Add to Chrome
Add to Firefox
Add to EdgePuru Sharma
AttentionStore: Cost-effective Attention Reuse across Multi-turn Conversations in Large Language Model Serving
Mar 23, 2024
Interacting with humans through multi-turn conversations is a fundamental feature of large language models (LLMs). However, existing LLM serving engines for executing multi-turn conversations are inefficient due to the need to repeatedly compute the key-value (KV) caches of historical tokens, incurring high serving costs. To address the problem, this paper proposes AttentionStore, a new attention mechanism that enables the reuse of KV caches (i.e., attention reuse) across multi-turn conversations, significantly reducing the repetitive computation overheads. AttentionStore maintains a hierarchical KV caching system that leverages cost-effective memory/storage mediums to save KV caches for all requests. To reduce KV cache access overheads from slow mediums, AttentionStore employs layer-wise pre-loading and asynchronous saving schemes to overlap the KV cache access with the GPU computation. To ensure that the KV caches to be accessed are placed in the fastest hierarchy, AttentionStore employs scheduler-aware fetching and eviction schemes to consciously place the KV caches in different layers based on the hints from the inference job scheduler. To avoid the invalidation of the saved KV caches incurred by context window overflow, AttentionStore enables the saved KV caches to remain valid via decoupling the positional encoding and effectively truncating the KV caches. Extensive experimental results demonstrate that AttentionStore significantly decreases the time to the first token (TTFT) by up to 88%, improves the prompt prefilling throughput by 8.2$\times$ for multi-turn conversations, and reduces the end-to-end inference cost by up to 56%. For long sequence inference, AttentionStore reduces the TTFT by up to 95% and improves the prompt prefilling throughput by 22$\times$.
Efficiently Supporting Hierarchy and Data Updates in DNA Storage
Dec 27, 2022
We propose a novel and flexible DNA-storage architecture that provides the notion of hierarchy among the objects tagged with the same primer pair and enables efficient data updates. In contrast to prior work, in our architecture a pair of PCR primers of length 20 does not define a single object, but an independent storage partition, which is internally managed in an independent way with its own index structure. We make the observation that, while the number of mutually compatible primer pairs is limited, the internal address space available to any pair of primers (i.e., partition) is virtually unlimited. We expose and leverage the flexibility with which this address space can be managed to provide rich and functional storage semantics, such as hierarchical data organization and efficient and flexible implementations of data updates. Furthermore, to leverage the full power of the prefix-based nature of PCR addressing, we define a methodology for transforming an arbitrary indexing scheme into a PCR-compatible equivalent. This allows us to run PCR with primers that can be variably extended to include a desired part of the index, and thus narrow down the scope of the reaction to retrieve a specific object (e.g., file or directory) within the partition with high precision. Our wetlab evaluation demonstrates the practicality of the proposed ideas and shows 140x reduction in sequencing cost retrieval of smaller objects within the partition.