JORA: JAX Tensor-Parallel LoRA Library for Retrieval Augmented Fine-Tuning

TL;DR

This paper introduces JORA, a JAX-based framework for efficient, scalable fine-tuning of large language models like Llama-2 for retrieval-augmented tasks, significantly reducing memory use and training time.

Contribution

It presents a novel JAX-based tensor-parallel LoRA library that enables efficient, distributed fine-tuning of LLMs for retrieval tasks, addressing memory constraints.

Findings

Over 12x faster runtime compared to existing methods

Uses less than half the VRAM per GPU

Enables fine-tuning on systems with limited GPU resources

Abstract

The scaling of Large Language Models (LLMs) for retrieval-based tasks, particularly in Retrieval Augmented Generation (RAG), faces significant memory constraints, especially when fine-tuning extensive prompt sequences. Current open-source libraries support full-model inference and fine-tuning across multiple GPUs but fall short of accommodating the efficient parameter distribution required for retrieved context. Addressing this gap, we introduce a novel framework for PEFT-compatible fine-tuning of Llama-2 models, leveraging distributed training. Our framework uniquely utilizes JAX's just-in-time (JIT) compilation and tensor-sharding for efficient resource management, thereby enabling accelerated fine-tuning with reduced memory requirements. This advancement significantly improves the scalability and feasibility of fine-tuning LLMs for complex RAG applications, even on systems with limited GPU resources. Our experiments show more than 12x improvement in runtime compared to Hugging Face/DeepSpeed implementation with four GPUs while consuming less than half the VRAM per GPU.