Daily Papers

Recent advances in long-text understanding have pushed the context length of large language models (LLMs) up to one million tokens. It boosts LLMs's accuracy and reasoning capacity but causes exorbitant computational costs and unsatisfactory Time to First Token (TTFT). KV cache reuse, which reuses the exact same KV cache of prefixes and templates or shares similar ones but with extra selective recomputation, offers a promising way to tackle this issue. However, prior studies overlook the cross-request KV reuse and the attention deviations introduced by new tokens during the decoding stage. In this paper, we present a KV cache management module that shares the KV cache across requests under multi-tenant scenarios without sacrificing model accuracy. Our system, KVShare, enables accurate and efficient LLM serving by 1) a Dual-Stage High Deviation algorithm (DHD) that conditionally selects a small portion of KV cache to be recomputed during both prefill and decode phases, and 2) a cache-aware scheduler that prioritizes requests based on their KV cache hit rates and orchestrates continuous batching to achieve enhanced system efficiency and faster TTFT. Multi-task experiments conducted on models such as Qwen2.5-7B,Llama3.1-8B and Yi1.5-9B demonstrate that KVShare reduces TTFT by up to 9.39x and increases 1.2x of the throughput compared to the full KV recompute. Moreover, KVShare achieves 20.38% boost in terms of accuracy compared to SOTA methods.

Long-context LLMs have enabled numerous downstream applications but also introduced significant challenges related to computational and memory efficiency. To address these challenges, optimizations for long-context inference have been developed, centered around the KV cache. However, existing benchmarks often evaluate in single-request, neglecting the full lifecycle of the KV cache in real-world use. This oversight is particularly critical, as KV cache reuse has become widely adopted in LLMs inference frameworks, such as vLLM and SGLang, as well as by LLM providers, including OpenAI, Microsoft, Google, and Anthropic. To address this gap, we introduce SCBench(SharedContextBench), a comprehensive benchmark for evaluating long-context methods from a KV cachecentric perspective: 1) KV cache generation, 2) KV cache compression, 3) KV cache retrieval, 4) KV cache loading. Specifically, SCBench uses test examples with shared context, ranging 12 tasks with two shared context modes, covering four categories of long-context capabilities: string retrieval, semantic retrieval, global information, and multi-task. With it, we provide an extensive KV cache-centric analysis of eight categories long-context solutions, including Gated Linear RNNs, Mamba-Attention hybrids, and efficient methods such as sparse attention, KV cache dropping, quantization, retrieval, loading, and prompt compression. The evaluation is conducted on 8 long-context LLMs. Our findings show that sub-O(n) memory methods suffer in multi-turn scenarios, while sparse encoding with O(n) memory and sub-O(n^2) pre-filling computation perform robustly. Dynamic sparsity yields more expressive KV caches than static patterns, and layer-level sparsity in hybrid architectures reduces memory usage with strong performance. Additionally, we identify attention distribution shift issues in long-generation scenarios. https://aka.ms/SCBench.

KV cache has traditionally been stored in GPU memory to accelerate the decoding phase of large language model (LLM) inference. However, it is increasingly necessary to move KV caches outside GPU devices, to enable cache reuse across different queries and inference engines. Our real-world usage statistics confirm this trend: over time, the total KV cache stored by users has grown rapidly, far exceeding the capacity of GPU memory. Despite this need, there lacks an efficient solution for offloading and transferring KV caches. We present LMCACHE, the first and so far the most efficient open-source KV caching solution, which extracts and stores KV caches generated by modern LLM engines (vLLM and SGLang) out of the GPU memory and shares them across engines and queries. LMCACHE supports both cache offloading (prefix reuse across queries) and prefill-decode (PD) disaggregation (cross-engine/GPU cache transfer). LMCACHE's high performance and wide adoption stem from the following contributions: (1) highly optimized KV cache data movement powered by batched data movement operations, compute and I/O pipelining; (2) a modular KV cache connector component, decoupling LMCACHE from the rapid evolution of inference engines; (3) a first-class control API for flexible cache orchestration across GPU, CPU, storage, and network layers. Our evaluation shows that combining LMCACHE with vLLM achieves up to 15x improvement in throughput across workloads such as multi-round question answering and document analysis. Large-scale adoption of LMCACHE in enterprise settings provides us valuable insights, for example, fetching KV cache from remote storage has unsurprisingly benefits to prefill delay, and that context truncation, which is a widely applied technique in industry, can greatly reduce prefix cache hit ratio by half. The source code of LMCACHE is at: https://github.com/LMCache/LMCache.

Diffusion-based large language models (Diffusion LLMs) have shown promise for non-autoregressive text generation with parallel decoding capabilities. However, the practical inference speed of open-sourced Diffusion LLMs often lags behind autoregressive models due to the lack of Key-Value (KV) Cache and quality degradation when decoding multiple tokens simultaneously. To bridge this gap, we introduce a novel block-wise approximate KV Cache mechanism tailored for bidirectional diffusion models, enabling cache reuse with negligible performance drop. Additionally, we identify the root cause of generation quality degradation in parallel decoding as the disruption of token dependencies under the conditional independence assumption. To address this, we propose a confidence-aware parallel decoding strategy that selectively decodes tokens exceeding a confidence threshold, mitigating dependency violations and maintaining generation quality. Experimental results on LLaDA and Dream models across multiple LLM benchmarks demonstrate up to 27.6times throughput improvement with minimal accuracy loss, closing the performance gap with autoregressive models and paving the way for practical deployment of Diffusion LLMs.

Autoregressive models (ARMs) are hindered by slow sequential inference. While masked diffusion models (MDMs) offer a parallel alternative, they suffer from critical drawbacks: high computational overhead from precluding Key-Value (KV) caching, and incoherent generation arising from learning dependencies over an intractable space of token combinations. To address these limitations, we introduce ReFusion, a novel masked diffusion model that achieves superior performance and efficiency by elevating parallel decoding from the token level to a higher slot level, where each slot is a fixed-length, contiguous sub-sequence. This is achieved through an iterative ``plan-and-infill'' decoding process: a diffusion-based planning step first identifies a set of weakly dependent slots, and an autoregressive infilling step then decodes these selected slots in parallel. The slot-based design simultaneously unlocks full KV cache reuse with a unified causal framework and reduces the learning complexity from the token combination space to a manageable slot-level permutation space. Extensive experiments on seven diverse benchmarks show that ReFusion not only overwhelmingly surpasses prior MDMs with 34% performance gains and an over 18times speedup on average, but also bridges the performance gap to strong ARMs while maintaining a 2.33times average speedup.

