Large Language Models (LLMs) require significant GPU memory when processing long texts, with the key value (KV) cache consuming up to 70% of total memory during inference. Although existing compression methods reduce memory by evaluating the importance of individual tokens, they overlook critical semantic relationships between tokens, resulting in fragmented context and degraded performance. We introduce ChunkKV, which fundamentally reimagines KV cache compression by treating semantic chunks - rather than isolated tokens - as basic compression units. This approach preserves complete linguistic structures and contextual integrity, ensuring that essential meaning is retained even under aggressive compression. Our innovation includes a novel layer-wise index reuse technique that exploits the higher cross-layer similarity of preserved indices in ChunkKV, reducing computational overhead and improving throughput by 26.5%. Comprehensive evaluations on challenging benchmarks: LongBench, Needle-InA-HayStack, GSM8K, and JailbreakV demonstrate that ChunkKV outperforms state-of-the-art methods by up to 8.7% in precision while maintaining the same compression ratio. These results confirm that semantic-aware compression significantly enhances both efficiency and performance for long-context LLM inference, providing a simple yet effective solution to the memory bottleneck problem. The code is available at link.

The proliferation of 100B+ parameter Large Language Models (LLMs) with 100k+ context length support have resulted in increasing demands for on-chip memory to support large KV caches. Techniques such as StreamingLLM and SnapKV demonstrate how to control KV cache size while maintaining model accuracy. Yet, these techniques are not commonly used within industrial deployments using frameworks like vLLM or SGLang. The reason is twofold: on one hand, the static graphs and continuous batching methodology employed by these frameworks make it difficult to admit modifications to the standard multi-head attention algorithm, while on the other hand, the accuracy implications of such techniques on modern instruction-following and reasoning models are not well understood, obfuscating the need for implementing these techniques. In this paper, we explore these accuracy implications on Llama-3.1-8B-Instruct and DeepSeek-R1, and develop SnapStream, a KV cache compression method that can be deployed at scale. We demonstrate the efficacy of SnapStream in a 16-way tensor-parallel deployment of DeepSeek-671B on SambaNova SN40L accelerators running at 128k context length and up to 1832 tokens per second in a real production setting. SnapStream enables $4\times$ improved on-chip memory usage and introduces minimal accuracy degradation on LongBench-v2, AIME24 and LiveCodeBench. To the best of our knowledge, this is the first implementation of sparse KV attention techniques deployed in a production inference system with static graphs and continuous batching.

Traditional KV cache eviction strategies, which discard less critical KV-pairs based on attention scores, often degrade generation quality, causing context loss or hallucinations. Recent efforts shift toward KV merging, merging eviction tokens with retention tokens based on similarity. However, in multimodal scenarios, distributional biases across modality tokens and attentional biases in cross-modal interactions limit its effectiveness. This work introduces FlowMM, an adaptive framework for cross-modal information flow-guided multimodal KV cache merging. FlowMM leverages cross-modal information flow to dynamically apply layer-specific merging strategies, capturing modality-specific patterns while preserving contextual integrity. Furthermore, we introduce a sensitivity-adaptive token matching mechanism that jointly evaluates token similarity and task-critical sensitivity, merging low-risk tokens while safeguarding high-sensitivity ones. Extensive experiments across diverse leading MLLMs show that FlowMM reduces KV cache memory by 80% to 95% and decoding latency by 1.3-1.8x, while maintaining competitive task performance.

KV cache compression seeks to discern the saliency of tokens, preserving vital information while aggressively compressing those of less importance.

Recent advancements in Large Visual Language Models (LVLMs) have gained significant attention due to their remarkable reasoning capabilities and proficiency in generalization. However, processing a large number of visual tokens and generating long-context outputs impose substantial computational overhead, leading to excessive demands for key-value (KV) cache. T o address this critical bottleneck, we propose AirCache, a novel KV cache compression method aimed at accelerating LVLMs inference. This work systematically investigates the correlations between visual and textual tokens within the attention mechanisms of LVLMs. Our empirical analysis reveals considerable redundancy in cached visual tokens, wherein strategically eliminating these tokens preserves model performance while significantly accelerating context generation. Inspired by these findings, we introduce an elite observation window for assessing the importance of visual components in the KV cache, focusing on stable inter-modal relevancy modeling with enhanced multi-perspective consistency. Additionally, we develop an adaptive layer-wise budget allocation strategy that capitalizes on the strength and skewness of token importance distribution, showcasing superior efficiency compared to uniform allocation. Comprehensive evaluations across multiple LVLMs and benchmarks demonstrate that our method achieves comparable performance to the full cache while retaining only 10% of visual KV cache, thereby reducing decoding latency by 29% to 66% across various batch size and prompt length of inputs. Notably, as cache retention rates decrease, our method exhibits increasing performance advantages over existing approaches.

