Federated fine-tuning of large language models with much less communication
FedFit: Federated Fine-Tuning of LLMs via Vector-Bank Parameterization and Quantization
Machine Learning
Summary
Training big language models on user devices while protecting privacy usually means sending lots of data back and forth, which is slow and costly. The authors propose FedFit, a new method that reduces this data transfer by sharing compact sets of vectors instead of large model parts. They also solve a tricky math problem about combining updates from different devices accurately and efficiently. Their tests show FedFit achieves similar language understanding quality as previous methods but with up to 100 times less communication.
What this means in practice
- •For mobile app developers: Reduce data sent from phones during personalized language model fine-tuning without losing accuracy using FedFit's compact vector updates.
- •For cloud service engineers: Implement highly compressed federated updating for large language models to cut bandwidth costs and speed up training cycles.
Authors
Hang Zou, Chao Zhang, Yuzhi Yang, Yu Tian, Samson Lasaulce, Mérouane Debbah
Abstract
Federated Learning (FL) enables privacy-preserving fine-tuning of Large Language Models (LLMs), yet the massive communication overhead remains a critical bottleneck. Furthermore, applying Low-Rank Adaptation (LoRA) in FL faces a fundamental "aggregation dilemma" between the accurate Sum-of-Products (SoP) and the communication-efficient Product-of-Sums (PoS) implementations. To tackle these challenges, we propose FedFit. First, to significantly reduce communication overhead, we introduce a disjoint shared vector-bank parameterization that reconstructs high-dimensional adapter matrices from two compact and disjoint global vector banks. Second, to address the aggregation dilemma, we devise an alternating optimization schedule. By cycling between decoupled single-bank updates (which allow for accurate aggregation) and joint updates corrected by a Residual Spectral Aggregation mechanism, we resolve the conflict between SoP and PoS. Additionally, we integrate blockwise quantization with client-side error feedback to further compress the transmitted vectors. Furthermore, we establish theoretical convergence guarantees for the proposed algorithm. Extensive experiments on Qwen2.5 models demonstrate that FedFit achieves perplexity performance comparable to standard federated LoRA methods, while providing compression ratios up to 100x higher.