input and uplink slots for output, with coordinated multi-layer machine learning processes, and utilize application and medium access control layer messaging for scheduling and result transmission.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If large language models are deployed at edge devices, then service capability is improved, but resource constraints (memory and computational power) worsen
Solution Approach 1:
The patent segments the large language model into multiple sub-layers that can be distributed across different edge devices. Each device runs only the necessary sub-layers for its specific service function, rather than deploying the entire model. This segmentation reduces the memory and computational power requirements at each individual edge device while maintaining the overall service capability of the network.
2Productivity
If multiple edge devices cooperate to provide machine learning processes as a service, then resource utilization is improved, but system architecture complexity worsens
Solution Approach 1:
The patent creates a universal service architecture where edge devices can dynamically assume different roles (service provider or service consumer) based on their capabilities and current workload. The same infrastructure and signaling mechanisms are used regardless of which device is providing which service, enabling multi-functionality without proportionally increasing system complexity.
Solution Approach 2:
The patent introduces standardized signaling protocols and message formats that act as intermediaries between cooperating edge devices. These intermediaries simplify the interaction complexity by providing predefined patterns for service discovery, capability exchange, and coordinated execution, rather than requiring custom peer-to-peer negotiation between each device pair.
3Productivity
If downlink slots are used for input and uplink slots for output in coordinated multi-layer machine learning processes, then communication efficiency is improved, but scheduling complexity worsens
Solution Approach 1:
The patent employs periodic slot allocation patterns where downlink and uplink slots are assigned in regular, predictable sequences. This periodic structure allows edge devices to anticipate when they will receive inputs or send outputs without requiring complex real-time scheduling decisions, thereby improving communication efficiency while keeping scheduling manageable through pattern-based allocation.
Data Source
AI summary
Methods, systems, and devices for wireless communication are described. A first user equipment (UE) may receive a configuration signal that indicates a set of downlink slots and a set of configured grants for communicating on a set of uplink slots. In some cases, the downlink and uplink slots may be for communications from a set of UEs including the first UE, where the set of UEs perform a coordinated multi-layer machine learning process. The first UE may transmit an uplink signal in an uplink slot from the set of uplink slots indicating a result from performing a first subset of processes of the coordinated multi-layer machine learning process. The first UE may then receive a group physical downlink control channel signal indicating that an upcoming slot is allocated to a second UE for performing a second subset of processes of the coordinated multi-layer machine learning process.