Skip to content

FrozenSFL: Model-Frozen Split Federated Learning for Resource-Efficient LLM Fine-Tuning over Heterogeneous Edge Devices

Nov 2026 · ACM Interational Symposium on Mobile Ad Hoc Networking and Computing · 0 citations · 6 references
Computer Science

TL;DR

This paper proposes FrozenSFL, a resource-efficient SFL framework for collaborative fine-tuning of LLMs across heterogeneous edge devices, and introduces a server-side feature caching mechanism that reuses invariant features, allowing subsequent training epochs to proceed offline and reducing redundant computation and communication.

Abstract

Collaborative fine-tuning of large language models (LLMs) at the network edge enables privacy-preserving and personalized generative intelligence, but is constrained by the limited memory and compute resources of edge devices. Traditional approaches, such as split federated learning (SFL), partition models between devices and a server, requiring iterative exchanges of activations/gradients while the server manages multiple submodels. This leads to substantial communication overhead, high activation storage costs, and severe server-side memory pressure when scaling to massive heterogeneous devices. In this paper, we propose FrozenSFL, a resource-efficient SFL framework for collaborative fine-tuning of LLMs across heterogeneous edge devices. At its core, FrozenSFL strictly freezes device-side submodels to eliminate both deviceside backpropagation and server-to-device gradient transmission, thereby removing on-device activation storage requirements and downlink communication overhead. Building upon this design, we improve the framework in two aspects. First, we design a layerwise LoRA aggregation strategy to handle structural heterogeneity introduced by device-specific model partitioning and enable efficient server-side updates. Second, we introduce a server-side feature caching mechanism that reuses invariant features, allowing subsequent training epochs to proceed offline and reducing redundant computation and communication. We implement the FrozenSFL system as an asynchronous multi-threaded pipeline, which serves as our real-world testbed for evaluation. On the device side, forward computation and data transmission are decoupled via a producer-consumer mechanism. On the server side, a first-in-first-out scheduling strategy serializes fine-tuning tasks across devices. A hybrid memory-disk streaming mechanism further buffers incoming data and ensures stable training under network variability. Extensive evaluations with representative models, including BERT-Base (0.11B), ViT-Large (0.3B), and LLaMA-2-7B (7B), validate the effectiveness of FrozenSFL. In particular, for LLaMA-2-7B, FrozenSFL can reduce device/server VRAM footprint by up to 14%/82.7% with only 1% accuracy loss, while achieving up to 6.7× fine-tuning speedup.

Read PDF

Similar papers

Review Open access Sep 2026

Federated Fine-Tuning of Large Language Models on Resource-Constrained Clients: Technical Approaches, Resource Costs, and Applicability

This review examines four composable routes---parameter-efficient and quantized adaptation, backpropagation-free adaptation, proxy or submodel adaptation, and split federated adaptation---through a common framework that traces the objects each endpoint retains, exchanges, and discloses.

Yun-Heng Shen, Xiao Liu, Yang Yang et al. · 1 citation
#federated learning Open access Sep 2026

Optimizing resource allocation for federated LLM training via workload forecasting and autoscaling in edge environments

An integrated framework that combines LLM fine-tuning using federated adaptive local low-rank adaptation (FedALoRA) with forecasting-driven autoscaling policies is proposed, making it well-suited for federated-LLM Autoscaling in dynamic edge environments.

Bablu Kumar, Anshul Verma, Rajkummar Buyya · 0 citations
#federated learning Book Open access Nov 2026

FedICL-KVc: Federated In-Context Learning via KV Cache

Large language models (LLMs) are increasingly constrained by the scarcity of high-quality public data, making private data essential for further improvement. However, private data is typically distributed across resource-constrained devices and cannot be shared due to strict privacy requirements, rendering conventional...

Yi-Ming Chen, Nan He, Zheng Jiang et al. · 0 citations
#artificial intelligence Preprint Oct 2026

Optimal Hydra: Model Heterogeneous Federated Learning for LLMs using Data-Informed Hydra-Nets

Large language models (LLMs) trained on vast amounts of data have demonstrated impressive performance over a range of tasks, but future performance gains will require increasing amounts of data and compute. Federated Learning (FL) facilitates training on decentralized private data, but client resource limitations can m...

G. Legate, Rahaf Aljundi, Eugene Belilovsky · 0 citations
#artificial intelligence Preprint Oct 2026

DySCo: Dynamic Sharding for Collaborative Edge-Cloud LLM Inference with Depth-Synchronized Batching

Pervasive intelligent applications are increasingly deployed on mobile and Internet of Things (IoT) edge devices. Consequently, Large Language Models (LLMs) are increasingly used to support these applications. Yet, due to their high resource demands, LLMs are mostly deployed in the cloud. Layer-wise edge-cloud inferenc...

Jing-Po Xu, Paul Joe Maliakel, Ivona Brandić et al. · 0 citations
#machine learning Preprint Sep 2026

Privacy-Preserving Split Learning for Federated LLM Fine-Tuning

This work addresses leakage through a learned obfuscate-and-recover scheme that protects participants' private datasets while still allowing an independently deployable model to be trained on the server side, making split-based federated LLM fine-tuning practically viable.

Heng Jin, Chao-Yu Zhang, He-Xuan Yu et al. · 1 citation

Related blog posts

MIT News · Artificial Intelligence Aug 27, 2026

Looking beyond natural sequences

A new machine-learning framework aims to improve the success rate of computational protein design while moving away from results that reproduce sequences found in nature.

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.