This paper presents VLCache, a cache reuse framework that exploits both Key-Value (KV) cache and encoder cache from prior multimodal inputs to eliminate costly recomputation when the same multimodal inputs recur. Unlike previous heuristic approaches, we formally identify the cumulative reuse error effect and demonstrate how to minimize the non-prefix cache reuse error effectively. We further analyze the varying importance of model layers and propose a dynamic, layer-aware recomputation strategy to balance accuracy and efficiency. Experimental results show that VLCache achieves an accuracy on par with full recomputation, while requiring only 2-5% of the tokens to compute, yielding 1.2x-16x TTFT speedups. The proposed VLCache pipeline has been integrated into SGLang, enabling significantly faster inference in practical deployments.

Vision-and-Language Navigation (VLN) in real-world settings requires agents to process continuous visual streams and generate actions with low latency grounded in language instructions. While Video-based Large Language Models (Video-LLMs) have driven recent progress, current VLN methods based on Video-LLM often face trade-offs among fine-grained visual understanding, long-term context modeling and computational efficiency. We introduce StreamVLN, a streaming VLN framework that employs a hybrid slow-fast context modeling strategy to support multi-modal reasoning over interleaved vision, language and action inputs. The fast-streaming dialogue context facilitates responsive action generation through a sliding-window of active dialogues, while the slow-updating memory context compresses historical visual states using a 3D-aware token pruning strategy. With this slow-fast design, StreamVLN achieves coherent multi-turn dialogue through efficient KV cache reuse, supporting long video streams with bounded context size and inference cost. Experiments on VLN-CE benchmarks demonstrate state-of-the-art performance with stable low latency, ensuring robustness and efficiency in real-world deployment. The project page is: https://streamvln.github.io/{https://streamvln.github.io/}.

Video large language models (LLMs) achieve strong video understanding by leveraging a large number of spatio-temporal tokens, but suffer from quadratic computational scaling with token count. To address this, we propose a training-free spatio-temporal token merging method, named STTM. Our key insight is to exploit local spatial and temporal redundancy in video data which has been overlooked in prior work. STTM first transforms each frame into multi-granular spatial tokens using a coarse-to-fine search over a quadtree structure, then performs directed pairwise merging across the temporal dimension. This decomposed merging approach outperforms existing token reduction methods across six video QA benchmarks. Notably, STTM achieves a 2times speed-up with only a 0.5% accuracy drop under a 50% token budget, and a 3times speed-up with just a 2% drop under a 30% budget. Moreover, STTM is query-agnostic, allowing KV cache reuse across different questions for the same video. The project page is available at https://www.jshyun.me/projects/sttm.

In recent multimodal research, the diffusion paradigm has emerged as a promising alternative to the autoregressive paradigm (AR), owing to its unique decoding advantages. However, due to the capability limitations of the base diffusion language model, the performance of the diffusion vision language model (dVLM) still lags significantly behind that of mainstream models. This leads to a simple yet fundamental question: Is it possible to construct dVLMs based on existing powerful AR models? In response, we propose DiffusionVL, a dVLM family that could be translated from any powerful AR models. Through simple fine-tuning, we successfully adapt AR pre-trained models into the diffusion paradigm. This approach yields two key observations: (1) The paradigm shift from AR-based multimodal models to diffusion is remarkably effective. (2) Direct conversion of an AR language model to a dVLM is also feasible, achieving performance competitive with LLaVA-style visual-instruction-tuning. Further, we introduce a block-decoding design into dVLMs that supports arbitrary-length generation and KV cache reuse, achieving a significant inference speedup. We conduct a large number of experiments. Despite training with less than 5% of the data required by prior methods, DiffusionVL achieves a comprehensive performance improvement-a 34.4% gain on the MMMU-Pro (vision) bench and 37.5% gain on the MME (Cog.) bench-alongside a 2x inference speedup. The model and code are released at https://github.com/hustvl/DiffusionVL.

Large language models (LLMs) are increasingly used for complex tasks requiring multiple chained generation calls, advanced prompting techniques, control flow, and interaction with external environments. However, efficient systems for programming and executing these applications are lacking. To bridge this gap, we introduce SGLang, a Structured Generation Language for LLMs. SGLang is designed for the efficient programming of LLMs and incorporates primitives for common LLM programming patterns. We have implemented SGLang as a domain-specific language embedded in Python, and we developed an interpreter, a compiler, and a high-performance runtime for SGLang. These components work together to enable optimizations such as parallelism, batching, caching, sharing, and other compilation techniques. Additionally, we propose RadixAttention, a novel technique that maintains a Least Recently Used (LRU) cache of the Key-Value (KV) cache for all requests in a radix tree, enabling automatic KV cache reuse across multiple generation calls at runtime. SGLang simplifies the writing of LLM programs and boosts execution efficiency. Our experiments demonstrate that SGLang can speed up common LLM tasks by up to 5x, while reducing code complexity and enhancing control.

Multi-token generation has emerged as a promising paradigm for accelerating transformer-based large model inference. Recent efforts primarily explore diffusion Large Language Models (dLLMs) for parallel decoding to reduce inference latency. To achieve AR-level generation quality, many techniques adapt AR models into dLLMs to enable parallel decoding. However, they suffer from limited speedup compared to AR models due to a pretrain-to-posttrain mismatch. Specifically, the masked data distribution in post-training deviates significantly from the real-world data distribution seen during pretraining, and dLLMs rely on bidirectional attention, which conflicts with the causal prior learned during pretraining and hinders the integration of exact KV cache reuse. To address this, we introduce Jacobi Forcing, a progressive distillation paradigm where models are trained on their own generated parallel decoding trajectories, smoothly shifting AR models into efficient parallel decoders while preserving their pretrained causal inference property. The models trained under this paradigm, Jacobi Forcing Model, achieves 3.8x wall-clock speedup on coding and math benchmarks with minimal loss in performance. Based on Jacobi Forcing Models' trajectory characteristics, we introduce multi-block decoding with rejection recycling, which enables up to 4.5x higher token acceptance count per iteration and nearly 4.0x wall-clock speedup, effectively trading additional compute for lower inference latency. Our code is available at https://github.com/hao-ai-lab/JacobiForcing.

