Source-linked AI summary

QA-LoRA: Quantization-Aware Low-Rank Adaptation of Large Language Models

Yuhui Xu, Lingxi Xie, Xiaotao Gu, Xin Chen, Heng Chang, Hengheng Zhang, Zhengsu Chen, Xiaopeng Zhang, Qi Tian

arXiv:2309.14717v2cs.LGcs.CL

TL;DR

LLMs’ computational and memory demands hinder deployment, especially on edge devices, while combining efficient adaptation with low-bit quantization remains challenging. QA-LoRA uses group-wise operators to balance quantization and adaptation, and experiments on LLaMA families show effective accuracy and efficiency across settings.

  • Problem

    LLMs’ heavy computational and memory requirements restrict deployment on edge devices, while post-training quantization after adaptation can reduce accuracy at low bit widths.

  • Method

    QA-LoRA introduces group-wise operators that individually quantize weight groups while sharing adaptation parameters across each group.

  • Results

    QA-LoRA consistently outperforms QLoRA with post-training quantization, matches QLoRA without it, and is more than 50% faster during inference.

  • Takeaways & Limitations

    QA-LoRA provides an efficient joint quantization-and-adaptation method whose fine-tuned model remains quantized after weight integration.

  • Takeaways & Limitations

    Low-bit quantization approximates pretrained weights and may harm language-understanding accuracy; QLoRA’s post-training quantization can cause significant accuracy loss at low bit widths.

Abstract

from arXiv · show

Recently years have witnessed a rapid development of large language models (LLMs). Despite the strong ability in many language-understanding tasks, the heavy computational burden largely restricts the application of LLMs especially when one needs to deploy them onto edge devices. In this paper, we propose a quantization-aware low-rank adaptation (QA-LoRA) algorithm. The motivation lies in the imbalanced degrees of freedom of quantization and adaptation, and the solution is to use group-wise operators which increase the degree of freedom of quantization meanwhile decreasing that of adaptation. QA-LoRA is easily implemented with a few lines of code, and it equips the original LoRA with two-fold abilities: (i) during fine-tuning, the LLM's weights are quantized (e.g., into INT4) to reduce time and memory usage; (ii) after fine-tuning, the LLM and auxiliary weights are naturally integrated into a quantized model without loss of accuracy. We apply QA-LoRA to the LLaMA and LLaMA2 model families and validate its effectiveness in different fine-tuning datasets and downstream scenarios. Code will be made available at https://github.com/yuhuixu1993/qa-lora.

1 INTRODUCTION

QA-LoRA addresses the difficulty of jointly fine-tuning and quantizing LLMs for efficient deployment. It uses group-wise operators to improve low-bit quantization and simplify integration of adaptation weights, while retaining performance across model scales and quantization settings.

  • LLM deployment on edge devices is constrained by the computational burden of their many parameters, motivating quantized and fine-tuned models.
  • Post-training quantization after parameter-efficient fine-tuning can lose accuracy at low bit widths, while existing alternatives may be costly or fail to preserve quantization after fine-tuning.
  • QA-LoRA balances quantization and adaptation degrees of freedom through group-wise operators: each group is quantized individually and shares adaptation parameters.
  • QA-LoRA quantizes LLM weights during fine-tuning and integrates auxiliary weights into a lightweight quantized model without requiring post-training quantization.
  • Across LLaMA models fine-tuned on Alpaca, QA-LoRA outperforms QLoRA with post-training quantization and matches QLoRA without it, with larger advantages at lower bit widths.

2 RELATED WORK

Related work combines parameter-efficient adaptation and LLM quantization, but existing approaches leave a gap between efficient fine-tuning and efficient deployment. QA-LoRA is positioned to address this joint objective.

  • Fine-tuning LLMs with adapters: LoRA adapts pretrained models by fine-tuning two low-rank matrices while keeping most pretrained parameters unchanged.
  • Quantization of LLMs: Quantization reduces parameter or activation bit width to improve LLM efficiency and scalability, while inference methods address memory costs and quantization-induced accuracy issues.
  • Joint adaptation and quantization: Joint adaptation and quantization seeks parameter-efficient adaptation together with computation-efficient tuning and deployment, but requires handling discrete values and quantization parameters.
  • Joint adaptation and quantization: QLoRA quantizes pretrained weights to NF4 during fine-tuning but restores merged weights to FP16 afterward, leaving the deployed model slower.

