ModuLoRA: Finetuning 2-Bit LLMs on Consumer GPUs by Integrating with Modular Quantizers

TL;DR

ModuLoRA introduces a memory-efficient finetuning method for large language models that enables 2-bit and 3-bit precision training on consumer GPUs by integrating modular quantization with low-rank adapters.

Contribution

It presents a novel quantization-agnostic backward pass allowing effective finetuning of ultra-low precision LLMs, outperforming previous methods in memory efficiency and performance.

Findings

Enables finetuning 2-bit and 3-bit LLMs for the first time

Achieves competitive performance on NLP tasks with less memory

Surpasses state-of-the-art ROUGE scores on summarization

Abstract

We propose a memory-efficient finetuning algorithm for large language models (LLMs) that supports finetuning LLMs with 65B parameters in 2/3/4-bit precision on as little as one 24GB GPU. Our method, modular low-rank adaptation (ModuLoRA), integrates any user-specified weight quantizer with finetuning via low-rank adapters (LoRAs). Our approach relies on a simple quantization-agnostic backward pass that adaptively materializes low-precision LLM weights from a custom black-box quantization module. This approach enables finetuning 2-bit and 3-bit LLMs for the first time -- leveraging state-of-the-art 2-bit QuIP\# quantization and 3-bit OPTQ quantization -- outperforming finetuning that relies on less sophisticated 4-bit and 8-bit methods. In our experiments, \lplora~attains competitive performance on text classification, natural language inference, and instruction following tasks using significantly less memory than existing approaches, and we also surpass the state-of-the-art ROUGE score on a popular summarization task. We release \lplora~together with a series of low-precision models as part of \llmtune, a user-friendly library for quantizing, running, and finetuning LLMs on consumer GPUs.