We present Datarus-R1-14B, a 14 B-parameter open-weights language model fine-tuned from Qwen 2.5-14B-Instruct to act as a virtual data analyst and graduate-level problem solver. Datarus is trained not on isolated question-answer pairs but on full analytical trajectories including reasoning steps, code execution, error traces, self-corrections, and final conclusions, all captured in a ReAct-style notebook format spanning finance, medicine, numerical analysis, and other quantitative domains. Our training pipeline combines (i) a trajectory-centric synthetic data generator that yielded 144 000 tagged notebook episodes, (ii) a dual-reward framework blending a lightweight tag-based structural signal with a Hierarchical Reward Model (HRM) that scores both single-step soundness and end-to-end coherence, and (iii) a memory-optimized implementation of Group Relative Policy Optimization (GRPO) featuring KV-cache reuse, sequential generation, and reference-model sharding. A cosine curriculum smoothly shifts emphasis from structural fidelity to semantic depth, reducing the format collapse and verbosity that often plague RL-aligned LLMs. A central design choice in Datarus is it dual reasoning interface. In agentic mode the model produces ReAct-tagged steps that invoke Python tools to execute real code; in reflection mode it outputs compact Chain-of-Thought (CoT) traces delimited by <think> and <answer> tags. On demanding postgraduate-level problems, Datarus exhibits an "AHA-moment" pattern: it sketches hypotheses, revises them once or twice, and converges avoiding the circular, token-inflating loops common to contemporary systems. Across standard public benchmarks Datarus surpasses similar size models and even reaches the level of larger reasoning models such as QwQ-32B achieving up to 30% higher accuracy on AIME 2024/2025 and LiveCodeBench while emitting 18-49% fewer tokens per solution.

Multi-agent large language model (LLM) systems are increasingly adopted for complex language processing tasks that require communication and coordination among agents. However, these systems often suffer substantial overhead from repeated reprocessing of overlapping contexts across agents. In typical pipelines, once an agent receives a message from its predecessor, the full context-including prior turns-must be reprocessed from scratch, leading to inefficient processing. While key-value (KV) caching is an effective solution for avoiding redundant computation in single-agent settings where prefixes remain unchanged, it cannot be directly reused in multi-agent scenarios due to diverging prefixes introduced by agent-specific context extensions. We identify that the core challenge lies in the offset variance of KV-caches across agents. To address this, we propose KVCOMM, a training-free framework that enables efficient prefilling in multi-agent inference by reusing KV-caches and aligning cache offsets of overlapping contexts under diverse prefix contexts. KVCOMM estimates and adjusts KV-caches for shared content by referencing a pool of cached examples-termed anchors-that store observed cache deviations under varying prefixes. The anchor pool is maintained and updated online, allowing dynamic adaptation to distinct user requests and context structures. KVCOMM achieves over 70% reuse rate across diverse multi-agent workloads, including retrieval-augmented generation, math reasoning, and collaborative coding tasks, all without quality degradation. Particularly, when each fully-connected agent receives 1K input tokens with 512 prefix tokens and 512 output tokens under a five-agent setting, KVCOMM achieves up to 7.8x speedup compared to the standard prefill pipeline, reducing TTFT from ~430 ms to ~55 ms.

Transformer-based large language models (LLMs) cache context as key-value (KV) pairs during inference. As context length grows, KV cache sizes expand, leading to substantial memory overhead and increased attention latency. This paper introduces KVzip, a query-agnostic KV cache eviction method enabling effective reuse of compressed KV caches across diverse queries. KVzip quantifies the importance of a KV pair using the underlying LLM to reconstruct original contexts from cached KV pairs, subsequently evicting pairs with lower importance. Extensive empirical evaluations demonstrate that KVzip reduces KV cache size by 3-4times and FlashAttention decoding latency by approximately 2times, with negligible performance loss in question-answering, retrieval, reasoning, and code comprehension tasks. Evaluations include various models such as LLaMA3.1-8B, Qwen2.5-14B, and Gemma3-12B, with context lengths reaching up to 170K tokens. KVzip significantly outperforms existing query-aware KV eviction methods, which suffer from performance degradation even at a 90% cache budget ratio under multi-query scenarios.

To reduce memory costs in long-context inference with Large Language Models (LLMs), many recent works focus on compressing the key-value (KV) cache of different tokens. However, we identify that the previous KV cache compression methods measure token importance individually, neglecting the dependency between different tokens in the real-world language characterics. In light of this, we introduce ChunkKV, grouping the tokens in a chunk as a basic compressing unit, and retaining the most informative semantic chunks while discarding the less important ones. Furthermore, observing that ChunkKV exhibits higher similarity in the preserved indices across different layers, we propose layer-wise index reuse to further reduce computational overhead. We evaluated ChunkKV on cutting-edge long-context benchmarks including LongBench and Needle-In-A-HayStack, as well as the GSM8K and JailbreakV in-context learning benchmark. Our experiments with instruction tuning and multi-step reasoning (O1 and R1) LLMs, achieve up to 10\% performance improvement under aggressive compression ratios compared to existing methods.

Large Language Models (LLMs) have revolutionized a wide range of domains such as natural language processing, computer vision, and multi-modal tasks due to their ability to comprehend context and perform logical reasoning. However, the computational and memory demands of LLMs, particularly during inference, pose significant challenges when scaling them to real-world, long-context, and real-time applications. Key-Value (KV) cache management has emerged as a critical optimization technique for accelerating LLM inference by reducing redundant computations and improving memory utilization. This survey provides a comprehensive overview of KV cache management strategies for LLM acceleration, categorizing them into token-level, model-level, and system-level optimizations. Token-level strategies include KV cache selection, budget allocation, merging, quantization, and low-rank decomposition, while model-level optimizations focus on architectural innovations and attention mechanisms to enhance KV reuse. System-level approaches address memory management, scheduling, and hardware-aware designs to improve efficiency across diverse computing environments. Additionally, the survey provides an overview of both text and multimodal datasets and benchmarks used to evaluate these strategies. By presenting detailed taxonomies and comparative analyses, this work aims to offer useful insights for researchers and practitioners to support the development of efficient and scalable KV cache management techniques, contributing to the practical deployment of LLMs in real-world applications. The curated paper list for KV cache management is in: https://github.com/TreeAI-Lab/Awesome-KV-Cache-Management{https://github.com/TreeAI-Lab/Awesome-KV-Cache-Management}.

Vision-Language-Action (VLA) model can process instructions and visual perception to directly generate actions as output in an end-to-end fashion due to its strong multi-modal reasoning capabilities. While the performance of VLA models is promising, their computational cost can be substantial. This raises challenge for applying them on robotics tasks, which requires real-time decision-making to respond quickly to environmental changes. Since robotic control involves sequential decision-making, the visual input often exhibits minimal variation between successive steps. A natural idea is to reuse the computational results of unchanged visual tokens from the last step. Motivated by this idea, we propose VLA-Cache, an efficient vision-language-action model. VLA-Cache incorporates a token-selection mechanism that compares the visual input at each step with the input from the previous step, adaptively identifying visual tokens with minimal changes. The computational results for these unchanged tokens are then reused in subsequent steps via KV-cache, thereby significantly improving the efficiency of the VLA-Cache model. Experimental results on both simulation (e.g., LIBERO benchmark and SIMPLER) and real-world robot valid VLA-Cache can achieve practical acceleration with minimal sacrifice in success rate.

