Daily Papers

The Mixture-of-Experts (MoE) architecture has demonstrated significant advantages in the era of Large Language Models (LLMs), offering enhanced capabilities with reduced inference costs. However, deploying MoE-based LLMs on memoryconstrained edge devices remains challenging due to their substantial memory requirements. While existing expertoffloading methods alleviate the memory requirements, they often incur significant expert-loading costs or compromise model accuracy. We present HOBBIT, a mixed precision expert offloading system to enable flexible and efficient MoE inference. Our key insight is that dynamically replacing less critical cache-miss experts with low precision versions can substantially reduce expert-loading latency while preserving model accuracy. HOBBIT introduces three innovative techniques that map the natural hierarchy of MoE computation: (1) a token-level dynamic expert loading mechanism, (2) a layer-level adaptive expert prefetching technique, and (3) a sequence-level multidimensional expert caching policy. These innovations fully leverage the benefits of mixedprecision expert inference. By implementing HOBBIT on top of the renowned LLM inference framework Llama.cpp, we evaluate its performance across different edge devices with representative MoE models. The results demonstrate that HOBBIT achieves up to a 9.93x speedup in decoding compared to state-of-the-art MoE offloading systems.

Mixture-of-Experts (MoE) models enable sparse expert activation, meaning that only a subset of the model's parameters is used during each inference. However, to translate this sparsity into practical performance, an expert caching mechanism is required. Previous works have proposed hardware-centric caching policies, but how these various caching policies interact with each other and different hardware specification remains poorly understood. To address this gap, we develop SpecMD, a standardized framework for benchmarking ad-hoc cache policies on various hardware configurations. Using SpecMD, we perform an exhaustive benchmarking of several MoE caching strategies, reproducing and extending prior approaches in controlled settings with realistic constraints. Our experiments reveal that MoE expert access is not consistent with temporal locality assumptions (e.g LRU, LFU). Motivated by this observation, we propose Least-Stale, a novel eviction policy that exploits MoE's predictable expert access patterns to reduce collision misses by up to 85times over LRU. With such gains, we achieve over 88% hit rates with up to 34.7% Time-to-first-token (TTFT) reduction on OLMoE at only 5% or 0.6GB of VRAM cache capacity.

Sparse Mixture of Experts (MoE) models, while outperforming dense Large Language Models (LLMs) in terms of performance, face significant deployment challenges during inference due to their high memory demands. Existing offloading techniques, which involve swapping activated and idle experts between the GPU and CPU, often suffer from rigid expert caching mechanisms. These mechanisms fail to adapt to dynamic routing, leading to inefficient cache utilization, or incur prohibitive costs for prediction training. To tackle these inference-specific challenges, we introduce ExpertFlow, a comprehensive system specifically designed to enhance inference efficiency by accommodating flexible routing and enabling efficient expert scheduling between CPU and GPU. This reduces overhead and boosts system performance. Central to our approach is a predictive routing path-based offloading mechanism that utilizes a lightweight predictor to accurately forecast routing paths before computation begins. This proactive strategy allows for real-time error correction in expert caching, significantly increasing cache hit ratios and reducing the frequency of expert transfers, thereby minimizing I/O overhead. Additionally, we implement a dynamic token scheduling strategy that optimizes MoE inference by rearranging input tokens across different batches. This method not only reduces the number of activated experts per batch but also improves computational efficiency. Our extensive experiments demonstrate that ExpertFlow achieves up to 93.72\% GPU memory savings and enhances inference speed by 2 to 10 times compared to baseline methods, highlighting its effectiveness and utility as a robust solution for resource-constrained inference scenarios.

Mixture-of-Experts (MoE) models activate few experts per token, yet their full expert sets can exceed GPU memory and require repeated weight transfers during decoding. We formulate expert-cache management as a model-side algorithmic problem and propose cache-aware post-training that jointly adapts the MoE backbone and lightweight auxiliary routers while preserving the native inference-time Top-K rule. The update-only Temporal Router learns same-layer retention across tokens without proactive loading. The full Spatio-Temporal Router adds a Spatio Router that uses the causal predecessor's hidden state to refine the temporal cache before target-layer access. We evaluate both modes on Qwen3 and GPT-OSS across GSM8K, MATH, and CommonsenseQA. Temporal Router consistently improves hit rate and reduces expert-weight traffic over matched LM-only baselines. On Qwen3, the full mode improves adjusted hit rate by 1.15--18.03 points and reduces traffic by 4.6--53.3\% relative to the strongest evaluated prefetching baseline; GPT-OSS results are competitive but task-dependent. Auxiliary-only training preserves baseline accuracy but yields modest coverage gains; joint post-training achieves substantially higher coverage. Sensitivity analyses distinguish the effects of cache capacity, refinement budget, and cache-loss weight on coverage, traffic, and quality.

Retrieval-Augmented Generation (RAG) has shown significant improvements in various natural language processing tasks by integrating the strengths of large language models (LLMs) and external knowledge databases. However, RAG introduces long sequence generation and leads to high computation and memory costs. We propose RAGCache, a novel multilevel dynamic caching system tailored for RAG. Our analysis benchmarks current RAG systems, pinpointing the performance bottleneck (i.e., long sequence due to knowledge injection) and optimization opportunities (i.e., caching knowledge's intermediate states). Based on these insights, we design RAGCache, which organizes the intermediate states of retrieved knowledge in a knowledge tree and caches them in the GPU and host memory hierarchy. RAGCache proposes a replacement policy that is aware of LLM inference characteristics and RAG retrieval patterns. It also dynamically overlaps the retrieval and inference steps to minimize the end-to-end latency. We implement RAGCache and evaluate it on vLLM, a state-of-the-art LLM inference system and Faiss, a state-of-the-art vector database. The experimental results show that RAGCache reduces the time to first token (TTFT) by up to 4x and improves the throughput by up to 2.1x compared to vLLM integrated with Faiss.

Mixture-of-Experts (MoE) models have gained popularity in achieving state-of-the-art performance in a wide range of tasks in computer vision and natural language processing. They effectively expand the model capacity while incurring a minimal increase in computation cost during training. However, deploying such models for inference is difficult due to their large size and complex communication pattern. In this work, we provide a characterization of two MoE workloads, namely Language Modeling (LM) and Machine Translation (MT) and identify their sources of inefficiencies at deployment. We propose three optimization techniques to mitigate sources of inefficiencies, namely (1) Dynamic gating, (2) Expert Buffering, and (3) Expert load balancing. We show that dynamic gating improves maximum throughput by 6.21-11.23times for LM, 5.75-10.98times for MT Encoder and 2.58-5.71times for MT Decoder. It also reduces memory usage by up to 1.36times for LM and up to 1.1times for MT. We further propose Expert Buffering, a new caching mechanism that only keeps hot, active experts in GPU memory while buffering the rest in CPU memory. This reduces static memory allocation by up to 1.47times. We finally propose a load balancing methodology that provides additional scalability to the workload.

