Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

LoRA Fine-Tuning Without GPUs: A CPU-Efficient Meta-Generation Framework for LLMs

Created by
  • Haebom

Author

Reza Arabpour, Haitz S aez de Oc ariz Borde, Anastasis Kratsios

Outline

In this paper, we propose a theoretically-backed LoRA tuning method designed for users with limited computing resources, especially those with only a CPU-based standard laptop. To overcome the limitations of existing GPU-based LoRA tuning, we propose a scheme that leverages a large number of pre-trained adapters for Mistral-7B-Instruct-v0.2 models to learn a meta-operator that maps an input dataset (represented as a probability distribution) to a set of LoRA weights. Instead of a novel gradient-based update, we generate the adapters via a lightweight composition of existing LoRAs on the CPU. Although not as performant as the GPU-trained adapters, our results consistently outperform the baseline Mistral model on subtasks, providing a practical and accessible alternative to existing GPU-based fine-tuning.

Takeaways, Limitations

•
Takeaways:
◦
Increased accessibility by enabling LLM fine-tuning even for users with limited computing resources.
◦
We present an efficient way to create a LoRA adapter using only CPU without GPU.
◦
Provides a practical alternative that delivers improved performance over the base model.
•
Limitations:
◦
Lower performance than GPU-based LoRA fine-tuning.
◦
Relies on a bank of pre-learned adapters.
◦
It may have limited generalization performance as it is specialized for a specific model (Mistral-7B-Instruct-v0.2).
👍
Made with Slashpage