We introduce AIBrix, a cloud-native, open-source framework designed to optimize and simplify large-scale LLM deployment in cloud environments. Unlike traditional cloud-native stacks, AIBrix follows a co-design philosophy, ensuring every layer of the infrastructure is purpose-built for seamless integration with inference engines like vLLM. AIBrix introduces several key innovations to reduce inference costs and enhance performance including high-density LoRA management for dynamic adapter scheduling, LLM-specific autoscalers, and prefix-aware, load-aware routing. To further improve efficiency, AIBrix incorporates a distributed KV cache, boosting token reuse across nodes, leading to a 50% increase in throughput and a 70% reduction in inference latency. AIBrix also supports unified AI runtime which streamlines model management while maintaining vendor-agnostic engine compatibility. For large-scale multi-node inference, AIBrix employs hybrid orchestration -- leveraging Kubernetes for coarse-grained scheduling and Ray for fine-grained execution -- to balance efficiency and flexibility. Additionally, an SLO-driven GPU optimizer dynamically adjusts resource allocations, optimizing heterogeneous serving to maximize cost efficiency while maintaining service guarantees. Finally, AIBrix enhances system reliability with AI accelerator diagnostic tools, enabling automated failure detection and mock-up testing to improve fault resilience. AIBrix is available at https://github.com/vllm-project/aibrix.

Large language models (LLMs) often incorporate multiple text chunks in their inputs to provide the necessary contexts. To speed up the prefill of the long LLM inputs, one can pre-compute the KV cache of a text and re-use the KV cache when the context is reused as the prefix of another LLM input. However, the reused text chunks are not always the input prefix, and when they are not, their precomputed KV caches cannot be directly used since they ignore the text's cross-attention with the preceding text in the LLM input. Thus, the benefits of reusing KV caches remain largely unrealized. This paper tackles just one question: when an LLM input contains multiple text chunks, how to quickly combine their precomputed KV caches in order to achieve the same generation quality as the expensive full prefill (i.e., without reusing KV cache)? We present CacheBlend, a scheme that reuses the pre-computed KV caches, regardless prefix or not, and selectively recomputes the KV values of a small subset of tokens to partially update each reused KV cache. In the meantime,the small extra delay for recomputing some tokens can be pipelined with the retrieval of KV caches within the same job,allowing CacheBlend to store KV caches in slower devices with more storage capacity while retrieving them without increasing the inference delay. By comparing CacheBlend with the state-of-the-art KV cache reusing schemes on three open-source LLMs of various sizes and four popular benchmark datasets of different tasks, we show that CacheBlend reduces time-to-first-token (TTFT) by 2.2-3.3X and increases the inference throughput by 2.8-5X, compared with full KV recompute, without compromising generation quality or incurring more storage cost.

In multi-agent systems utilizing Large Language Models (LLMs), communication between agents traditionally relies on natural language. This communication often includes the full context of the query so far, which can introduce significant prefill-phase latency, especially with long contexts. We introduce DroidSpeak, a novel framework to target this cross-LLM communication by leveraging the reuse of intermediate data, such as input embeddings (E-cache) and key-value caches (KV-cache). We efficiently bypass the need to reprocess entire contexts for fine-tuned versions of the same foundational model. This approach allows faster context integration while maintaining the quality of task performance. Experimental evaluations demonstrate DroidSpeak's ability to significantly accelerate inter-agent communication, achieving up to a 2.78x speedup in prefill latency with negligible loss in accuracy. Our findings underscore the potential to create more efficient and scalable multi-agent systems.

Many LLM tasks are performed in large batches or even offline, and the performance indictor for which is throughput. These tasks usually show the characteristic of prefix sharing, where different prompt input can partially show the common prefix. However, the existing LLM inference engines tend to optimize the streaming requests and show limitations of supporting the large batched tasks with the prefix sharing characteristic. The existing solutions use the LRU-based cache to reuse the KV context of common prefix. The KV context that is about to be reused may prematurely be evicted with the implicit cache management. Even if not evicted, the lifetime of the shared KV context is extended since requests sharing the same context are not scheduled together, resulting in larger memory usage. These streaming oriented systems schedule the requests in the first-come-first-serve or similar order. As a result, the requests with larger ratio of decoding steps may be scheduled too late to be able to mix with the prefill chunks to increase the hardware utilization. Besides, the token and request number based batching can limit the size of token-batch, which keeps the GPU from saturating for the iterations dominated by decoding tokens. We propose BatchLLM to address the above problems. BatchLLM explicitly identifies the common prefixes globally. The requests sharing the same prefix will be scheduled together to reuse the KV context the best, which also shrinks the lifetime of common KV memory. BatchLLM reorders the requests and schedules the requests with larger ratio of decoding first to better mix the decoding tokens with the latter prefill chunks and applies memory-centric token batching to enlarge the token-batch sizes, which helps to increase the GPU utilization. Extensive evaluation shows that BatchLLM outperforms vLLM by 1.1x to 2x on a set of microbenchmarks and two typical industry workloads.

We introduce a decoder-decoder architecture, YOCO, for large language models, which only caches key-value pairs once. It consists of two components, i.e., a cross-decoder stacked upon a self-decoder. The self-decoder efficiently encodes global key-value (KV) caches that are reused by the cross-decoder via cross-attention. The overall model behaves like a decoder-only Transformer, although YOCO only caches once. The design substantially reduces GPU memory demands, yet retains global attention capability. Additionally, the computation flow enables prefilling to early exit without changing the final output, thereby significantly speeding up the prefill stage. Experimental results demonstrate that YOCO achieves favorable performance compared to Transformer in various settings of scaling up model size and number of training tokens. We also extend YOCO to 1M context length with near-perfect needle retrieval accuracy. The profiling results show that YOCO improves inference memory, prefill latency, and throughput by orders of magnitude across context lengths and model sizes. Code is available at https://aka.ms/YOCO.

