PagedAttention is an attention algorithm for efficient serving of large language models (LLMs). It was introduced in 2023 by Woosuk Kwon and colleagues in the paper Efficient Memory Management for Large Language Model Serving with PagedAttention, alongside the vLLM serving engine. The method stores the key–value cache used during autoregressive decoding in fixed-size blocks that can be mapped to non-contiguous physical memory, borrowing ideas from virtual memory, paging, and operating system design.
Background In transformer inference, the key–value cache grows with sequence length and the number of concurrent requests. Kwon et al. argued that earlier serving systems typically reserved contiguous cache regions in advance, which caused reserved space, internal fragmentation, and external fragmentation. In their experiments, the paper reported that the effective memory utilization of previous systems could fall as low as 20.4%.
Description PagedAttention partitions the cache of each sequence into fixed-size KV blocks. A request's cache is represented as a sequence of logical blocks, while a block table maps those logical blocks to physical GPU-memory blocks. As a result, neighboring logical blocks do not need to be contiguous in physical memory, and new blocks can be allocated on demand as generation proceeds. The design also makes it easier to share cache state across related decoding paths. In vLLM, physical blocks can be reference-counted and shared among requests or branches, with block-granularity copy-on-write used when a shared block must be modified. The original paper applied this design to parallel sampling, beam search, and prompts with shared prefixes.
Mathematical formulation For a query token i {\displaystyle i} in causal self-attention, the standard attention output can be written as
a i j = exp ( q i ⊤ k j / d ) ∑ t = 1 i exp ( q i ⊤ k t / d ) , o i = ∑ j = 1 i a i j v j {\displaystyle a_{ij}={\frac {\exp(\mathbf {q} _{i}^{\top }\mathbf {k} _{j}/{\sqrt {d}})}{\sum _{t=1}^{i}\exp(\mathbf {q} _{i}^{\top }\mathbf {k} _{t}/{\sqrt {d}})}},\;\mathbf {o} _{i}=\sum _{j=1}^{i}a_{ij}\mathbf {v} _{j}}
where q i {\displaystyle \mathbf {q} _{i}} , k j {\displaystyle \mathbf {k} _{j}} , and v j {\displaystyle \mathbf {v} _{j}} are the query, key, and value vectors, and d {\displaystyle d} is the attention dimension. If the cache is partitioned into blocks of size B {\displaystyle B} , the key and value blocks may be written as
K j = ( k ( j − 1 ) B + 1 , … , k j B ) , V j = ( v ( j − 1 ) B + 1 , … , v j B ) {\displaystyle \mathbf {K} _{j}=(\mathbf {k} _{(j-1)B+1},\ldots ,\mathbf {k} _{jB}),\;\mathbf {V} _{j}=(\mathbf {v} _{(j-1)B+1},\ldots ,\mathbf {v} _{jB})}
PagedAttention then performs the computation blockwise:
… excerpt ends here. Continue reading the full article.