This paper presents MoE-Infinity, a cost-efficient mixture-of-expert (MoE) serving system that realizes activation-aware expert offloading. MoE-Infinity features sequence-level expert activation tracing, a new approach adept at identifying sparse activations and capturing the temporal locality of MoE inference. By analyzing these traces, MoE-Infinity performs novel activation-aware expert prefetching and caching, substantially reducing the latency overheads usually associated with offloading experts for improved cost performance. Extensive experiments in a cluster show that MoE-Infinity outperforms numerous existing systems and approaches, reducing latency by 4 - 20X and decreasing deployment costs by over 8X for various MoEs. MoE-Infinity's source code is publicly available at https://github.com/TorchMoE/MoE-Infinity

Mixture-of-Experts (MoE) enables efficient scaling of large language models (LLMs) with sparsely activated experts during inference. To effectively deploy large MoE models on memory-constrained devices, many systems introduce *expert offloading* that caches a subset of experts in fast memory, leaving others on slow memory to run on CPU or load on demand. While some research has exploited the locality of expert activations, where consecutive tokens activate similar experts, the degree of this **local routing consistency** varies across models and remains understudied. In this paper, we propose two metrics to measure local routing consistency of MoE models: (1) **Segment Routing Best Performance (SRP)**, which evaluates how well a fixed group of experts can cover the needs of a segment of tokens, and (2) **Segment Cache Best Hit Rate (SCH)**, which measures the optimal segment-level cache hit rate under a given cache size limit. We analyzed 20 MoE LLMs with diverse sizes and architectures and found that models that apply MoE on every layer and do not use shared experts exhibit the highest local routing consistency. We further showed that domain-specialized experts contribute more to routing consistency than vocabulary-specialized ones, and that most models can balance between cache effectiveness and efficiency with cache sizes approximately 2x the active experts. These findings pave the way for memory-efficient MoE design and deployment without compromising inference speed. We publish the code for replicating experiments at https://github.com/ljcleo/moe-lrc .

Fine-grained Mixture-of-Experts (MoE) models sparsely activate only a subset of experts per token, reducing activated computation while maintaining high model capacity. However, in memory-constrained inference scenarios, only a small set of experts can be cached. Experts not in the cache must be fetched from slow external storage (e.g., UFS), leading to frequent evictions and substantial I/O overhead. We propose ReMoE, a router fine-tuning framework designed to boost token-wise expert reuse. ReMoE biases the router toward recently selected experts, producing temporally stable routing that better matches cache locality constraints. By increasing short-horizon expert reuse, ReMoE reduces expert fetches from storage without adding inference-time computation. Experiments on DeepSeek and Qwen models show that ReMoE improves expert reuse by 26% while maintaining downstream task performance. Real-system evaluations further confirm these benefits, improving output throughput by 8.4% under vLLM GPU-CPU expert offloading and reducing TPOT by 43.6-49.8% under llama.cpp on Jetson Orin NX, corresponding to a 1.77-1.99times decode speedup across diverse workloads. Checkpoints and usage instructions are available at https://github.com/BUAA-OSCAR/ReMoE.

Recent research has developed several LLM architectures suitable for inference on end-user devices, such as the Mixture of Lookup Experts (MoLE)~jie_mixture_2025. A key feature of MoLE is that each token id is associated with a dedicated group of experts. For a given input, only the experts corresponding to the input token id will be activated. Since the communication overhead of loading this small number of activated experts into RAM during inference is negligible, expert parameters can be offloaded to storage, making MoLE suitable for resource-constrained devices. However, MoLE's context-independent expert selection mechanism, based solely on input ids, may limit model performance. To address this, we propose the Mixture of Lookup Key-Value Experts (MoLKV) model. In MoLKV, each expert is structured as a key-value pair. For a given input, the input-derived query interacts with the cached key-value experts from the current sequence, generating a context-aware expert output. This context-aware mechanism alleviates the limitation of MoLE, and experimental results demonstrate that MoLKV achieves significantly lower validation loss in small-scale evaluations.

LLM serving systems process heterogeneous query workloads where different categories exhibit different characteristics. Code queries cluster densely in embedding space while conversational queries distribute sparsely. Content staleness varies from minutes (stock data) to months (code patterns). Query repetition patterns range from power-law (code) to uniform (conversation), producing long tail cache hit rate distributions: high-repetition categories achieve 40-60% hit rates while low-repetition or volatile categories achieve 5-15% hit rates. Vector databases must exclude the long tail because remote search costs (30ms) require 15--20% hit rates to break even, leaving 20-30% of production traffic uncached. Uniform cache policies compound this problem: fixed thresholds cause false positives in dense spaces and miss valid paraphrases in sparse spaces; fixed TTLs waste memory or serve stale data. This paper presents category-aware semantic caching where similarity thresholds, TTLs, and quotas vary by query category. We present a hybrid architecture separating in-memory HNSW search from external document storage, reducing miss cost from 30ms to 2ms. This reduction makes low-hit-rate categories economically viable (break-even at 3-5% versus 15-20%), enabling cache coverage across the entire workload distribution. Adaptive load-based policies extend this framework to respond to downstream model load, dynamically adjusting thresholds and TTLs to reduce traffic to overloaded models by 9-17% in theoretical projections.

Mixture-of-Experts (MoE) architectures employ sparse activation to deliver faster training and inference with higher accuracy than dense LLMs. However, in production serving, MoE models require batch inference to optimize hardware efficiency, which may cause excessive expert activation and thus slow the memory-bound decoding stage. To address the fundamental tension between batch decoding and expert sparsity, we present SERE, a Similarity-based Expert Re-routing method for Efficient batch decoding in MoE models. SERE dynamically reduces the number of active experts in an input-aware manner by re-routing tokens from secondary experts to their most similar primary counterparts. It also leverages similarity patterns to identify and preserve critical experts, thereby preventing capability loss. Notably, SERE avoids static expert pruning or merging, instead enabling dynamic expert skipping based on batch-level expert redundancy. Additionally, we provide an efficient custom CUDA kernel for SERE, enabling plug-and-play use in vLLM with only a single-line code change. Extensive experiments on various complex reasoning benchmarks demonstrate that SERE achieves up to 2.0x speedup with minimal quality loss, providing a practical solution for cost-efficient and latency-sensitive large-scale MoE deployment. Code implementation of SERE can be found in https://github.com/JL-Cheng/SERE.

