Large Language Models (LLMs) are complex artificial intelligence tools capable of amazing feats in natural language processing. However, these large models require extensive fine-tuning to adapt to specific tasks, a process that usually involves adjusting a considerable number of parameters and consequently consuming significant computational resources and memory. This means the fine-tuning of LLMs is very resource-intensive, especially for complex and information-rich tasks. Therefore, fine-tuning could possibly go beyond the capacities of ordinary computing setups.
To address this hurdle, scientists from Shandong University, Carnegie Mellon University, the Academy of Mathematics and Systems Science, and Leiden University have developed the Memory-Efficient Fine-Tuning (MEFT) method. This novel approach leverages the inherent sparsity in the Feed-Forward Networks (FFNs) of LLMs and the larger capacity of Central Processing Unit (CPU) memory compared to Graphics Processing Unit (GPU) memory, which is ordinarily used.
The MEFT approach stores and updates larger adapter parameters on the CPU. It employs a Mixture of Experts (MoE)-like architecture to optimize computations and reduce the amount of communication between the CPU and GPU. This approach makes use of sparse activation, in which only neurons that are crucial for the input are activated. These computations ensure that most parameters remain on the CPU, which reduces the volume of communication and memory usage on the GPU, making the process more efficient while using fewer resources.
The researchers tested MEFT on two models, LLaMA-7B and Mistral-7B, using four different datasets: Natural Questions (NQ), SQuAD, ToolBench, and GSM8K. They found that MEFT reduced GPU memory usage by a staggering 50%, bringing the demands from 48GB down to 24GB. Meanwhile, performance was comparable to other, more resource-heavy full fine-tuning methods.
The researchers have demonstrated the efficiency and potential of MEFT for scaling model fine-tuning tasks by significantly reducing computational resources and memory requirements without compromising performance. This innovation paves the way for a more cost-effective and efficient means of adapting LLMs to specific tasks, which could be a huge step forward for natural language processing.
