WActiGrad: Structured Pruning for Efficient Finetuning and Inference of Large Language Models on AI Accelerators

Abstract

Large Language Models (LLMs) have shown remarkable performance across various language processing applications. Nevertheless, their extensive computational requirements could hinder their deployment in real-time applications or resource-constrained environments. Pruning is a powerful technique to reduce the model size and make it computationally efficient. In this paper, we propose a structured pruning algorithm, Weight Activation and Gradient (WActiGrad), to obtain smaller LLMs from large pre-trained models. We investigate the level of granularity at which structured pruning techniques can be applied to an LLM and identify the challenges in applying these techniques across different parts of the transformer. Finally, based on these observations, we develop a pruning methodology that is adaptable to various attention and feedforward network modules. We comprehensively assess our WActiGrad method on state-of-the-art LLMs, LLaMA (7B and 13B), LLaMA-2 (7B and 13B), and Mistral-7B models across several language benchmarks for post-pretraining. This approach can prune close to 20% of the original model size without significantly compromising the model validation accuracy. We evaluate the hardware performance of our structurally pruned LLMs on different AI accelerators such as Nvidia A100 GPU, Groq LPU, Cerebras CS-2, and Graphcore Bow systems to show the effectiveness of the structured pruning technique. The findings presented in this paper offer insights into the integration of structured pruning techniques deployment on AI accelerators.