Efficient Beam Search for Large Language Models Using Trie-Based Decoding

Abstract: In Transformer-based sequence-to-sequence generation, beam search has proven effective in enhancing the quality of generated sequences compared to greedy decoding. Conventional beam search methods typically adopt either a sequential or batch-based approach. The sequential approach, while memory-efficient, requires multiple decoding passes to construct a complete search tree, leading to significantly slower inference. On the other hand, the batch-based approach enables parallel computation across beams, but at the expense of high memory consumption due to the need to maintain separate key-value (KV) caches for each beam. In this study, we introduce a novel trie (prefix-tree)-based parallel decoding method that addresses the memory inefficiency of batch-based beam search. By sharing a single KV cache among all beams that share the same prefix, the proposed method not only reduces memory consumption dramatically but also enables parallel decoding across all branches. This innovative use of a prefix tree offers an efficient alternative for beam search, achieving significant memory savings while preserving inference speed, making it particularly well-suited for memory-constrained environments or large-scale model deployments.