Improving Real-Time Inference with Anchor Tokens

Table of Links

Abstract and 1 Introduction

2 Related Work

3 Anchor-based Large Language Models

3.1 Background

3.2 Anchor-based Self-Attention Networks

3.3 Anchor-based Inference

4 Experiments and 4.1 Our Implementation

4.2 Data and Training Procedure

4.3 Evaluation

5 Results

6 Analysis

7 Conclusion, Limitations, Ethics Statement, and References

A More Experimental Results

B Data Settings

3.3 Anchor-based Inference

By training the model to compress information into the anchor token of a natural language sequence, we can optimize the inference process by modifying the keys/values caching mechanism. Specifically, during inference, upon encountering an anchor token that condenses the comprehensive semantic information of preceding tokens in the current sequence, the model can reduce the keys/values caches by deleting the caches of non-anchor tokens within that sequence.

We introduce the inference method in Algorithm 1. The function “REDUCTION” in Line 1 is utilized to remove keys/values caches when the model processes prefix texts in Line 10 or generates an anchor token during the prediction of the next

token in Line 16. This approach aims to reduce the keys/values caches for both prefix tokens and generated outputs during real-time inference.