Retrieval-Augmented Generation (RAG) is often used with Large Language Models (LLMs) to infuse domain knowledge or user-specific information. In RAG, given a user query, a retriever extracts chunks of relevant text from a knowledge base. These chunks are sent to an LLM as part of the input prompt. Typically, any given chunk is repeatedly retrieved across user questions. However, currently, for every question, attention-layers in LLMs fully compute the key values (KVs) repeatedly for the input chunks, as state-of-the-art methods cannot reuse KV-caches when chunks appear at arbitrary locations with arbitrary contexts. Naive reuse leads to output quality degradation. This leads to potentially redundant computations on expensive GPUs and increases latency. In this work, we propose Cache-Craft, a system for managing and reusing precomputed KVs corresponding to the text chunks (we call chunk-caches) in RAG-based systems. We present how to identify chunk-caches that are reusable, how to efficiently perform a small fraction of recomputation to fix the cache to maintain output quality, and how to efficiently store and evict chunk-caches in the hardware for maximizing reuse while masking any overheads. With real production workloads as well as synthetic datasets, we show that Cache-Craft reduces redundant computation by 51% over SOTA prefix-caching and 75% over full recomputation. Additionally, with continuous batching on a real production workload, we get a 1.6X speed up in throughput and a 2X reduction in end-to-end response latency over prefix-caching while maintaining quality, for both the LLaMA-3-8B and LLaMA-3-70B models.

Vision-language models (VLMs) could power real-time assistants and autonomous agents, but they face a critical challenge: understanding near-infinite video streams without escalating latency and memory usage. Processing entire videos with full attention leads to quadratic computational costs and poor performance on long videos. Meanwhile, simple sliding window methods are also flawed, as they either break coherence or suffer from high latency due to redundant recomputation. In this paper, we introduce StreamingVLM, a model designed for real-time, stable understanding of infinite visual input. Our approach is a unified framework that aligns training with streaming inference. During inference, we maintain a compact KV cache by reusing states of attention sinks, a short window of recent vision tokens, and a long window of recent text tokens. This streaming ability is instilled via a simple supervised fine-tuning (SFT) strategy that applies full attention on short, overlapped video chunks, which effectively mimics the inference-time attention pattern without training on prohibitively long contexts. For evaluation, we build Inf-Streams-Eval, a new benchmark with videos averaging over two hours that requires dense, per-second alignment between frames and text. On Inf-Streams-Eval, StreamingVLM achieves a 66.18% win rate against GPT-4O mini and maintains stable, real-time performance at up to 8 FPS on a single NVIDIA H100. Notably, our SFT strategy also enhances general VQA abilities without any VQA-specific fine-tuning, improving performance on LongVideoBench by +4.30 and OVOBench Realtime by +5.96. Code is available at https://github.com/mit-han-lab/streaming-vlm.

We introduce GoldFinch, a hybrid Linear Attention/Transformer sequence model that uses a new technique to efficiently generate a highly compressed and reusable KV-Cache in linear time and space with respect to sequence length. GoldFinch stacks our new GOLD transformer on top of an enhanced version of the Finch (RWKV-6) architecture. We train up to 1.5B parameter class models of the Finch, Llama, and GoldFinch architectures, and find dramatically improved modeling performance relative to both Finch and Llama. Our cache size savings increase linearly with model layer count, ranging from 756-2550 times smaller than the traditional transformer cache for common sizes, enabling inference of extremely large context lengths even on limited hardware. Although autoregressive generation has O(n) time complexity per token because of attention, pre-fill computation of the entire initial cache state for a submitted context costs only O(1) time per token due to the use of a recurrent neural network (RNN) to generate this cache. We release our trained weights and training code under the Apache 2.0 license for community use.

Large language models (LLMs) have achieved remarkable success on various aspects of human life. However, one of the major challenges in deploying these models is the substantial memory consumption required to store key-value pairs (KV), which imposes significant resource demands. Recent research has focused on KV cache budget allocation, with several approaches proposing head-level budget distribution by evaluating the importance of individual attention heads. These methods, however, assess the importance of heads independently, overlooking their cooperative contributions within the model, which may result in a deviation from their true impact on model performance. In light of this limitation, we propose CoKV, a novel method that models the cooperation between heads in model inference as a cooperative game. By evaluating the contribution of each head within the cooperative game, CoKV can allocate the cache budget more effectively. Extensive experiments show that CoKV achieves state-of-the-art performance on the LongBench benchmark using LLama-3-8B-Instruct and Mistral-7B models.

Recent advances in large language models (LLMs) have extended context lengths, enabling assistants to sustain long histories for coherent, personalized responses. This ability, however, hinges on Key-Value (KV) caching, whose memory grows linearly with dialogue length and quickly dominates under strict resource constraints. An active line of research for reducing this overhead is KV cache compression, which seeks to limit cache size while preserving accuracy. Yet existing methods face two major limitations: (i) evicting entries after full-context prefill causes unbounded peak memory, and (ii) query-dependent eviction narrows the cache to a single query, leading to degraded accuracy in multi-turn conversations. We introduce EpiCache, a training-free KV cache management framework for long conversational question answering (LongConvQA) under fixed memory budgets. EpiCache bounds cache growth through block-wise prefill and preserves topic-relevant context via episodic KV compression, which clusters conversation history into coherent episodes and applies episode-specific KV cache eviction. We further design an adaptive layer-wise budget allocation strategy that measures each layer's sensitivity to eviction and distributes the memory budget across layers accordingly. Across three LongConvQA benchmarks, EpiCache improves accuracy by up to 40% over recent baselines, sustains near-full KV accuracy under 4-6x compression, and reduces latency and memory by up to 2.4x and 3.5x, thereby enabling efficient multi-turn interaction under strict resource constraints.

Large Language models (LLMs) have become a research hotspot. To accelerate the inference of LLMs, storing computed caches in memory has become the standard technique. However, as the inference length increases, growing KV caches might lead to out-of-memory issues. Many existing methods address this issue through KV cache compression, primarily by preserving key tokens throughout all layers to reduce information loss. Most of them allocate a uniform budget size for each layer to retain. However, we observe that the minimum budget sizes needed to retain essential information vary across layers and models based on the perspectives of attention and hidden state output. Building on this observation, this paper proposes a simple yet effective KV cache compression method that leverages layer uncertainty to allocate budget size for each layer. Experimental results show that the proposed method can reduce memory usage of the KV caches to only sim20\% when compared to Full KV inference while achieving nearly lossless performance.

