Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse natural language processing tasks, from conversational AI to code generation and content creation [1, 2, 3]. However, the deployment of these models in production environments presents significant engineering challenges. The computational demands of autoregressive text generation, combined with the massive parameter counts of modern LLMs, necessitate specialized serving infrastructure that can efficiently manage GPU resources while meeting application-specific performance requirements. The serving infrastructure for LLMs must address several competing objectives: maximizing throughput to serve many concurrent users, minimizing latency for responsive user experiences, and efficiently utilizing expensive GPU resources. Different applications prioritize these objectives differently--a chatbot requires low latency for individual requests, while a batch document processing system prioritizes throughput. This variation in requirements has led to the development of specialized serving frameworks, each making different design trade-offs. Among the available open-source solutions, vLLM [4] and HuggingFace Text Generation Inference (TGI) [5] have emerged as leading frameworks, widely adopted in both research and production settings.

KV caching significantly improves the efficiency of Large Language Model (LLM) inference by storing attention states from previously processed tokens, enabling faster generation of subsequent tokens. However, as sequence length increases, the KV cache quickly becomes a major memory bottleneck. To address this, we propose PagedEviction, a novel fine-grained, structured KV cache pruning strategy that enhances the memory efficiency of vLLM's PagedAttention. Unlike existing approaches that rely on attention-based token importance or evict tokens across different vLLM pages, PagedEviction introduces an efficient block-wise eviction algorithm tailored for paged memory layouts. Our method integrates seamlessly with PagedAttention without requiring any modifications to its CUDA attention kernels. We evaluate PagedEviction across Llama-3.1-8B-Instruct, Llama-3.2-1B-Instruct, and Llama-3.2-3B-Instruct models on the LongBench benchmark suite, demonstrating improved memory usage with better accuracy than baselines on long context tasks.

Large Language Models (LLMs) encounter severe memory inefficiencies during long-context inference due to conventional handling of key-value (KV) caches. In this work, we introduce a novel integration of PagedAttention with PyTorch's FlexAttention, addressing internal fragmentation and inefficiencies associated with monolithic KV cache allocations. Implemented within IBM's Foundation Model Stack (FMS), our fused attention kernel efficiently gathers scattered KV data. Our benchmarks on an NVIDIA L4 GPU (24GB) demonstrate significantly reduced inference latency, growing only linearly (~2x) with sequence length from 128 to 2048 tokens when utilizing a global KV cache, compared to exponential latency increases without caching. While peak memory usage remains largely unchanged for single-step evaluations (dominated by model weights and activations), paged attention causes minimal incremental memory usage, observable only at sequence lengths exceeding 2048 tokens due to its power-of-two cache allocations. We open-source the full implementation and discuss its implications for future long-context model deployment.

Efficient management of GPU memory is essential for high throughput LLM inference. Prior systems used to reserve KV-cache memory ahead-of-time that resulted in wasted capacity due to internal fragmentation. Inspired by demand paging, vLLM proposed PagedAttention to enable dynamic memory allocation for KV-cache. This approach eliminates fragmentation and improves serving throughout. However, to be able to allocate physical memory dynamically, PagedAttention changes the layout of KV-cache from contiguous virtual memory to non-contiguous virtual memory. As a consequence, one needs to rewrite the attention kernels to support paging, and implement a memory manager in the serving framework. This results in both performance and programming overheads, as well as portability challenges in adopting state-of-the-art attention kernels. In this paper, we propose vAttention, a new approach for dynamic KV-cache memory management. In contrast to PagedAttention, vAttention stores KV-cache in contiguous virtual memory and leverages OS support for on-demand allocation of physical memory. vAttention thus enables one to use state-of-the art attention kernels out-of-the-box by adding support for dynamic allocation of physical memory without having to re-write their code. We implement vAttention in the vLLM serving stack to show that it also helps improve decode throughput by up to 1.99x over vLLM, and the end-to-end serving throughput by up to 1.22x and 1.29x, compared to using the state-of-the-art PagedAttention based kernels of FlashAttention and FlashInfer.

High throughput serving of large language models (LLMs) requires batching sufficiently many requests at a time. However, existing systems struggle because the key-value cache (KV cache) memory for each request is huge and grows and shrinks dynamically. When managed inefficiently, this memory can be significantly wasted by fragmentation and redundant duplication, limiting the batch size. To address this problem, we propose PagedAttention, an attention algorithm inspired by the classical virtual memory and paging techniques in operating systems. On top of it, we build vLLM, an LLM serving system that achieves (1) near-zero waste in KV cache memory and (2) flexible sharing of KV cache within and across requests to further reduce memory usage. Our evaluations show that vLLM improves the throughput of popular LLMs by 2-4$\times$ with the same level of latency compared to the state-of-the-art systems, such as FasterTransformer and Orca. The improvement is more pronounced with longer sequences, larger models, and more complex decoding algorithms. vLLM's source code is publicly available at https://github.com/vllm-project/vllm