The growth of million-token LLMs exposes the scalability limits of inference systems, where the KVCache dominates memory usage and data transfer overhead. Recent offloading systems migrate the KVCache to CPU memory and incorporate top-k attention to reduce the volume of data transferred from the CPU, while further applying system-level optimizations such as on-GPU caching and prefetching to lower transfer overhead. However, they overlook the CPU bottleneck in three aspects: (1) substantial overhead of fine-grained dynamic cache management performed on the CPU side, (2) significant transfer overhead from poor PCIe bandwidth utilization caused by heavy gathering operations at the CPU side, and (3) GPU runtime bubbles introduced by coarse-grained CPU-centric synchronization. To address these challenges, we propose CLO, a CPU-light KVCache offloading system via algorithm-system co-design. CLO features: (1) a coarse-grained head-wise approximate on-GPU caching strategy with negligible cache management cost, (2) seamless combination of data prefetching and on-GPU persistent caching for lower transfer overhead, (3) a zero-copy transfer engine to fully exploit PCIe bandwidth, and a GPU-centric synchronization method to eliminate GPU stalls. Evaluation on two widely-used LLMs demonstrates that CLO achieves comparable accuracy to state-of-the-art systems, while substantially minimizing CPU overhead, fully utilizing PCIe bandwidth, thus improving decoding throughput by 9.3%-66.6%. Our results highlight that algorithm-system co-design is essential for memory-constrained LLM inference on modern GPU platforms. We open source CLO at https://github.com/CommediaJW/CLO.

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.

The Transformer architecture, underpinned by the Multi-Head Attention (MHA) mechanism, has become the de facto standard for state-of-the-art models in artificial intelligence. However, the quadratic computational complexity of MHA with respect to sequence length presents a significant barrier to scaling, particularly for applications involving long contexts. Prevailing solutions, such as Multi-Query Attention (MQA) and Grouped-Query Attention (GQA), have effectively addressed the memory bandwidth bottleneck that dominates autoregressive inference latency by sharing Key and Value projections. While highly successful, these methods do not reduce the fundamental number of floating-point operations (FLOPs) required for the attention score computation, which remains a critical bottleneck for training and full-sequence processing. This paper introduces Sparse Query Attention (SQA), a novel attention architecture that pursues an alternative and complementary optimization path. Instead of reducing Key/Value heads, SQA reduces the number of Query heads. This architectural modification directly decreases the computational complexity of the attention mechanism by a factor proportional to the reduction in query heads, thereby lowering the overall FLOPs. This work presents the theoretical foundation of SQA, its mathematical formulation, and a family of architectural variants. Empirical benchmarks on long sequences (32k-200k tokens) demonstrate that SQA can achieve significant throughput improvements of up to 3x in computation-bound scenarios such as model pre-training, fine-tuning, and encoder-based tasks, with only a minimal impact on model quality in preliminary smallscale experiments. SQA was discovered serendipitously during the development of the upcoming Reactive Transformer architecture, suggesting its potential as a powerful tool for building more efficient and scalable models

LLM decoding is bottlenecked for large batches and long contexts by loading the key-value (KV) cache from high-bandwidth memory, which inflates per-token latency, while the sequential nature of decoding limits parallelism. We analyze the interplay among arithmetic intensity, parallelization, and model quality and question whether current architectures fully exploit modern hardware. This work redesigns attention to perform more computation per byte loaded from memory to maximize hardware efficiency without trading off parallel scalability. We first propose Grouped-Tied Attention (GTA), a simple variant that combines and reuses key and value states, reducing memory transfers without compromising model quality. We then introduce Grouped Latent Attention (GLA), a parallel-friendly latent attention paired with low-level optimizations for fast decoding while maintaining high model quality. Experiments show that GTA matches Grouped-Query Attention (GQA) quality while using roughly half the KV cache and that GLA matches Multi-head Latent Attention (MLA) and is easier to shard. Our optimized GLA kernel is up to 2$\times$ faster than FlashMLA, for example, in a speculative decoding setting when the query length exceeds one. Furthermore, by fetching a smaller KV cache per device, GLA reduces end-to-end latency and increases throughput in online serving benchmarks by up to 2$\times$.

