FSLoRA enables collaborative fine-tuning of large language models across devices with varying computational capabilities, preserving privacy while maximizing performance.

• Addresses the challenge of resource heterogeneity in on-device LLM fine-tuning
• Uses sketching techniques to compress high-rank adaptation matrices into lower dimensions
• Achieves performance comparable to centralized training while reducing communication costs
• Preserves data privacy through federated learning principles

This innovation brings enterprise-grade LLM fine-tuning capabilities to edge computing scenarios, enabling customization of large models without compromising user privacy or requiring powerful hardware.

Federated Sketching LoRA: On-Device Collaborative Fine-Tuning of Large Language Models