Search papers, labs, and topics across Lattice.
This paper investigates the impact of factor sharing strategies in federated low-rank adaptation (LoRA) for fine-tuning large language models, specifically comparing the approaches of sharing the input factor $A$ while keeping the output factor $B$ local, versus sharing $B$ while keeping $A$ local. By analyzing projection residuals, the authors establish that the optimal strategy minimizes aggregate residuals across clients, leading to the development of Federated Adaptive Factor Sharing Low-Rank Adaptation (FedAS-LoRA). The proposed method incorporates a Rank-Aware Shared-Subspace Sufficiency (RSS) metric to assess the adequacy of shared input subspaces, demonstrating superior performance across various tasks and data distributions in experiments.
Sharing the right LoRA factor can significantly enhance fine-tuning performance in federated learning, with a novel adaptive strategy that outperforms traditional methods.
Low-rank adaptation (LoRA) represents large language model (LLM) updates with two compact matrix factors, i.e., $A$ and $B$, providing an efficient way to fine-tune large models in federated learning paradigm. Inspired by the asymmetric roles of the LoRA factors, we study whether $A$ should be shared across clients while $B$ remains client-specific (Share-A/Local-B), or whether $B$ should instead be shared while $A$ remains client-specific (Share-B/Local-A). With a least-squares surrogate, we reveal that Share-A/Local-B requires the client-specific LoRA update matrices to use a common rank-$r$ input-side space, whereas Share-B/Local-A requires a common rank-$r$ output-side space. The two strategies therefore incur different projection residuals, indicating that the preferred strategy is the one with the smaller aggregate residual across clients. With this insight, we propose Federated Adaptive Factor Sharing Low-Rank Adaptation (FedAS-LoRA), which selects the sharing side before training to enhance fine-tuning performance. To enable adaptive factor selection before training, we design a Rank-Aware Shared-Subspace Sufficiency (RSS) metric, which effectively assesses whether a shared rank-$r$ input subspace is sufficient for the local data distributions using representations extracted from a frozen LLM backbone. Experiments across different tasks, data distributions, LoRA ranks, and participation settings confirm the effectiveness of RSS and the superior performance of FedAS-LoRA.
