AhaKV: Adaptive Holistic Attention-Driven KV Cache Eviction for Efficient Inference of Large Language Models

Large Language Models (LLMs) have significantly advanced the field of Artificial Intelligence. However, their deployment is resource-intensive, not only due to the large number of model parameters but also because the (Key-Value) KV cache consumes a lot of memory during inference. While several works propose reducing the KV cache by evicting the unnecessary tokens, these approaches rely on accumulated attention score as eviction score to quantify the importance of the token. We identify the accumulated attention score is biased and it decreases with the position of the tokens in the mathematical expectation. As a result, the retained tokens concentrate on the initial positions, limiting model's access to global contextual information. To address this issue, we propose Adaptive holistic attention KV (AhaKV), it addresses the bias of the accumulated attention score by adaptively tuning the scale of softmax according the expectation of information entropy of attention scores. To make use of the holistic attention information in self-attention mechanism, AhaKV utilize the information of value vectors, which is overlooked in previous works, to refine the adaptive score. We show theoretically that our method is well suited for bias reduction. We deployed AhaKV on different models with a fixed cache budget. Experiments show that AhaKV successfully mitigates bias and retains crucial tokens across global context and achieve state-of-the-art results against other related work on several benchmark tasks.

Essence Knowledge Distillation for Speech Recognition

It is well known that a speech recognition system that combines multiple acoustic models trained on the same data significantly outperforms a single-model system. Unfortunately, real time speech recognition using a whole ensemble of models is too computationally expensive. In this paper, we propose to distill the knowledge of essence in an ensemble of models (i.e. the teacher model) to a single model (i.e. the student model) that needs much less computation to deploy. Previously, all the soften outputs of the teacher model are used to optimize the student model. We argue that not all the outputs of the ensemble are necessary to be distilled. Some of the outputs may even contain noisy information that is useless or even harmful to the training of the student model. In addition, we propose to train the student model with a multitask learning approach by utilizing both the soften outputs of the teacher model and the correct hard labels. The proposed method achieves some surprising results on the Switchboard data set. When the student model is trained together with the correct labels and the essence knowledge from the teacher model, it not only significantly outperforms another single model with the same architecture that is trained only with the correct labels, but also consistently outperforms the teacher model that is used to generate the soft labels.