Large Language Models (LLMs) have demonstrated impressive performance across various tasks, and their application in edge scenarios has attracted significant attention. However, sparse-activated Mixture-of-Experts (MoE) models, which are well suited for edge scenarios, have received relatively little attention due to their high memory demands. Offload-based methods have been proposed to address this challenge, but they face difficulties with expert prediction. Inaccurate expert predictions can result in prolonged inference delays. To promote the application of MoE models in edge scenarios, we propose Fate, an offloading system designed for MoE models to enable efficient inference in resource-constrained environments. The key insight behind Fate is that gate inputs from adjacent layers can be effectively used for expert prefetching, achieving high prediction accuracy without additional GPU overhead. Furthermore, Fate employs a shallow-favoring expert caching strategy that increases the expert hit rate to 99\%. Additionally, Fate integrates tailored quantization strategies for cache optimization and IO efficiency. Experimental results show that, compared to Load on Demand and Expert Activation Path-based method, Fate achieves up to 4.5x and 1.9x speedups in prefill speed and up to 4.1x and 2.2x speedups in decoding speed, respectively, while maintaining inference quality. Moreover, Fate's performance improvements are scalable across different memory budgets.

We propose Chain-of-Experts (CoE), a new Mixture-of-Experts (MoE) architecture that introduces sequential expert communication within each layer. Unlike traditional MoE models, where experts operate independently in parallel, CoE processes tokens iteratively across a chain of experts inside a layer. To support dynamic expert selection across iterations, CoE employs a dedicated router at each iteration step within a layer. This design allows tokens to re-evaluate and select different experts during each iteration, rather than being statically assigned. As a result, CoE introduces a flexible routing mechanism that increases the diversity of expert combinations and enriches the model's representational capacity. CoE demonstrates improved performance under fixed compute: on math reasoning tasks, it reduces validation loss from 1.20 to 1.12 compared to a standard MoE. Beyond performance, CoE offers a new scaling axis: depth through expert iteration, which complements conventional width/depth scaling. For example, using 2x iterations matches the performance of 3x expert selections (in width), while reducing memory usage by 17.6-42% relative to other scaling strategies. Our analysis reveals that CoE's benefits stem from its iterative residual structure and enhanced expert specialization empowered by iterative routing, which together unlock more expressive representations. Code is available at https://github.com/ZihanWang314/coe.

This report investigates enhancing semantic caching effectiveness by employing specialized, fine-tuned embedding models. Semantic caching relies on embedding similarity rather than exact key matching, presenting unique challenges in balancing precision, query latency, and computational efficiency. We propose leveraging smaller, domain-specific embedding models, fine-tuned with targeted real-world and synthetically generated datasets. Our empirical evaluations demonstrate that compact embedding models fine-tuned for just one epoch on specialized datasets significantly surpass both state-of-the-art open-source and proprietary alternatives in precision and recall. Moreover, we introduce a novel synthetic data generation pipeline for the semantic cache that mitigates the challenge of limited domain-specific annotated data, further boosting embedding performance. Our approach effectively balances computational overhead and accuracy, establishing a viable and efficient strategy for practical semantic caching implementations.

Mixture-of-Experts (MoE) architectures scale large language models (LLMs) by activating only a subset of experts per token, but the standard TopK routing assigns the same fixed number of experts to all tokens, ignoring their varying complexity. Prior adaptive routing methods introduce additional modules and hyperparameters, often requiring costly retraining from scratch. We propose Sequence-level TopK (SeqTopK), a minimal modification that shifts the expert budget from the token level to the sequence level. By selecting the top T cdot K experts across all T tokens, SeqTopK enables end-to-end learned dynamic allocation -- assigning more experts to difficult tokens and fewer to easy ones -- while preserving the same overall budget. SeqTopK requires only a few lines of code, adds less than 1% overhead, and remains fully compatible with pretrained MoE models. Experiments across math, coding, law, and writing show consistent improvements over TopK and prior parameter-free adaptive methods, with gains that become substantially larger under higher sparsity (up to 16.9%). These results highlight SeqTopK as a simple, efficient, and scalable routing strategy, particularly well-suited for the extreme sparsity regimes of next-generation LLMs. Code is available at https://github.com/Y-Research-SBU/SeqTopK.

Long-context ability, has become one of the most important iteration direction of next-generation Large Language Models, particularly in semantic understanding/reasoning, code agentic intelligence and recommendation system. However, the standard softmax attention exhibits quadratic time complexity with respect to sequence length. As the sequence length increases, this incurs substantial overhead in long-context settings, leading the training and inference costs of extremely long sequences deteriorate rapidly. Existing solutions mitigate this issue through two technique routings: i) Reducing the KV cache per layer, such as from the head-level compression GQA, and the embedding dimension-level compression MLA, but the KV cache remains linearly dependent on the sequence length at a 1:1 ratio. ii) Interleaving with KV Cache friendly architecture, such as local attention SWA, linear kernel GDN, but often involve trade-offs among KV Cache and long-context modeling effectiveness. Besides the two technique routings, we argue that there exists an intermediate path not well explored: {Maintaining a linear relationship between the KV cache and sequence length, but performing semantic-level compression through a specific ratio k}. This O(n/k) path does not pursue a ``minimum KV cache'', but rather trades acceptable memory costs for complete, referential, and interpretable retention of long distant dependency. Motivated by this, we propose Kwai Summary Attention (KSA), a novel attention mechanism that reduces sequence modeling cost by compressing historical contexts into learnable summary tokens.

The proliferation of large language models (LLMs) has led to the adoption of Mixture-of-Experts (MoE) architectures that dynamically leverage specialized subnetworks for improved efficiency and performance. Despite their benefits, MoE models face significant challenges during inference, including inefficient memory management and suboptimal batching, due to misaligned design choices between the model architecture and the system policies. Furthermore, the conventional approach of training MoEs from scratch is increasingly prohibitive in terms of cost. In this paper, we propose a novel framework Read-ME that transforms pre-trained dense LLMs into smaller MoE models (in contrast to "upcycling" generalist MoEs), avoiding the high costs of ground-up training. Our approach employs activation sparsity to extract experts. To compose experts, we examine the widely-adopted layer-wise router design and show its redundancy, and thus we introduce the pre-gating router decoupled from the MoE backbone that facilitates system-friendly pre-computing and lookahead scheduling, enhancing expert-aware batching and caching. Our codesign therefore addresses critical gaps on both the algorithmic and system fronts, establishing a scalable and efficient alternative for LLM inference in resource-constrained settings. Read-ME outperforms other popular open-source dense models of similar scales, achieving improvements of up to 10.1% on MMLU, and improving mean end-to-end latency up to 6.1%. Codes are available at: https://github.com/VITA-Group/READ-ME.

