We introduce WINA (Weight-Informed Neuron Activation), a novel framework designed to enhance the efficiency of Large Language Models (LLMs) without requiring retraining. Traditional methods often struggle with the significant computational and memory demands of LLMs, leading to high costs and slow performance. WINA addresses these challenges by employing a dual-criterion approach for sparse activation, which intelligently selects neurons based on both their current signal strength and their potential downstream influence. This training-free optimization dramatically reduces inference costs and speeds up processing, offering substantial improvements in accuracy and efficiency compared to previous sparse activation techniques. Ultimately, WINA enables LLMs to operate more efficiently on various platforms, from cloud services to mobile devices, signaling a significant step toward more practical and cost-effective AI deployment.

New tracks tagged #SparseActivation