KV cache is a widely used acceleration technique for large language models (LLMs) inference. However, its memory requirement grows rapidly with input length. Previous studies have reduced the size of KV cache by either removing the same number of unimportant tokens for all attention heads or by allocating differentiated KV cache budgets for pre-identified attention heads. However, due to the importance of attention heads varies across different tasks, the pre-identified attention heads fail to adapt effectively to various downstream tasks. To address this issue, we propose Task-KV, a method that leverages the semantic differentiation of attention heads to allocate differentiated KV cache budgets across various tasks. We demonstrate that attention heads far from the semantic center (called heterogeneous heads) make an significant contribution to task outputs and semantic understanding. In contrast, other attention heads play the role of aggregating important information and focusing reasoning. Task-KV allocates full KV cache budget to heterogeneous heads to preserve comprehensive semantic information, while reserving a small number of recent tokens and attention sinks for non-heterogeneous heads. Furthermore, we innovatively introduce middle activations to preserve key contextual information aggregated from non-heterogeneous heads. To dynamically perceive semantic differences among attention heads, we design a semantic separator to distinguish heterogeneous heads from non-heterogeneous ones based on their distances from the semantic center. Experimental results on multiple benchmarks and different model architectures demonstrate that Task-KV significantly outperforms existing baseline methods.

Memory consumption of the Key-Value (KV) cache represents a major bottleneck for efficient large language model inference. While attention-score-based KV cache pruning shows promise, it faces critical practical limitations: attention scores from future tokens are unavailable during compression, and modern implementations like Flash Attention do not materialize the full attention matrix, making past scores inaccessible. To overcome these challenges, we introduce Expected Attention, a training-free compression method that estimates KV pairs importance by predicting how future queries will attend to them. Our approach leverages the distributional properties of LLM activations to compute expected attention scores in closed form for each KV pair. These scores enable principled ranking and pruning of KV pairs with minimal impact on the residual stream, achieving effective compression without performance degradation. Importantly, our method operates seamlessly across both prefilling and decoding phases, consistently outperforming state-of-the-art baselines in both scenarios. Finally, we release KVPress, a comprehensive library to enable researchers to implement and benchmark KV cache compression methods, already including more than 20 techniques.

This work studies how to adaptively recompute key-value (KV) caches for diffusion large language models (DLMs) to maximize prediction accuracy while minimizing decoding latency. Prior methods' decoders recompute QKV for all tokens at every denoising step and layer, despite KV states changing little across most steps, especially in shallow layers, leading to substantial redundancy. We make three observations: (1) distant {bf MASK} tokens primarily act as a length-bias and can be cached block-wise beyond the active prediction window; (2) KV dynamics increase with depth, suggesting that selective refresh starting from deeper layers is sufficient; and (3) the most-attended token exhibits the smallest KV drift, providing a conservative lower bound on cache change for other tokens. Building on these, we propose {bf Elastic-Cache}, a training-free, architecture-agnostic strategy that jointly decides {when} to refresh (via an attention-aware drift test on the most-attended token) and {where} to refresh (via a depth-aware schedule that recomputes from a chosen layer onward while reusing shallow-layer caches and off-window MASK caches). Unlike fixed-period schemes, Elastic-Cache performs adaptive, layer-aware cache updates for diffusion LLMs, reducing redundant computation and accelerating decoding with negligible loss in generation quality. Experiments on LLaDA-Instruct, LLaDA-1.5, and LLaDA-V across mathematical reasoning and code generation tasks demonstrate consistent speedups: 8.7times on GSM8K (256 tokens), 45.1times on longer sequences, and 4.8times on HumanEval, while consistently maintaining higher accuracy than the baseline. Our method achieves significantly higher throughput (6.8times on GSM8K) than existing confidence-based approaches while preserving generation quality, enabling practical deployment of diffusion LLMs.

Recent developments in Video Large Language Models (Video LLMs) have enabled models to process long video sequences and demonstrate remarkable performance. Nonetheless, studies predominantly focus on offline video question answering, neglecting memory usage and response speed that are essential in various real-world applications, such as Deepseek services, autonomous driving, and robotics. To mitigate these challenges, we propose LiveVLM, a training-free framework specifically designed for streaming, online video understanding and real-time interaction. Unlike existing works that process videos only after one question is posed, LiveVLM constructs an innovative streaming-oriented KV cache to process video streams in real-time, retain long-term video details and eliminate redundant KVs, ensuring prompt responses to user queries. For continuous video streams, LiveVLM generates and compresses video key-value tensors (video KVs) to reserve visual information while improving memory efficiency. Furthermore, when a new question is proposed, LiveVLM incorporates an online question-answering process that efficiently fetches both short-term and long-term visual information, while minimizing interference from redundant context. Extensive experiments demonstrate that LiveVLM enables the foundation LLaVA-OneVision model to process 44times number of frames on the same device, and achieves up to 5times speedup in response speed compared with SoTA online methods at an input of 256 frames, while maintaining the same or better model performance.

Large Language Model (LLM) inference is typically memory-intensive, especially when processing large batch sizes and long sequences, due to the large size of key-value (KV) cache. Vector Quantization (VQ) is recently adopted to alleviate this issue, but we find that the existing approach is susceptible to distribution shift due to its reliance on calibration datasets. To address this limitation, we introduce NSNQuant, a calibration-free Vector Quantization (VQ) technique designed for low-bit compression of the KV cache. By applying a three-step transformation-1) a token-wise normalization (Normalize), 2) a channel-wise centering (Shift), and 3) a second token-wise normalization (Normalize)-with Hadamard transform, NSNQuant effectively aligns the token distribution with the standard normal distribution. This alignment enables robust, calibration-free vector quantization using a single reusable codebook. Extensive experiments show that NSNQuant consistently outperforms prior methods in both 1-bit and 2-bit settings, offering strong generalization and up to 3times throughput gain over full-precision baselines.

Recently, large language models (LLM) based on transformers are facing memory bottleneck issues due to KV cache, especially in long sequence handling. Previous researches proposed KV cache compression techniques that identify insignificant tokens based on Accumulative Attention Scores and removes their items from KV cache, noting that only few tokens play an important role in attention operations. However, we have observed that the existing Accumulative Attention Score is not suitable for the transformer decoder structure. In the decoder model, the number of times the Attention Score accumulates varies depending on the order of token appearance due to the effect of masking, causing an uneven comparison between tokens. To solve this, we propose Accumulative Attention Score with Forgetting Factor (A2SF) technique, which introduces a Forgetting Factor in the Attention Score accumulation process. A2SF applies a penalty to the past Attention Score generated from old tokens by repeatedly multiplying the Forgetting Factor to the Attention Score over time. Therefore, older tokens receive a larger penalty, providing fairness among different ages of tokens. Through the fair comparison among tokens, we can more effectively select important tokens. We have verified the accuracy improvement through A2SF in the OPT and LLaMA models and A2SF improves the accuracy of LLaMA 2 by up to 7.8% and 5.1% on 1-shot and 0-shot.