The promising applications of large language models are often limited by the constrained GPU memory capacity available on edge devices. Mixture-of-Experts (MoE) models help address this issue by activating only a subset of the model's parameters during computation. This approach allows the unused parameters to be offloaded to host memory, thereby reducing the overall GPU memory demand. However, existing cache-based offloading solutions handle cache misses reactively, which significantly impacts system performance. In this paper, we introduce ProMoE, a novel proactive caching system that utilizes intermediate results to predict subsequent expert usage. By proactively fetching experts in advance, ProMoE eliminates passive cache misses, removes loading time from the critical path, and reduces the performance overhead associated with offloading. Our evaluations demonstrate that ProMoE achieves an average speedup of 2.20x (up to 3.21x) and 2.07x (up to 5.02x) in the prefill and decode stages, respectively, compared to existing offloading solutions.

Among parallel decoding paradigms, diffusion large language models (dLLMs) have emerged as a promising candidate that balances generation quality and throughput. However, their integration with Mixture-of-Experts (MoE) architectures is constrained by an expert explosion: as the number of tokens generated in parallel increases, the number of distinct experts activated grows nearly linearly. This results in substantial memory traffic that pushes inference into a memory-bound regime, negating the efficiency gains of both MoE and parallel decoding. To address this challenge, we propose Dynamic Expert Sharing (DES), a novel technique that shifts MoE optimization from token-centric pruning and conventional expert skipping methods to sequence-level coreset selection. To maximize expert reuse, DES identifies a compact, high-utility set of experts to satisfy the requirements of an entire parallel decoding block. We introduce two innovative selection strategies: (1) Intra-Sequence Sharing (DES-Seq), which adapts optimal allocation to the sequence level, and (2) Saliency-Aware Voting (DES-Vote), a novel mechanism that allows tokens to collectively elect a coreset based on aggregated router weights. Extensive experiments on MoE dLLMs demonstrate that DES reduces unique expert activations by over 55% and latency by up to 38%, while retaining 99% of vanilla accuracy, effectively decoupling memory overhead from the degree of parallelism.

Serving transformer language models with high throughput requires caching Key-Values (KVs) to avoid redundant computation during autoregressive generation. The memory footprint of KV caching is significant and heavily impacts serving costs. This work proposes to lessen these memory requirements. While recent work has largely addressed KV cache reduction via compression and eviction along the temporal axis, we argue that the depth dimension offers an orthogonal and robust avenue for optimization. Although prior research suggests that a full cache for every layer is redundant, implementing cross-layer cache sharing remains a practical challenge; existing methods typically suffer from reduced throughput or increased time-to-first-token. In this paper, we demonstrate that dropping a layer's cache offers efficient optimization without information loss. We propose a simple training approach: random cross-layer attention. During training, layers randomly choose to attend either to their own KV states or those of a preceding layer. This stochastic process adapts the model to be robust to various depth-wise cache sharing strategies, ensuring flexibility for unknown hardware constraints at deployment time. Our evaluations show that applying this scheme during pre-training or fine-tuning enables depth-wise cache sharing for various model families. Furthermore, for larger models in data-constrained settings, this approach is suggestive of a regularization-like effect, frequently preserving or improving performance while significantly reducing the cache's memory footprint.

Mixture-of-experts (MoE) language models often exceed the memory of a single GPU. Expert offloading keeps most experts in host memory and loads them on demand, so decoding speed depends on how many experts each token must fetch. Caching and prefetching reduce this cost only as far as the routing allows. Router-only fine-tuning can reshape the routing to reuse experts, but it keeps the experts frozen, so they cannot adapt to the tokens the new routing sends them. We propose MaskCoFT, a masked co-adaptive fine-tuning method that trains routers and experts together with the cross-entropy loss alone. During fine-tuning, a learnable binary mask restricts the Top-K routing of each layer to a subset of experts, and the experts adapt to the tokens redirected to them. At inference, the learned mask becomes a soft prior that re-ranks experts, so every expert remains selectable. We simulate a GPU cache of 4 experts per layer for Mixtral-8x7B and 12 for DeepSeek-V2-Lite. MaskCoFT cuts expert fetches per token by 23.7% and 10.1% relative to the base model. In real offloading system serving, it lowers the time per output token by up to 16.4% and 5.5%, respectively. Its average accuracy over nine benchmarks stays above the base model by 0.92 and 0.53 points.

ExpertRAG is a novel theoretical framework that integrates Mixture-of-Experts (MoE) architectures with Retrieval Augmented Generation (RAG) to advance the efficiency and accuracy of knowledge-intensive language modeling. We propose a dynamic retrieval gating mechanism coupled with expert routing, enabling the model to selectively consult an external knowledge store or rely on specialized internal experts based on the query's needs. The paper lays out the theoretical foundations of ExpertRAG, including a probabilistic formulation that treats retrieval and expert selection as latent decisions, and mathematical justifications for its efficiency in both computation and knowledge utilization. We derive formulae to quantify the expected computational cost savings from selective retrieval and the capacity gains from sparse expert utilization. A comparative analysis positions ExpertRAG against standard RAG (with always-on retrieval) and pure MoE models (e.g., Switch Transformer, Mixtral) to highlight its unique balance between parametric knowledge and non-parametric retrieval. We also outline an experimental validation strategy, proposing benchmarks and evaluation protocols to test ExpertRAG's performance on factual recall, generalization, and inference efficiency. The proposed framework, although presented theoretically, is supported by insights from prior work in RAG and MoE, and is poised to provide more factual, efficient, and adaptive generation by leveraging the best of both paradigms. In summary, ExpertRAG contributes a new perspective on scaling and augmenting language models, backed by a thorough analysis and a roadmap for empirical validation.

