Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Shears: Unstructured Sparsity with Neural Low-rank Adapter Search

J. Pablo Muñoz, Jinjie Yuan, Nilesh Jain

TL;DR

Shears addresses the cost and performance barriers of fine-tuning large language models by fusing unstructured sparsification with a Neural Low-Rank Adapter Search (NLS) to optimize elastic LoRA adapters on a sparsified base model. The method prunes the base model with Wanda, freezes the pruned weights, and trains a super-adapter network whose sub-adapters are discovered via NLS and a lightweight sub-adapter search. Across math and commonsense tasks on LLaMA and MPT models, Shears achieves high sparsity (up to 50–60% in some setups) with accuracy comparable to or exceeding traditional PEFT baselines and substantially better than full fine-tuning in many cases, while requiring far less compute and memory. This yields practical, deployment-friendly PEFT with compression, enabling efficient fine-tuning and inference on limited hardware, with code and models made available for reproducibility.

Abstract

Recently, several approaches successfully demonstrated that weight-sharing Neural Architecture Search (NAS) can effectively explore a search space of elastic low-rank adapters (LoRA), allowing the parameter-efficient fine-tuning (PEFT) and compression of large language models. In this paper, we introduce a novel approach called Shears, demonstrating how the integration of cost-effective sparsity and a proposed Neural Low-rank adapter Search (NLS) algorithm can further improve the efficiency of PEFT approaches. Results demonstrate the benefits of Shears compared to other methods, reaching high sparsity levels while improving or with little drop in accuracy, utilizing a single GPU for a pair of hours.

Shears: Unstructured Sparsity with Neural Low-rank Adapter Search

TL;DR

Shears addresses the cost and performance barriers of fine-tuning large language models by fusing unstructured sparsification with a Neural Low-Rank Adapter Search (NLS) to optimize elastic LoRA adapters on a sparsified base model. The method prunes the base model with Wanda, freezes the pruned weights, and trains a super-adapter network whose sub-adapters are discovered via NLS and a lightweight sub-adapter search. Across math and commonsense tasks on LLaMA and MPT models, Shears achieves high sparsity (up to 50–60% in some setups) with accuracy comparable to or exceeding traditional PEFT baselines and substantially better than full fine-tuning in many cases, while requiring far less compute and memory. This yields practical, deployment-friendly PEFT with compression, enabling efficient fine-tuning and inference on limited hardware, with code and models made available for reproducibility.

Abstract

Recently, several approaches successfully demonstrated that weight-sharing Neural Architecture Search (NAS) can effectively explore a search space of elastic low-rank adapters (LoRA), allowing the parameter-efficient fine-tuning (PEFT) and compression of large language models. In this paper, we introduce a novel approach called Shears, demonstrating how the integration of cost-effective sparsity and a proposed Neural Low-rank adapter Search (NLS) algorithm can further improve the efficiency of PEFT approaches. Results demonstrate the benefits of Shears compared to other methods, reaching high sparsity levels while improving or with little drop in accuracy, utilizing a single GPU for a pair of hours.
Paper Structure (28 sections, 3 equations, 2 figures, 9 tables)

This paper contains 28 sections, 3 equations, 2 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Sparsification
  4. Low-Rank Adaptation
  5. Methodology
  6. Unstructured Sparsification
  7. Super-Adapter Training
  8. Neural Low-Rank Adapter Search (NLS)
  9. Sub-Adapter Search
  10. Experiments
  11. Experimental Setup
  12. Datasets.
  13. Models.
  14. Baselines.
  15. Comparison to LLM-Adapters
  16. ...and 13 more sections

Figures (2)

  • Figure 1: Shears workflow. Initially, Shears employs a zeroth-order pruning algorithm to induce sparsity in the given LLM. Subsequently, the framework generates a super-adapter network trained by activating subnetworks within the search space of elastic adapters. Finally, Shears yields sub-adapter networks that exhibit high performance.
  • Figure 2: Comparison of Shears and Sparse Fine-tuning (SparseFT) kurtic2023sparse on the GSM8K test dataset.