Resource Management for Low-latency Cooperative Fine-tuning of Foundation Models at the Network Edge

TL;DR

This paper proposes a resource management framework for low-latency cooperative fine-tuning of foundation models at the network edge, optimizing computation and communication to enable efficient multi-device adaptation.

Contribution

It introduces the DEFT paradigm with a novel CRUNCH algorithm and a joint bandwidth-and-block allocation method for low-latency model fine-tuning at the edge.

Findings

Significant latency reduction demonstrated on GLUE benchmark.

Effective multi-device cooperation for fine-tuning large models.

Optimized resource allocation improves edge device performance.

Abstract

The emergence of large-scale foundation models (FoMo's) that can perform human-like intelligence motivates their deployment at the network edge for devices to access state-of-the-art artificial intelligence. For better user experiences, the pre-trained FoMo's need to be adapted to specialized downstream tasks through fine-tuning techniques. To transcend a single device's memory and computation limitations, we advocate multi-device cooperation within the device-edge cooperative fine-tuning (DEFT) paradigm, where edge devices cooperate to simultaneously optimize different parts of fine-tuning parameters within a FoMo. However, the parameter blocks reside at different depths within a FoMo architecture, leading to varied computation latency-and-memory cost due to gradient backpropagation-based calculations. The heterogeneous on-device computation and memory capacities and channel conditions necessitate an integrated communication-and-computation allocation of local computation loads and communication resources to achieve low-latency (LoLa) DEFT. To this end, we consider the depth-ware DEFT block allocation problem. The involved optimal block-device matching is tackled by the proposed low-complexity Cutting-RecoUNting-CHecking (CRUNCH) algorithm, which is designed by exploiting the monotone-increasing property between block depth and computation latency-and-memory cost. Next, the joint bandwidth-and-block allocation makes the problem more sophisticated. We observe a splittable Lagrangian expression through the transformation and analysis of the original problem, where the variables indicating device involvement are introduced. Then, the dual ascent method is employed to tackle this problem iteratively. Through extensive experiments conducted on the GLUE benchmark, our results demonstrate significant latency reduction achievable by LoLa DEFT for fine-tuning a RoBERTa model.