When a Mixture-of-Experts (MoE) model does not fit in GPU memory, batch-1 decoding is dominated by copying expert weights from host memory. Cache-aware routing reduces these copies by biasing the router toward resident experts, but combined with asynchronous prefetching it makes the model's outputs depend on transfer timing, so a harness that does not simulate that timing no longer evaluates the deployed model. We propose prefetch-steered routing (PSR): routing conditions on a deterministic logical cache, into which at most n experts predicted by the next layer's router are inserted per layer and token; physical copies affect only latency. On OLMoE-1B-7B with 50% and 25% of experts resident on an RTX 4090, PSR decodes 2.2-2.3x faster than exact routing with LRU and 25-30% faster than the Cache-Prior operating point, with mean benchmark differences of +0.1 and +0.5 points (95% lower bounds -0.8 and -0.3). In greedy generation its lead over Cache-Prior depends on the workload and reverses on HumanEval at 50% residency. On Moonlight-16B-A3B, PSR is 1.9x faster than exact routing with LRU at 37.5% residency and, at the latency of Cache-Prior, significantly better on three of eight benchmarks. On Qwen3.6-35B-A3B, where copies are cheap relative to computation, PSR keeps its advantage over Cache-Prior on the likelihood benchmarks but not its speed. Distillation-trained gates did not help. Speculative decoding with block-level expert selection was slower than PSR with a 1B autoregressive drafter on OLMoE, but with a block drafter on Qwen3.6 it was 2.1-2.9x faster than any autoregressive configuration with at most its GPU memory on the 64 prompts we timed.

As model parameters scale up, sparsely activated Mixture-of-Experts (MoE) approaches have become essential for model scaling. However, mainstream models currently adopt hierarchical architectures, where experts in different layers, although topologically independent, exhibit substantial similarity and redundancy in their stored content. Compared to the hierarchically arranged Local Expert in existing mainstream models, we hypothesize that inter-layer shared Meta Expert hold greater potential. To this end, we conduct a multi-dimensional empirical comparison between Local and Meta Experts, and find that models incorporating Meta Experts exhibit lower parameter redundancy, together with lower cross-entropy loss and higher downstream task accuracy. With the activated quantity and the expert granularity held fixed, our model matches the performance of Vanilla MoE using only 70--80\% of the experts, substantially improving the model's scaling potential. Beyond removing inter-layer redundancy among experts and improving information storage efficiency, our theoretical analysis further shows that, compared with Local Experts, Meta Experts provide hidden states with more flexible residual pathways, thereby improving information retrieval efficiency.

This paper introduces Neurocache, an approach to extend the effective context size of large language models (LLMs) using an external vector cache to store its past states. Like recent vector retrieval approaches, Neurocache uses an efficient k-nearest-neighbor (kNN) algorithm to retrieve relevant past states and incorporate them into the attention process. Neurocache improves upon previous methods by (1) storing compressed states, which reduces cache size; (2) performing a single retrieval operation per token which increases inference speed; and (3) extending the retrieval window to neighboring states, which improves both language modeling and downstream task accuracy. Our experiments show the effectiveness of Neurocache both for models trained from scratch and for pre-trained models such as Llama2-7B and Mistral-7B when enhanced with the cache mechanism. We also compare Neurocache with text retrieval methods and show improvements in single-document question-answering and few-shot learning tasks. We made the source code available under: https://github.com/alisafaya/neurocache

Multimodal Large Language Models (MLLMs) achieve strong performance through instruction tuning, but real-world deployment requires them to continually expand their capabilities, making Multimodal Continual Instruction Tuning (MCIT) essential. Recent methods leverage sparse expert routing to promote task specialization, but we find that the expert routing process suffers from drift as the data distribution evolves. For example, a grounding query that previously activated localization experts may instead be routed to irrelevant experts after learning OCR tasks. Meanwhile, the grounding-related experts can be overwritten by new tasks and lose their original functionality. Such failure reflects two problems: router drift, where expert selection becomes inconsistent over time, and expert drift, where shared experts are overwritten across tasks. Therefore, we propose StAbilized Mixture-of-Experts (SAME) for MCIT. To address router drift, SAME stabilizes expert selection by decomposing routing dynamics into orthogonal subspaces and updating only task-relevant directions. To mitigate expert drift, we regulate expert updates via curvature-aware scaling using historical input covariance in a rehearsal-free manner. SAME also introduces adaptive expert activation to freeze selected experts during training, reducing redundant computation and cross-task interference. Extensive experiments demonstrate its SOTA performance.

Conditional computation can decouple language model quality from per-token inference cost, yet leading techniques act on a single axis in isolation: Mixture-of-Experts (MoE) sparsifies the FFN, Mixture-of-Depths (MoD) skips whole transformer blocks, and KV-cache quantization compresses attention memory. We argue these three decisions (attention resolution, expert selection, and cache bit-width) are strongly coupled and should be made jointly: a token rare enough to warrant full attention may also need high-precision caching regardless of which expert processes it. We introduce TriRoute, a single lightweight controller shared across all three axes that, for every token at every layer, emits a coordinated policy: (i) an attention mode (skip/local/full), (ii) a sparse set of FFN experts (with a null expert recovering MoD), and (iii) a KV-cache bit-width. The controller trains end-to-end via a heterogeneous relaxation (Gumbel-Softmax with straight-through estimation for categorical decisions and load-balanced top-k gating for experts) under a Lagrangian budget constraint that turns the average compute and memory cost into a controllable knob. We identify a cross-axis routing-collapse cascade in naive joint training, where collapse on one axis propagates to the others, and address it with per-axis normalization and a coupling-aware balancing loss. On decoder-only models from 160M to 1.3B parameters at compute-optimal token counts, TriRoute Pareto-dominates the best independent MoD+MoE+KV-quantization combination at matched inference FLOPs and memory, while better preserving tail-case robustness on rare entities, code, and arithmetic that pure perplexity optimization erodes. Post-hoc analysis reveals interpretable structure: the controller allocates full attention and high-precision cache to sentence-initial positions, rare subwords, and named entities, while cheaply routing function words.

Mixture-of-experts (MoE) models activate few experts per token, yet batched decoding can access nearly the entire expert pool, making expert-weight traffic a major bottleneck. Expert pruning reduces this traffic, but conventional approaches also prune compute-bound prefill, sacrificing model quality for little throughput benefit. We present SlimWise, a serving framework that tailors the expert pool to each inference phase. SlimWise performs prefill with the full model and decode with a pruned model that directly reuses the prefill-generated KV cache without conversion. Across two MoE backbones and three pruning criteria, this training-free KV cache handoff substantially narrows accuracy gaps relative to the full model in many settings. We also show that benchmark accuracy can conceal substantial pruning-induced changes in generation length. To address these distortions and residual accuracy loss, SlimWise introduces a low-cost distillation stage that trains the decoder to continue from full-model KV caches while updating only a small subset of parameters. Implemented in vLLM, SlimWise supports both prefill-decode (PD) disaggregation and PD-colocated serving. On Qwen3.6-35B-A3B, SlimWise improves decode throughput by up to 1.81x at 50% expert pruning with minimal accuracy loss.

