ScholarWorks@중앙대학교

초록

In recent years, the rapid advancement of deep learning technologies in the field of Natural Language Processing (NLP) has led to the emergence of large language models (LLM). While these large language models have demonstrated exceptional performance across various tasks, they require huge computational resources due to their massive parameter sizes. To address the high operational costs associated with training and deploying these models, techniques such as pruning and mixed-precision computation have been proposed. Pruning reduces the number and size of model parameters while mixed-precision computation optimizes computational speed and memory usage. This study evaluates the impact of mixed-precision techniques on the performance of LLM when applied at various levels of sparsity induced by pruning.

키워드