Efficient use of GPU memory is essential for high throughput LLM inference. Prior systems reserved memory for the KV-cache ahead-of-time, resulting in wasted capacity due to internal fragmentation. Inspired by OS-based virtual memory systems, vLLM proposed PagedAttention to enable dynamic memory allocation for KV-cache. This approach eliminates fragmentation, enabling high-throughput LLM serving with larger batch sizes. 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. This change requires attention kernels to be rewritten to support paging, and serving framework to implement a memory manager. Thus, the PagedAttention model leads to software complexity, portability issues, redundancy and inefficiency. In this paper, we propose vAttention for dynamic KV-cache memory management. In contrast to PagedAttention, vAttention retains KV-cache in contiguous virtual memory and leverages low-level system support for demand paging, that already exists, to enable on-demand physical memory allocation. Thus, vAttention unburdens the attention kernel developer from having to explicitly support paging and avoids re-implementation of memory management in the serving framework. We show that vAttention enables seamless dynamic memory management for unchanged implementations of various attention kernels. vAttention also generates tokens up to 1.97x faster than vLLM, while processing input prompts up to 3.92x and 1.45x faster than the PagedAttention variants of FlashAttention and FlashInfer.

We present CIFLEX (Contextual Instruction Flow for Sub-task Execution), which is a novel execution system for efficient sub-task handling in multi-turn interactions with a single on-device large language model (LLM). As LLMs become increasingly capable, a single model is expected to handle diverse sub-tasks that more effectively and comprehensively support answering user requests. Naive approach reprocesses the entire conversation context when switching between main and sub-tasks (e.g., query rewriting, summarization), incurring significant computational overhead. CIFLEX mitigates this overhead by reusing the key-value (KV) cache from the main task and injecting only task-specific instructions into isolated side paths. After sub-task execution, the model rolls back to the main path via cached context, thereby avoiding redundant prefill computation. To support sub-task selection, we also develop a hierarchical classification strategy tailored for small-scale models, decomposing multi-choice decisions into binary ones. Experiments show that CIFLEX significantly reduces computational costs without degrading task performance, enabling scalable and efficient multi-task dialogue on-device.

Mixture-of-Experts (MoE) models promise efficient scaling of large language models (LLMs) by activating only a small subset of experts per token, but their parallelized inference pipelines make elastic serving challenging. Existing strategies fall short: horizontal scaling provisions entire replicas of the current configuration, often tens to hundreds of accelerators, leading to coarse granularity, long provisioning delays, and costly overprovisioning. Vertical scaling offers finer adjustments but typically requires instance restarts, incurring downtime. These limitations make current approaches ill-suited for the bursty, short-lived traffic patterns common in cloud deployments. We present ElasticMoE, an elastic scaling framework for MoE LLMs that achieves fine-grained, low-latency, and zero-downtime scaling. ElasticMoE decouples inference execution from memory operations, enabling scaling steps to proceed concurrently with serving. An HBM Management Module (HMM) reuses weights and KV caches via zero-copy remapping, while high-bandwidth peer-to-peer transfers bring newly added accelerators online without interrupting service. A virtual memory based expert redistribution mechanism migrates MoE experts without costly buffer reallocations, reducing peak memory usage during expert parallelism reconfiguration. Our evaluation on Ascend NPUs with three popular MoE LLMs shows that ElasticMoE achieves up to 9x lower scale-up latency, up to 2x better throughput during scaling, and significantly improves SLO attainment compared to baselines. By enabling fine-grained, concurrent scaling with minimal disruption, ElasticMoE advances the practicality of deploying massive MoE LLMs in dynamic cloud environments.

Large Language Models (LLMs) rely on the Key-Value (KV) Cache to store token history, enabling efficient decoding of tokens. As the KV-Cache grows, it becomes a major memory and computation bottleneck, however, there is an opportunity to alleviate this bottleneck, especially because prior research has shown that only a small subset of tokens contribute meaningfully to each decoding step. A key challenge in finding these critical tokens is that they are dynamic, and heavily input query-dependent. Existing methods either risk quality by evicting tokens permanently, or retain the full KV-Cache but rely on retrieving chunks (pages) of tokens at generation, failing at dense, context-rich tasks. Additionally, many existing KV-Cache sparsity methods rely on inaccurate proxies for token importance. To address these limitations, we introduce TokenButler, a high-granularity, query-aware predictor that learns to identify these critical tokens. By training a light-weight predictor with less than 1.2% parameter overhead, TokenButler prioritizes tokens based on their contextual, predicted importance. This improves perplexity & downstream accuracy by over 8% relative to SoTA methods for estimating token importance. We evaluate TokenButler on a novel synthetic small-context co-referential retrieval task, demonstrating near-oracle accuracy. Code, models and benchmarks: https://github.com/abdelfattah-lab/TokenButler

Generating texts with a large language model (LLM) consumes massive amounts of memory. Apart from the already-large model parameters, the key/value (KV) cache that holds information about previous tokens in a sequence can grow to be even larger than the model itself. This problem is exacerbated in one of the current LLM serving frameworks which reserves the maximum sequence length of memory for the KV cache to guarantee generating a complete sequence as they do not know the output sequence length. This restricts us to use a smaller batch size leading to lower GPU utilization and above all, lower throughput. We argue that designing a system with a priori knowledge of the output sequence can mitigate this problem. To this end, we propose S^{3}, which predicts the output sequence length, schedules generation queries based on the prediction to increase device resource utilization and throughput, and handle mispredictions. Our proposed method achieves 6.49times throughput over those systems that assume the worst case for the output sequence length.

As large language models (LLMs) continue to scale, the memory footprint of key-value (KV) caches during inference has become a significant bottleneck. Existing approaches primarily focus on compressing KV caches within a single prompt or reusing shared prefixes or frequently ocurred text segments across prompts. However, such strategies are limited in scenarios where prompts are semantically similar but lexically different, which frequently occurs in tasks such as multi-document summarization and conversational agents. We propose SemShareKV, a KV cache sharing and compression framework that accelerates LLM inference by reusing KVCache in semantically similar prompts. Instead of relying on exact token matches, SemShareKV applies fuzzy token matching using locality-sensitive hashing (LSH) on token embeddings and incorporates Rotary Position Embedding (RoPE) to better preserve positional information. By selectively reusing relevant key-value pairs from a reference prompt's cache, SemShareKV reduces redundant computation while maintaining output quality. Experiments on diverse summarization datasets show up to 6.25times speedup and 42\% lower GPU memory usage with 5k tokens input, with negligible quality degradation. These results highlight the potential of semantic-aware cache sharing for efficient LLM inference.