Generating long sequences of tokens given a long-context input imposes a heavy computational burden for large language models (LLMs). One of the computational bottleneck comes from computing attention over a long sequence of input at each generation step. In this paper, we propose Recycled Attention, an inferencetime method which alternates between full context attention and attention over a subset of input tokens. When performing partial attention, we recycle the attention pattern of a previous token that has performed full attention and attend only to the top K most attended tokens, reducing the cost of data movement and attention computation. Compared to previously proposed inference-time acceleration method which attends only to local context or tokens with high accumulative attention scores, our approach flexibly chooses tokens that are relevant to the current decoding step. We evaluate our methods on RULER, a suite of tasks designed to comprehensively evaluate long-context abilities, and long-context language modeling tasks. Applying our method to off-the-shelf LLMs achieves comparable speedup to baselines which only consider local context while improving the performance by 2x. We further explore two ideas to improve performance-efficiency trade-offs: (1) dynamically decide when to perform recycled or full attention step based on the query similarities and (2) continued pre-training the model with Recycled Attention. Large language models (LLMs) are trained to ingest extremely long inputs and generate long outputs (Meta, 2024; Gemini, 2024) to support a wide range of applications. However, deploying such long-context LLMs can be very costly. As the context length increases, LLMs see a linear increase in memory to store the Key-Value (KV) cache and a quadratic increase in time for attention computation.

Auto-regressive inference of transformers benefit greatly from Key-Value (KV) caching, but can lead to major memory bottlenecks as model size, batch size, and sequence length grow at scale. We introduce Multi-Layer Key-Value (MLKV) sharing, a novel approach extending KV sharing across transformer layers to reduce memory usage beyond what was possible with Multi-Query Attention (MQA) and Grouped-Query Attention (GQA). Evaluations on various NLP benchmarks and inference metrics using uptrained Pythia-160M variants demonstrate that MLKV significantly reduces memory usage with minimal performance loss, reducing KV cache size down Figure 1: Simplified overview of current KV sharing to a factor of 6x compared to MQA. These methods, vanilla MHA (top left), MQA (bottom left), results highlight MLKV's potential for efficient and GQA (top right). All of them share KV heads deployment of transformer models at within the same layer. Our proposed KV sharing scheme scale. We provide code at https://github. MLKV (bottom right) shares KV heads between layers.

Excessive memory requirements of key and value features (KV-cache) present significant challenges in the autoregressive inference of large language models (LLMs), restricting both the speed and length of text generation. Approaches such as Multi-Query Attention (MQA) and Grouped Query Attention (GQA) mitigate these challenges by grouping query heads and consequently reducing the number of corresponding key and value heads. However, MQA and GQA decrease the KV-cache size requirements at the expense of LLM accuracy (quality of text generation). These methods do not ensure an optimal tradeoff between KV-cache size and text generation quality due to the absence of quality-aware grouping of query heads. To address this issue, we propose Quality and Capacity-Aware Grouped Query Attention (QCQA), which identifies optimal query head groupings using an evolutionary algorithm with a computationally efficient and inexpensive fitness function. We demonstrate that QCQA achieves a significantly better tradeoff between KV-cache capacity and LLM accuracy compared to GQA. For the Llama2 $7\,$B model, QCQA achieves $\mathbf{20}$\% higher accuracy than GQA with similar KV-cache size requirements in the absence of fine-tuning. After fine-tuning both QCQA and GQA, for a similar KV-cache size, QCQA provides $\mathbf{10.55}\,$\% higher accuracy than GQA. Furthermore, QCQA requires $40\,$\% less KV-cache size than GQA to attain similar accuracy. The proposed quality and capacity-aware grouping of query heads can serve as a new paradigm for KV-cache optimization in autoregressive LLM inference.

Multi-query attention (MQA), which only uses a single key-value head, drastically speeds up decoder inference. However, MQA can lead to quality degradation, and moreover it may not be desirable to train a separate model just for faster inference. We (1) propose a recipe for uptraining existing multi-head language model checkpoints into models with MQA using 5% of original pre-training compute, and (2) introduce grouped-query attention (GQA), a generalization of multi-query attention which uses an intermediate (more than one, less than number of query heads) number of key-value heads. We show that uptrained GQA achieves quality close to multi-head attention with comparable speed to MQA.