Recent advance in sparse attention mechanisms has demonstrated strong potential for reducing the computational cost of long-context training and inference in large language models (LLMs). Native Sparse Attention (NSA), one state-of-the-art approach, introduces natively trainable, hardware-aligned sparse attention that delivers substantial system-level performance boost while maintaining accuracy comparable to full attention. However, the kernel implementation of NSA forces a loop order that is only efficient with a relatively large number of query heads in each Grouped Query Attention (GQA) group, whereas existing LLMs widely adopt much smaller number of query heads in each GQA group -- such an inconsistency significantly limits the applicability of this sparse algorithmic advance. In this work, we propose Flash Sparse Attention (FSA), an alternative kernel implementation that enables efficient NSA computation across a wide range of popular LLMs with varied smaller number of heads in each GQA group on modern GPUs. Compared to vanilla NSA kernel implementation, our empirical evaluation demonstrates that FSA achieves (i) up to 3.5x and on average 1.6x kernel-level latency reduction, (ii) up to 1.25x and 1.09x on average end-to-end training speedup on state-of-the-art LLMs, and (iii) up to 1.36x and 1.11x on average for prefill-phase speedup in LLM generative inference. Github Repo at https://github.com/Relaxed-System-Lab/Flash-Sparse-Attention.

Processing long contexts has become a critical capability for modern large language models (LLMs). However, serving long-context LLMs comes with significant inference costs due to the high memory overhead of the key-value (KV) cache. Existing work leverages dynamic sparse attention algorithms (DSAes) to mitigate the KV cache overhead, but these algorithms rely on top-$k$ KV cache selection, which results in a trade-off between accuracy and efficiency. A larger $k$ improves accuracy but decreases efficiency, while a smaller $k$ boosts efficiency but compromises accuracy. To overcome this trade-off, this paper presents PSA, a $\underline{P}$rogressive $\underline{S}$parse $\underline{A}$ttention mechanism that integrates algorithmic innovations with system co-design to achieve both high inference accuracy and improved efficiency in LLM serving. The PSA algorithm adaptively adjusts the KV cache budget of different tokens and layers according to their real attention weight distributions, rather than relying on a fixed budget $k$. This enables high accuracy while minimizing KV cache usage. To further enhance execution efficiency, we introduce a pipelined iteration scheme that reduces CPU-GPU interleaving and synchronization overhead during PSA computation. Additionally, we implement unified GPU memory management that optimizes PSA's memory utilization by accounting for uneven memory requirements across different model layers. Extensive experimental results demonstrate that PSA reduces KV cache usage for attention computation by up to 2.4$\times$ and 8.8$\times$, and increases end-to-end serving throughput by up to 1.4$\times$ and 2.0$\times$, compared to state-of-the-art DSAes and systems without sparse attention, respectively.

The expanding context windows in large language models (LLMs) have greatly enhanced their capabilities in various applications, but they also introduce significant challenges in maintaining low latency, particularly in Time to First Token (TTFT). This paper identifies that the sharp rise in TTFT as context length increases is predominantly driven by queuing delays, which are caused by the growing demands for GPU Key-Value (KV) cache allocation clashing with the limited availability of KV cache blocks. To address this issue, we propose LayerKV, a simple yet effective plug-in method that effectively reduces TTFT without requiring additional hardware or compromising output performance, while seamlessly integrating with existing parallelism strategies and scheduling techniques. Specifically, LayerKV introduces layer-wise KV block allocation, management, and offloading for fine-grained control over system memory, coupled with an SLO-aware scheduler to optimize overall Service Level Objectives (SLOs). Comprehensive evaluations on representative models, ranging from 7B to 70B parameters, across various GPU configurations, demonstrate that LayerKV improves TTFT latency up to 69x and reduces SLO violation rates by 28.7%, significantly enhancing the user experience.

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