This paper investigates an under-explored challenge in large language models (LLMs): the impact of KV cache compression methods on LLMs' fundamental capabilities. While existing methods achieve impressive compression ratios on long-context benchmarks, their effects on core model capabilities remain understudied. We present a comprehensive empirical study evaluating prominent KV cache compression methods across diverse tasks, spanning world knowledge, commonsense reasoning, arithmetic reasoning, code generation, safety, and long-context understanding and generation.Our analysis reveals that KV cache compression methods exhibit task-specific performance degradation. Arithmetic reasoning tasks prove particularly sensitive to aggressive compression, with different methods showing performance drops of $17.4\%$-$43.3\%$. Notably, the DeepSeek R1 Distill model exhibits more robust compression tolerance compared to instruction-tuned models, showing only $9.67\%$-$25.53\%$ performance degradation. Based on our analysis of attention patterns and cross-task compression performance, we propose ShotKV, a novel compression approach that distinctly handles prefill and decoding phases while maintaining shot-level semantic coherence. Empirical results show that ShotKV achieves $9\%$-$18\%$ performance improvements on long-context generation tasks under aggressive compression ratios.

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.

Key-Value (KV) caching is a common technique to enhance the computational efficiency of Large Language Models (LLMs), but its memory overhead grows rapidly with input length. Prior work has shown that not all tokens are equally important for text generation, proposing layer-level KV cache compression to selectively retain key information. Recognizing the distinct roles of attention heads in generation, we propose HeadKV, a head-level KV cache compression method, and HeadKV-R2, which leverages a novel contextual reasoning ability estimation for compression. Our approach operates at the level of individual heads, estimating their importance for contextual QA tasks that require both retrieval and reasoning capabilities. Extensive experiments across diverse benchmarks (LongBench, LooGLE), model architectures (e.g., Llama-3-8B-Instruct, Mistral-7B-Instruct), and long-context abilities tests demonstrate that our head-level KV cache compression significantly outperforms strong baselines, particularly in low-resource settings (KV size = 64 & 128). Notably, our method retains just 1.5% of the KV cache while achieving 97% of the performance of the full KV cache on the contextual question answering benchmark. Modern Large Language Models (LLMs) increasingly support extremely long inputs: GPT-4 (Achiam et al., 2023), Llama-3 (Dubey et al., 2024), and Qwen-2 (Yang et al., 2024) handle up to 128K tokens, while Claude (Anthropic, 2024) supports up to 1 million tokens. These extended capacities improve performance on tasks like dialogue generation (Li et al., 2024a; Yi et al., 2024), question answering (Ho et al., 2020; Xu et al., 2023), and summarization (Xiao & Carenini, 2019; Koh et al., 2022). As input lengths increase, memory usage and latency grow significantly due to self-attention in transformers.

Large language models (LLMs) (Hadi et al., 2023; Brown et al., 2020; OpenAI, 2023; Touvron et al., 2023a,b; Scao et al., 2022; Reid et al., 2024) have emerged as a dominant paradigm in natural language processing, achieving state-of-the-art performance across various tasks. A key principle, the Scaling Law (Kaplan et al., 2020), suggests that LLMs exhibit emergent abilities as model size increases, enhancing their capacity to understand context and handle long sequences (Xiong et al., 2023). This capacity growth allows LLMs to generate coherent and contextually accurate responses and enables various downstream applications, such as document summarization (Zhang et al., 2019, 2024a), code generation (Chen et al., 2021b), and conversational AI (Bordes et al., 2016; OpenAI, 2022),. Despite their success in various applications, the generation of LLMs incurs significant expenses, which escalate with increasing model size and sequence length. Notably, both the training (Strubell et al., 2020; Hoffmann et al., 2022; Dong et al., 2024a) and inference (Ainslie et al., 2023) stages involve frequent generation by LLMs, further contributing to these costs. Consequently, efficient LLMs have gained popularity in recent years (Hu et al., 2021; Wan et al., 2023). To address these challenges, quantization (Frantar et al., 2022; Lin et al., 2024; Dettmers et al., 2024; Xu et al., 2023) and pruning methods (Frankle and Carbin, 2018; Blalock et al., 2020) are employed to reduce model size. Additionally, managing long sequences presents another cost due to the transformer attention mechanism.