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.

The linear memory growth of the KV cache poses a significant bottleneck for LLM inference in long-context tasks. Existing static compression methods often fail to preserve globally important information. Although recent dynamic retrieval approaches attempt to address this issue, they typically suffer from coarse-grained caching strategies and incur high I/O overhead. To overcome these limitations, we propose HeteroCache, a training-free dynamic compression framework. Our method is built on two key insights: attention heads exhibit diverse temporal heterogeneity, and there is significant spatial redundancy among heads within the same layer. Guided by these insights, HeteroCache categorizes heads based on stability and similarity, applying a fine-grained weighting strategy that allocates larger cache budgets to heads with rapidly shifting attention to capture context changes. Furthermore, it features a hierarchical storage mechanism where representative heads monitor attention drift to trigger asynchronous, on-demand context retrieval, thereby hiding I/O latency. Experiments demonstrate that HeteroCache achieves state-of-the-art performance on long-context benchmarks and accelerates decoding by up to 3times compared to the original model with a 224K context. Our code is available at https://github.com/ponytaill/HeteroCache.

Hybrid models that combine the language modeling capabilities of Attention layers with the efficiency of Recurrent layers (e.g., State Space Models) have gained traction in practically supporting long contexts in Large Language Model serving. Yet, the unique properties of these models complicate the usage of complementary efficiency optimizations such as prefix caching that skip redundant computations across requests. Most notably, their use of in-place state updates for recurrent layers precludes rolling back cache entries for partial sequence overlaps, and instead mandates only exact-match cache hits; the effect is a deluge of (large) cache entries per sequence, most of which yield minimal reuse opportunities. We present Marconi, the first system that supports efficient prefix caching with Hybrid LLMs. Key to Marconi are its novel admission and eviction policies that more judiciously assess potential cache entries based not only on recency, but also on (1) forecasts of their reuse likelihood across a taxonomy of different hit scenarios, and (2) the compute savings that hits deliver relative to memory footprints. Across diverse workloads and Hybrid models, Marconi achieves up to 34.4times higher token hit rates (71.1% or 617 ms lower TTFT) compared to state-of-the-art prefix caching systems.

The increasing complexity of AI tasks has shifted the paradigm from monolithic models toward multi-agent large language model (LLM) systems. However, these collaborative architectures introduce a critical bottleneck: redundant prefill computation for shared content generated by previous agents, which significantly increases KV cache memory usage and time-to-first-token (TTFT). While various KV cache methods have been proposed to mitigate prefill redundancy, they either fail to maintain accuracy on agent-generated outputs or exhibit low reuse rates due to rigid constraints. We present RelayCaching, a training-free inference method that directly reuses decoding phase KV caches from previous agents in subsequent prefill phases. Our key insight is that KV caches for identical content are highly consistent across phases, while prefix-induced deviations are sparse and localized within a limited range of layers and token positions. By selectively recomputing KV caches at these positions, RelayCaching preserves model accuracy with minimal overhead, yielding a superior accuracy-efficiency trade-off over existing methods. Experiments on diverse collaborative LLM tasks spanning mathematical reasoning, general knowledge, and code generation demonstrate that RelayCaching achieves over 80% KV cache reuse, reduces TTFT by up to 4.7times compared to the standard pipeline, all with negligible accuracy degradation.

Self-attention gives LLMs fine-grained, query-dependent access to context, but dense token interactions incur quadratic prefill cost and a key--value cache growing with context length. Research thus spans explicit-memory compression, sparse access, recurrent state construction, structured state dynamics, and heterogeneous mechanism composition. This survey analyzes these developments as model-internal contextual memory. We introduce a five-dimensional lens---Memory Representation, Memory Update, Access, Readout, and Integration---describing what is represented, how it changes, what is query-eligible, how it is read, and how readouts form outputs. This lens compares overlapping research lines without imposing one computational model. We reconstruct mechanism-level developments and architectural adoption using 59 release-level records from 14 major model lineages and 11 high-performing open-weight endpoints. First, explicit-memory and recurrent-state methods retain distinct interfaces but increasingly control overlapping memory functions. Second, heterogeneous architectures increasingly coordinate across network depth: layer-wise composition distributes complementary memory processing across representational stages, while cross-layer reuse carries selected memory and routing artifacts forward. Depth thus becomes a dimension along which contextual memory is constructed and managed. Third, these developments motivate a stateful multidimensional memory-routing hypothesis: persistent memory is organized across temporal scope, network depth, substrate type, and representation granularity, while coordinated Sparse Write and Sparse Read determine what is maintained and what contributes to each query. Overall, efficient sequence architecture design increasingly concerns the organization, lifecycle, and selective use of contextual memory rather than an isolated Attention operator.

We present Sparse Interpolated Mixture-of-Experts (SIMoE) instruction-tuning, an end-to-end algorithm designed to fine-tune a dense pre-trained Large Language Model (LLM) into a MoE-style model that possesses capabilities in multiple specialized domains. During instruction-tuning, SIMoE automatically identifies multiple specialized experts under a specified sparsity constraint, with each expert representing a structurally sparse subset of the seed LLM's parameters that correspond to domain-specific knowledge within the data. SIMoE simultaneously learns an input-dependent expert merging strategy via a router network, leveraging rich cross-expert knowledge for superior downstream generalization that surpasses existing baselines. Empirically, SIMoE consistently achieves state-of-the-art performance on common instruction-tuning benchmarks while maintaining an optimal performance-compute trade-off compared to all baselines.

We consider the problem of ranking n experts based on their performances on d tasks. We make a monotonicity assumption stating that for each pair of experts, one outperforms the other on all tasks. We consider the sequential setting where in each round, the learner has access to noisy evaluations of actively chosen pair of expert-task, given the information available up to the actual round. Given a confidence parameter delta in (0, 1), we provide strategies allowing to recover the correct ranking of experts and develop a bound on the total number of queries made by our algorithm that hold with probability at least 1 -- delta. We show that our strategy is adaptive to the complexity of the problem (our bounds are instance dependent), and develop matching lower bounds up to a poly-logarithmic factor. Finally, we adapt our strategy to the relaxed problem of best expert identification and provide numerical simulation consistent with our theoretical results.

