GRASS: Gradient-based Adaptive Layer-wise Importance Sampling for Memory-efficient Large Language Model Fine-tuning

TL;DR

GRASS is a gradient-based adaptive importance sampling method that improves memory efficiency and performance in large language model fine-tuning by dynamically estimating layer importance.

Contribution

It introduces a task-aware, training-stage-aware importance metric and an adaptive sampling strategy for layer-wise fine-tuning, outperforming existing methods.

Findings

Achieves up to 4.38 point accuracy improvement.

Reduces memory usage by up to 19.97%.

Outperforms state-of-the-art methods across multiple benchmarks.

Abstract

Full-parameter fine-tuning of large language models is constrained by substantial GPU memory requirements. Low-rank adaptation methods mitigate this challenge by updating only a subset of parameters. However, these approaches often limit model expressiveness and yield lower performance than full-parameter fine-tuning. Layer-wise fine-tuning methods have emerged as an alternative, enabling memory-efficient training through static layer importance sampling strategies. However, these methods overlook variations in layer importance across tasks and training stages, resulting in suboptimal performance on downstream tasks. To address these limitations, we propose GRASS, a gradient-based adaptive layer-wise importance sampling framework. GRASS utilizes mean gradient norms as a task-aware and training-stage-aware metric for estimating layer importance. Furthermore, GRASS adaptively adjusts layer sampling probabilities through an adaptive training strategy. We also introduce a layer-wise optimizer state offloading mechanism that overlaps computation and communication to further reduce memory usage while maintaining comparable training throughput. Extensive experiments across multiple models and benchmarks demonstrate that GRASS consistently outperforms state-of-the-art methods, achieving an average accuracy improvement of up to 4.38 points and reducing memory usage by up to 19.97\%.