航天控制软件大语言模型轻量化方法OA

To address the contradiction between limited computational resources and massive parameter scales for large language models(LLMs)in aerospace control software,as well as the performance degradation issues in existing lightweight methods,a collaborative method combining knowledge asset-based layer activation pruning and adaptive quantization is proposed.First,a calibration dataset is constructed based on domain-specific knowledge assets.A layer activation metric is designed to precisely evaluate the structural importance of LLMs for pruning.Subsequently,optimal quantization bits are adaptively selected according to perplexity metrics.The entire process is training-free.Experiments on two aerospace control software LLMs,BiceLLM-7B and BiceLLM-72B,demonstrate that the proposed method achieves significant reductions of 6.6%and 7.3%in parameters,60.7%and 70.3%in storage space,45.0%and 72.1%in memory consumption,and 25%and 64.5%in inference latency,respectively,while model performance is maintained with negligible loss.This study provides an efficient and practical technical solution for lightweight deployment of LLMs in the aerospace domain by substantially reducing resource consumption while preserving model capabilities.