3 THE PROPOSED APPROACH

QA-LoRA jointly quantizes pretrained weights during fine-tuning and preserves a quantized merged model for efficient deployment. It balances quantization and adaptation degrees of freedom through group-wise operations, retaining useful low-rank capacity while enabling quantized integration.

  • Baseline: Low-Rank Adaptation and Low-Bit Quantization: LoRA supplements fixed pretrained weights W with adjustable low-rank matrices A and B, computing y = W^⊤x + s · (AB)^⊤x.A and B have dimensions D_in × D_int and D_int × D_out, with D_int much smaller than both input and output dimensions.
  • Baseline: Low-Rank Adaptation and Low-Bit Quantization: Min-max weight quantization reduces storage and accelerates computation, but approximating W with quantized W̃ can harm language-understanding accuracy.The paper describes quantization such as FP16 to INT4 and uses scaling and zero factors with integer rounding.
  • Objective: Efficient Adaptation and Deployment: QA-LoRA targets quantized weights during fine-tuning and keeps the fine-tuned merged weights W′ quantized after adaptation for computationally efficient deployment.These are the paper’s two stated objectives for efficient adaptation and deployment.
  • Objective: Efficient Adaptation and Deployment: QLoRA quantizes weights during fine-tuning but restores FP16 after merging LoRA weights, while post-training quantization can significantly reduce accuracy at low bit widths.The paper also notes that QLoRA’s NF4 representation lacks operator-level optimization for accelerating fine-tuning and inference.
  • Solution: Group-Wise Quantization with Low-Rank Adaptation: Column-wise quantization with unconstrained LoRA weights cannot preserve quantization after merging; forcing identical rows in A makes rank(AB) = 1 and causes a significant accuracy drop.The resulting rank limitation reduces the adaptation capacity needed for new data.
  • Solution: Group-Wise Quantization with Low-Rank Adaptation: QA-LoRA partitions each weight column into L groups, using group-specific quantization factors and requiring equal A rows only within each group.A parameter-free summation reduces x from D_in dimensions to L, allowing A to have dimensions L × D_int without further constraints.
  • Solution: Group-Wise Quantization with Low-Rank Adaptation: QA-LoRA increases quantization parameters from D_out to L×D_out while reducing adaptation parameters from D_in×D_int + D_int×D_out to L×D_int + D_int×D_out.The paper presents this balance as the basis for preserving quantized integration and adaptation capacity.

4 EXPERIMENTS

Experiments evaluate QA-LoRA across LLaMA and LLaMA2 models, quantization settings, datasets, and downstream tasks. QA-LoRA generally improves accuracy and efficiency, with especially strong gains at lower bit widths and over post-quantized QLoRA.

  • 4.1 SETTINGS: Experiments use LLaMA models from 7B to 65B and LLaMA2 models at 7B and 13B, with MMLU and commonsense QA evaluation.The main quantization setting is group-wise asymmetric quantization with group size 32 using GPTQ.
  • 4.2 MAIN RESULTS AND EFFICIENCY: QA-LoRA consistently outperforms QLoRA with GPTQ and PEQA on LLaMA MMLU in both 0-shot and 5-shot accuracy.The advantage is more significant for smaller models and lower bit widths, including INT3 and INT2.
  • 4.2 MAIN RESULTS AND EFFICIENCY: More than 50% faster inference than QLoRA follows because QA-LoRA remains in INT4 after weight integration instead of converting back to FP16.INT4 operators also contribute to faster fine-tuning than QLoRA’s NF4 operators.
  • 4.2 MAIN RESULTS AND EFFICIENCY: 15.0% accuracy gain is reported for 2-bit QA-LoRA over 2-bit post-quantized QLoRA on 0-shot commonsense QA.The 4-bit version is comparable with mixed-precision QLoRA and averages a 2.0% advantage over post-quantized QLoRA.
  • 4.2 MAIN RESULTS AND EFFICIENCY: QA-LoRA generalizes to LLaMA2, where INT4 models fine-tuned on FLAN v2 consistently outperform the original FP16 models.LLaMA2 models fine-tuned on Alpaca report slightly lower accuracy than the original FP16 models.
  • 4.3 ABLATIVE STUDIES: Larger L, corresponding to smaller group size, often yields higher accuracy, especially when the quantization bit width is small.Larger L increases quantization and adaptation freedom while adding storage and computation that remain negligible when L ≫1.
  • 4.3 ABLATIVE STUDIES: On smaller fine-tuning datasets, QA-LoRA consistently outperforms mixed-precision QLoRA on LLaMA-13B and is much faster during inference.Self-instruct, Longform, and Chip2 produce somewhat weaker MMLU accuracy than Alpaca and FLAN v2.
  • 4.3 ABLATIVE STUDIES: 320K examples are sufficient for both INT2 and INT4 QA-LoRA variants, although lower-bit quantization requires more data.The dataset-size study uses LLaMA-7B and subsets of FLAN v2 ranging from 160K to 480K examples.

