Research of inference speed optimization methods of large language models for function calling task

This work is devoted to study and practical implementation of optimization methods (especially pruning) for large language models (LLM) in the context of function calling task, as well as comparison of accuracy and speed of the obtained models. Authors chose Mistral-7B as the basic model; glaive-function-calling-v2 – as dataset for training. 4-bit quantization in nf4 format and double quantization were used in combination with QLoRA (Quantized Low-Rank Adaptation) method. Four different pruning methods were applied for model optimization. The first method, ShortGPT, focuses on reducing the model size by trimming less significant parts. The second method is based on Taylor’s criterion for layer-by-layer pruning. The third method, LLM-Pruner, removes parameters channel-by-channel maintaining the total number of layers. The fourth method, PowerInfer, uses contextual sparsity of large language models. Optimized models were implemented for all these methods; the accuracy and speed of resulting models were compared. Results of experiments show that the highest accuracy was achieved using the layer-by-layer pruning according to Taylor’s criterion of layer importance. This method was tested with different placement of gates within the decoder layer and different ways of aggregation of layer importance on the gates. Experiments show that best results were achieved by placing the gates after Multi-Head Attention blocks and using the L2 norm of the gradient vector to aggregate layer importance. Scholarly importance of the work includes comparison of advanced pruning methods in the context of quality/speed ratio and obtaining a speed up version model for the function calling task.