Mobile terminal AI efficient deployment method and system based on multi-mode large model compression

Abstract

The invention relates to a mobile terminal AI efficient deployment method and system based on multi-modal large model compression, and the method comprises the steps: obtaining an original multi-modal large model which can process various types of data, including but not limited to texts, images and audios; and carrying out model pruning on the original multi-modal large model, and calculating the influence degree of each parameter on model output by adopting a pruning algorithm based on sensitivity analysis through analyzing the importance of each parameter in the model. According to the method, the computing resource demand is reduced, the number of parameters and the calculation amount of the multi-modal large model are effectively reduced through compression technologies such as model pruning, quantification and knowledge distillation, the model can efficiently run under limited computing resources of mobile terminal equipment, the reasoning speed of the model is increased, the storage capacity occupation is reduced, and the method is suitable for large-scale popularization and application. Model parameters are converted into low-precision representation through the quantification technology, the storage requirement of the model is greatly reduced, and the occupied storage space is further reduced through a miniaturized model obtained through knowledge distillation.

Images