Incorporating external knowledge in large language models (LLMs) enhances their utility across diverse applications, but existing methods have trade-offs. Retrieval-Augmented Generation (RAG) fetches evidence via similarity search, but key information may fall outside top ranked results. Long-context models can process multiple documents but are computationally expensive and limited by context window size. Inspired by students condensing study material for open-book exams, we propose task-aware key-value (KV) cache compression, which compresses external knowledge in a zero- or few-shot setup. This enables LLMs to reason efficiently over a compacted representation of all relevant information. Experiments show our approach outperforms both RAG and task-agnostic compression methods. On LongBench v2, it improves accuracy by up to 7 absolute points over RAG with a 30x compression rate, while reducing inference latency from 0.43s to 0.16s. A synthetic dataset highlights that RAG performs well when sparse evidence suffices, whereas task-aware compression is superior for broad knowledge tasks.

While Key-Value (KV) cache succeeds in reducing redundant computations in auto-regressive models, it introduces significant memory overhead, limiting its practical deployment in long-sequence scenarios. Existing KV retrieval methods mitigate this by dynamically retaining only a subset of KV entries on the GPU. However, they still suffer from notable efficiency and accuracy bottlenecks due to per-token retrieval and coarse-grained page-level KV management, especially in long-output reasoning scenarios. With the emergence of large reasoning models, efficiently handling such scenarios has become increasingly important. To address this issue, we present two key observations: (1) critical KVs exhibit strong temporal locality during decoding, and (2) these KVs exhibit distinct distribution patterns across the input prompt and generated output. Building on these observations, we propose LouisKV, an efficient KV cache retrieval framework designed for various long-sequence scenarios. Specifically, LouisKV introduces a semantic-aware retrieval strategy leveraging temporal locality to trigger retrieval only at semantic boundaries, drastically reducing computation and data transfer overhead. LouisKV also designs a decoupled, fine-grained management scheme that tailors differentiated strategies for input and output sequences to create retrieval units that better match the model's attention patterns, enabling precise identification of critical KVs. Furthermore, to boost efficiency, LouisKV incorporates several kernel-level optimizations, including custom Triton and CUDA kernels to accelerate the KV clustering and retrieval. Evaluations show that LouisKV achieves up to 4.7times speedup over state-of-the-art KV retrieval methods while maintaining near-lossless accuracy across diverse long-sequence tasks, including long-input short-output, short-input long-output, and long-input long-output scenarios.

Mixture-of-Experts (MoE) models are designed to enhance the efficiency of large language models (LLMs) without proportionally increasing the computational demands. However, their deployment on edge devices still faces significant challenges due to high on-demand loading overheads from managing sparsely activated experts. This paper introduces AdapMoE, an algorithm-system co-design framework for efficient MoE inference. AdapMoE features adaptive expert gating and management to reduce the on-demand loading overheads. We observe the heterogeneity of experts loading across layers and tokens, based on which we propose a sensitivity-based strategy to adjust the number of activated experts dynamically. Meanwhile, we also integrate advanced prefetching and cache management techniques to further reduce the loading latency. Through comprehensive evaluations on various platforms, we demonstrate AdapMoE consistently outperforms existing techniques, reducing the average number of activated experts by 25% and achieving a 1.35x speedup without accuracy degradation. Code is available at: https://github.com/PKU-SEC-Lab/AdapMoE.

Diffusion-based large language models (dLLMs) rely on bidirectional attention, which prevents lossless KV caching and requires a full forward pass at every denoising step. Existing approximate KV caching methods reduce this cost by selectively updating cached states, but their decision overhead scales with context length or model depth. We propose EntropyCache, a training-free KV caching method that uses the maximum entropy of newly decoded token distributions as a constant-cost signal for deciding when to recompute. Our design is grounded in two empirical observations: (1) decoded token entropy correlates with KV cache drift, providing a cheap proxy for cache staleness, and (2) feature volatility of decoded tokens persists for multiple steps after unmasking, motivating recomputation of the k most recently decoded tokens. The skip-or-recompute decision requires only O(V) computation per step, independent of context length and model scale. Experiments on LLaDA-8B-Instruct and Dream-7B-Instruct show that EntropyCache achieves 15.2times-26.4times speedup on standard benchmarks and 22.4times-24.1times on chain-of-thought benchmarks, with competitive accuracy and decision overhead accounting for only 0.5% of inference time. Code is available at https://github.com/mscheong01/EntropyCache.

We present significant extensions to diffusion-based sequence generation models, blurring the line with autoregressive language models. We introduce hyperschedules, which assign distinct noise schedules to individual token positions, generalizing both autoregressive models (e.g., GPT) and conventional diffusion models (e.g., SEDD, MDLM) as special cases. Second, we propose two hybrid token-wise noising processes that interpolate between absorbing and uniform processes, enabling the model to fix past mistakes, and we introduce a novel inference algorithm that leverages this new feature in a simplified context inspired from MDLM. To support efficient training and inference, we design attention masks compatible with KV-caching. Our methods achieve state-of-the-art perplexity and generate diverse, high-quality sequences across standard benchmarks, suggesting a promising path for autoregressive diffusion-based sequence generation.

Diffusion models achieve state-of-the-art video generation quality, but their inference remains expensive due to the large number of sequential denoising steps. This has motivated a growing line of research on accelerating diffusion inference. Among training-free acceleration methods, caching reduces computation by reusing previously computed model outputs across timesteps. Existing caching methods rely on heuristic criteria to choose cache/reuse timesteps and require extensive tuning. We address this limitation with a principled sensitivity-aware caching framework. Specifically, we formalize the caching error through an analysis of the model output sensitivity to perturbations in the denoising inputs, i.e., the noisy latent and the timestep, and show that this sensitivity is a key predictor of caching error. Based on this analysis, we propose Sensitivity-Aware Caching (SenCache), a dynamic caching policy that adaptively selects caching timesteps on a per-sample basis. Our framework provides a theoretical basis for adaptive caching, explains why prior empirical heuristics can be partially effective, and extends them to a dynamic, sample-specific approach. Experiments on Wan 2.1, CogVideoX, and LTX-Video show that SenCache achieves better visual quality than existing caching methods under similar computational budgets.

