RetrievalAttention: Accelerating Long-Context LLM Inference via Vector Retrieval

Transformer-based large Language Models (LLMs) become increasingly important in various domains. However, the quadratic time complexity of attention operation poses a significant challenge for scaling to longer contexts due to the extremely high inference latency and GPU memory consumption for caching key-value (KV) vectors. This paper proposes RetrievalAttention, a training-free approach to accelerate attention computation. To leverage the dynamic sparse property of attention, RetrievalAttention builds approximate nearest neighbor search (ANNS) indexes upon KV vectors in CPU memory and retrieves the most relevant ones via vector search during generation. Due to the out-of-distribution (OOD) between query vectors and key vectors, off-the-shelf ANNS indexes still need to scan O(N) (usually 30% of all keys) data for accurate retrieval, which fails to exploit the high sparsity. RetrievalAttention first identifies the OOD challenge of ANNS-based attention, and addresses it via an attention-aware vector search algorithm that can adapt to queries and only access 1--3% of data, thus achieving a sub-linear time complexity. RetrievalAttention greatly reduces the inference cost of long-context LLM with much lower GPU memory requirements while maintaining the model accuracy. Especially, RetrievalAttention only needs 16GB GPU memory for serving 128K tokens in LLMs with 8B parameters, which is capable of generating one token in 0.188 seconds on a single NVIDIA RTX4090 (24GB).

Efficient Retrieval with Learned Similarities

Retrieval plays a fundamental role in recommendation systems, search, and natural language processing by efficiently finding relevant items from a large corpus given a query. Dot products have been widely used as the similarity function in such retrieval tasks, thanks to Maximum Inner Product Search (MIPS) that enabled efficient retrieval based on dot products. However, state-of-the-art retrieval algorithms have migrated to learned similarities. Such algorithms vary in form; the queries can be represented with multiple embeddings, complex neural networks can be deployed, the item ids can be decoded directly from queries using beam search, and multiple approaches can be combined in hybrid solutions. Unfortunately, we lack efficient solutions for retrieval in these state-of-the-art setups. Our work investigates techniques for approximate nearest neighbor search with learned similarity functions. We first prove that Mixture-of-Logits (MoL) is a universal approximator, and can express all learned similarity functions. We next propose techniques to retrieve the approximate top K results using MoL with a tight bound. We finally compare our techniques with existing approaches, showing that MoL sets new state-of-the-art results on recommendation retrieval tasks, and our approximate top-k retrieval with learned similarities outperforms baselines by up to two orders of magnitude in latency, while achieving > .99 recall rate of exact algorithms.