GRASS: Gradient-based Adaptive Layer-wise Importance Sampling for Memory-efficient Large Language Model Fine-tuning
2026-04-09 • Computation and Language
Computation and LanguageMachine Learning
AI summaryⓘ
The authors address the problem of fine-tuning large language models, which usually needs a lot of GPU memory. They propose GRASS, a method that decides which parts (layers) of the model to update by looking at the size of the gradients, which reflect how important each layer is for the current task and training stage. This adaptive approach improves efficiency and performance compared to fixed layer selection methods. Additionally, they introduce a way to manage memory better during training without slowing it down. Their experiments show that GRASS both improves accuracy and reduces memory usage.
large language modelsfine-tuningGPU memorylow-rank adaptationlayer-wise importance samplinggradient normadaptive trainingoptimizer state offloading
Authors
Kaiyuan Tian, Yu Tang, Gongqingjian Jiang, Baihui Liu, Yifu Gao, Xialin Su, Linbo Qiao, Dongsheng Li
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\%.