Long-context language models exhibit impressive performance but remain challenging to deploy due to high GPU memory demands during inference. We propose Memory-efficient Offloaded Mini-sequence Inference (MOM), a method that partitions critical layers into smaller "mini-sequences" and integrates seamlessly with KV cache offloading. Experiments on various Llama, Qwen, and Mistral models demonstrate that MOM reduces peak memory usage by over 50\% on average. On Meta-Llama-3.2-8B, MOM extends the maximum context length from 155k to 455k tokens on a single A100 80GB GPU, while keeping outputs identical and not compromising accuracy. MOM also maintains highly competitive throughput due to minimal computational overhead and efficient last-layer processing. Compared to traditional chunked prefill methods, MOM achieves a 35\% greater context length extension. More importantly, our method drastically reduces prefill memory consumption, eliminating it as the longstanding dominant memory bottleneck during inference. This breakthrough fundamentally changes research priorities, redirecting future efforts from prefill-stage optimizations to improving decode-stage residual KV cache efficiency.

Motivated by reducing the computational and storage costs of LLMs, model compression and KV cache compression have attracted much attention from researchers. However, current methods predominantly emphasize maintaining the performance of compressed LLMs, as measured by perplexity or simple accuracy on tasks of common sense knowledge QA and basic arithmetic reasoning. In this blog, we present a brief review of recent advancements in LLMs related to retrieval-augmented generation, multi-step reasoning, external tools, and computational expressivity, all of which substantially enhance LLM performance. Then, we propose a lottery LLM hypothesis suggesting that for a given LLM and task, there exists a smaller lottery LLM capable of producing the same performance as the original LLM with the assistance of multi-step reasoning and external tools. Based on the review of current progress in LLMs, we discuss and summarize the essential capabilities that the lottery LLM and KV cache compression must possess, which are currently overlooked in existing methods.

Test-time scaling has emerged as a prominent research direction in machine learning, enabling models to enhance their expressive capabilities during inference.Transformers, renowned for striking a delicate balance between efficiency and expressiveness, have benefited from test-time scaling techniques that leverage an expanding key-value (KV) cache to significantly improve performance.In this paper, we introduce a novel state-based approach to test-time scaling, which we term state tuning, tailored to the RNN-based RWKV-7 model.By exploiting the unique strengths of RWKV-7, our method achieves state-of-the-art performance on the target task without altering the model's pre-trained weights. Our approach centers on three key innovations. First, we develop an observer framework that allows a smaller model to replicate and learn the state dynamics of the RWKV-7 model. Second, we employ a kernel method to dynamically upscale the state size, enhancing the model's capacity to capture intricate patterns. Third, we integrate Decorrelated Backpropagation (DBP) to optimize the upscaled state matrix, thereby improving convergence and expressivity. By tuning only the state matrix, we demonstrate that a smaller model can outperform larger models on the given task. This method preserves the efficiency of the original RWKV-7 architecture while harnessing the power of test-time scaling to deliver superior results. Our findings underscore the potential of state tuning as an effective strategy for advancing model performance in resource-constrained settings. Our code is https://github.com/TorchRWKV/flash-linear-attention.

Recent progress in large language models (LLMs) has focused on test-time scaling to improve reasoning via increased inference computation, but often at the cost of efficiency. We revisit test-time behavior and uncover a simple yet underexplored phenomenon: reasoning uncertainty is highly localized-only a small subset of high-entropy tokens dominantly affects output correctness. Motivated by this, we propose Minimal Test-Time Intervention (MTI), a training-free framework that enhances reasoning accuracy and stability with minimal overhead. MTI includes: (i) Selective CFG intervention, applying classifier-free guidance only at uncertain positions; and (ii) Lightweight negative-prompt guidance, reusing the main model's KV cache to approximate unconditional decoding efficiently. MTI yields consistent gains across general, coding, and STEM tasks-e.g., +1.35% average improvement on eight benchmarks for Qwen3-8B-Base and +5% on AIME2024 using Qwen3-32B-Reasoning-while remaining highly efficient.

Current Retrieval-Augmented Generation (RAG) systems concatenate and process numerous retrieved document chunks for prefill which requires a large volume of computation, therefore leading to significant latency in time-to-first-token (TTFT). To reduce the computation overhead as well as TTFT, we introduce TurboRAG, a novel RAG system that redesigns the inference paradigm of the current RAG system by first pre-computing and storing the key-value (KV) caches of documents offline, and then directly retrieving the saved KV cache for prefill. Hence, online computation of KV caches is eliminated during inference. In addition, we provide a number of insights into the mask matrix and positional embedding mechanisms, plus fine-tune a pretrained language model to maintain model accuracy of TurboRAG. Our approach is applicable to most existing large language models and their applications without any requirement in modification of models and inference systems. Experimental results across a suite of RAG benchmarks demonstrate that TurboRAG reduces TTFT by up to 9.4x compared to the conventional RAG systems (on an average of 8.6x), but reserving comparable performance to the standard RAG systems.

Despite the recent progress in long-context language models, it remains elusive how transformer-based models exhibit the capability to retrieve relevant information from arbitrary locations within the long context. This paper aims to address this question. Our systematic investigation across a wide spectrum of models reveals that a special type of attention heads are largely responsible for retrieving information, which we dub retrieval heads. We identify intriguing properties of retrieval heads:(1) universal: all the explored models with long-context capability have a set of retrieval heads; (2) sparse: only a small portion (less than 5\%) of the attention heads are retrieval. (3) intrinsic: retrieval heads already exist in models pretrained with short context. When extending the context length by continual pretraining, it is still the same set of heads that perform information retrieval. (4) dynamically activated: take Llama-2 7B for example, 12 retrieval heads always attend to the required information no matter how the context is changed. The rest of the retrieval heads are activated in different contexts. (5) causal: completely pruning retrieval heads leads to failure in retrieving relevant information and results in hallucination, while pruning random non-retrieval heads does not affect the model's retrieval ability. We further show that retrieval heads strongly influence chain-of-thought (CoT) reasoning, where the model needs to frequently refer back the question and previously-generated context. Conversely, tasks where the model directly generates the answer using its intrinsic knowledge are less impacted by masking out retrieval heads. These observations collectively explain which internal part of the model seeks information from the input tokens. We believe our insights will foster future research on reducing hallucination, improving reasoning, and compressing the KV cache.