Language-Specific Pruning for Efficient Reduction of Large Language Models

AbstractDelving into pruning techniques is essential to boost the efficiency of Large Language Models (LLMs) by reducing their size and computational demands, resulting in faster and more cost-effective inference. In this work, our key contribution lies in recognizing that LLMs trained on diverse languages manifest distinct language-specific weight distributions. Exploiting this insight, we illustrate that pruning LLMs using language-specific data results in a more potent model compression. Empirical evidence underscores the critical nature of pruning on language-specific data, highlighting a noteworthy impact on the perplexity of Ukrainian texts compared to pruning on English data. The proposed methodology significantly reduces the size of LLaMA, LLaMA 2 and Mistral models while preserving competitive performance. This research underscores the significance of linguistic considerations in LLM pruning and advocates for language-specific optimization, establishing a framework for more efficient and tailored language models across diverse linguistic contexts. Additionally, all experiments were conducted using a single consumer-grade NVIDIA RTX 3090 GPU, and the code is available at https://github.com/mshamrai/language-specific-pruning.