5 CONCLUSION

QA-LoRA introduces quantization-awareness into low-rank adaptation through group-wise operations that balance quantization and adaptation degrees of freedom. Experiments on LLaMA families validate its effectiveness across language-understanding tasks.

  • QA-LoRA introduces quantization-awareness into low-rank adaptation using group-wise operations.
  • Balancing the degrees of freedom of quantization and adaptation is QA-LoRA’s key insight.
  • Extensive experiments on the LLaMA model families validate QA-LoRA’s effectiveness.

A QUALITATIVE STUDIES

The qualitative studies compare QA-LoRA with QLoRA across quantization bit widths and conversational, categorization, HTML-generation, and dialogue prompts. QA-LoRA generally produces coherent outputs, while QLoRA with GPTQ can behave unstably, including a reported 4-bit and 2-bit crash in one case.

  • QLoRA with GPTQ shows unstable behavior across qualitative cases and crashes in one case at 4-bit and 2-bit quantization.The passage attributes this mainly to uncontrollable quantization loss during post-processing.
  • For expense-budgeting prompts, QA-LoRA produces five complete suggestions at 4-bit and 2-bit quantization.The corresponding QLoRA 2-bit output is corrupted, while its 4-bit output is shorter and less complete.
  • For social-media responses, QA-LoRA gives a detailed de-escalatory answer, whereas QLoRA’s 4-bit answer is limited to an apology.
  • For fruit categorization, QA-LoRA consistently assigns explicit colors to the listed fruits across shown bit widths.QLoRA outputs include incorrect or malformed categorizations in the displayed examples.
  • For HTML generation and beach-dialogue prompts, QA-LoRA produces structured outputs at low bit widths, while QLoRA includes malformed or corrupted text in some cases.

B SIMPLE PROOF

This section presents a simple proof explaining why QA-LoRA can be merged with quantized base models. It also notes that some notation conflicts with the main script.

  • The proof explains why QA-LoRA can be merged with quantized base models.
  • The section uses a notation set that conflicts with some letters in the main script.
  • The proof is presented as a simple explanation of the proposed approach’s mergeability.

B.1 QUANTIZATION

The quantization procedure applies group-wise quantization to columns of the pretrained weight matrix. Scaling and zero factors are defined per group, while the scaling matrix and zero matrix support de-quantization.

  • Pretrained LLM weights are quantized using post-training quantization methods such as GPTQ.
  • Each weight-matrix column is divided into groups with group size g for quantization.
  • The scaling matrix A and zero matrix B provide the factors used in the de-quantization process.

B.2 QA-LORA TRAINING

QA-LoRA aggregates input features before they enter the LoRA adapters, using windows matched to the quantization groups. The aggregated features then pass through low-rank adapters and augment the quantized-weight output.

  • QA-LoRA inserts an aggregation operation before the input reaches the LoRA adapters.
  • The aggregation window size g equals the quantization group size.
  • Features within each non-overlapping aggregation window are added, reducing the input feature dimension from h to h/g.
  • The aggregated input passes through low-rank adapters and augments the output produced by the quantized weight.

B.3 MERGE FOR INFERENCE

QA-LoRA supports merging learned LoRA adapters into quantized weights while preserving the quantization characteristic. The merged weights can therefore be deployed directly for low-bit inference.

  • QA-LoRA merges learned LoRA adapters with quantized weights while keeping their quantization characteristic.
  • The resulting merged weights support direct low-bit inference deployment.
  • The merge procedure derives the left term of the adapter expression from the quantized-weight representation.
  • The right term is expressed using P = [p_i,j] as a substitute for L1L2, with P ∈ R^(d/l)×o.
  • The adapter weights are perfectly merged by updating the zero-point matrix B to B − sL1L2 ⊘ A.
Loading 2309.14717v2…