Mixture-of-Experts architectures have revolutionized scaling, yet the underlying logic of their routing remains a black box. In this paper, we uncover a fundamental governing principle: MoE routing is not merely selection, but a manifestation of Huffman Coding. We introduce the Frequency-Diversity Law, revealing that state-of-the-art models, such as Phi-3.5-MoE and Gemma-4-27B-A4B, spontaneously act as information-theoretic engines. These models allocate sparse expert resources for common tokens while invoking high-diversity expert committees for rare, complex tasks found in chain-of-thought trajectories. However, we identify a critical redundancy trap in Qwen3.5-35B-A3B: when effective sparsity (k/E_eff) is sufficiently low, load-balancing inadvertently imposes functional redundancy, masking the underlying Huffman efficiency signal. To bridge this gap, we propose Subset Difference Pruning, a surgical strategy to eliminate functional duplicates. We demonstrate that pruning does not degrade reasoning; instead, it unleashes the model's latent Huffman efficiency, forcing the logic to collapse into streamlined, high-density paths. Our findings suggest that the next generation of MoEs should move beyond forced load-balancing toward Minimum Description Length (MDL) optimality, assigning shorter expert-routing codes to high-frequency information and longer, more diverse codes to low-frequency information, thereby transforming routing from a heuristic into a principled compression engine.

The Mixture of Experts (MoE) architecture has demonstrated significant advantages as it enables to increase the model capacity without a proportional increase in computation. However, the large MoE model size still introduces substantial memory demands, which usually requires expert offloading on resource-constrained platforms and incurs significant overhead. Hybrid CPU-GPU inference has been proposed to leverage CPU computation to reduce expert loading overhead but faces major challenges: on one hand, the expert activation patterns of MoE models are highly unstable, rendering the fixed mapping strategies in existing works inefficient; on the other hand, the hybrid CPU-GPU schedule for MoE is inherently complex due to the diverse expert sizes, structures, uneven workload distribution, etc. To address these challenges, in this paper, we propose HybriMoE, a hybrid CPU-GPU inference framework that improves resource utilization through a novel CPU-GPU scheduling and cache management system. HybriMoE introduces (i) a dynamic intra-layer scheduling strategy to balance workloads across CPU and GPU, (ii) an impact-driven inter-layer prefetching algorithm, and (iii) a score-based caching algorithm to mitigate expert activation instability. We implement HybriMoE on top of the kTransformers framework and evaluate it on three widely used MoE-based LLMs. Experimental results demonstrate that HybriMoE achieves an average speedup of 1.33times in the prefill stage and 1.70times in the decode stage compared to state-of-the-art hybrid MoE inference framework. Our code is available at: https://github.com/PKU-SEC-Lab/HybriMoE.

Despite the significant success of large language models (LLMs), their extensive memory requirements pose challenges for deploying them in long-context token generation. The substantial memory footprint of LLM decoders arises from the necessity to store all previous tokens in the attention module, a requirement imposed by key-value (KV) caching. In this work, our focus is on developing an efficient compression technique for the KV cache. Empirical evidence indicates a significant clustering tendency within key embeddings in the attention module. Building on this key insight, we have devised a novel caching method with sublinear complexity, employing online clustering on key tokens and online ell_2 sampling on values. The result is a provably accurate and efficient attention decoding algorithm, termed SubGen. Not only does this algorithm ensure a sublinear memory footprint and sublinear time complexity, but we also establish a tight error bound for our approach. Empirical evaluations on long-context question-answering tasks demonstrate that SubGen significantly outperforms existing and state-of-the-art KV cache compression methods in terms of performance and efficiency.

Mixture-of-Experts (MoE) models have become the dominant architecture for large-scale language models, yet on-premises serving remains fundamentally memory-bound as batching turns sparse per-token compute into dense memory activation. Memory-centric architectures (PIM, NMP) improve bandwidth but leave compute underutilized under MoE's low arithmetic intensity at high batch sizes. Speculative decoding (SD) trades idle compute for fewer target invocations, yet verification must load experts even for rejected tokens, severely limiting its benefit in MoE especially at low batch sizes. We propose ELMoE-3D, a hybrid-bonding (HB)-based HW-SW co-designed framework that unifies cache-based acceleration and speculative decoding to offer overall speedup across batch sizes. We identify two intrinsic elasticity axes of MoE-expert and bit-and jointly scale them to construct Elastic Self-Speculative Decoding (Elastic-SD), which serves as both an expert cache and a strongly aligned self-draft model accelerated by high HB bandwidth. Our LSB-augmented bit-sliced architecture exploits inherent redundancy in bit-slice representations to natively support bit-nested execution. On our 3D-stacked hardware, ELMoE-3D achieves an average 6.6times speedup and 4.4times energy efficiency gain over naive MoE serving on xPU across batch sizes 1-16, and delivers 2.2times speedup and 1.4times energy efficiency gain over the best-performing prior accelerator baseline.

Mixture-of-Experts (MoE) architectures have emerged as a promising approach to scale Large Language Models (LLMs). MoE boosts the efficiency by activating a subset of experts per token. Recent works show that fine-grained experts substantially enriches the combinatorial flexibility of active experts and enhances model expressiveness. However, such a design is fundamentally limited by the layer-local routing mechanism: each layer is restricted to its own expert pool. This requires a careful trade-off between expert dimensionality and routing diversity given fixed parameter budgets. We describe ReXMoE, a novel MoE architecture that improves routing beyond the existing layer-local approaches by allowing routers to reuse experts across adjacent layers. ReXMoE decouples expert dimensionality from per-layer budgets, enabling richer expert combinations without sacrificing individual expert capacity or inflating overall parameters. To this end, we propose a new progressive scaling routing (PSR) strategy to gradually increase the candidate expert pool during training. As a result, ReXMoE improves both language modeling and downstream task performance. Extensive experiments on models ranging from 0.5B to 7B parameters across different architectures demonstrate that ReXMoE consistently improves performance under fixed architectural dimensions, confirming ReXMoE as new design paradigm for parameter-efficient and scalable MoE-based LLMs.

Semantic caching cuts LLM inference costs by serving a cached response to semantically similar queries. Standard practice evaluates these systems using PR-AUC, a metric that only measures how well scores rank and ignores whether they are usable at a fixed threshold. We show this mismatch leads to systematically poor deployment choices, as models with the highest PR-AUC are often the worst in operation. We introduce Precision--Cache Hit Ratio (P-CHR) AUC, a cache-aware metric that measures precision across cache utilization levels, and Operational Retention Rate (ORR), which captures how much offline ranking quality survives at deployment. We decompose the operational gap between offline and deployed quality into a recoverable threshold-utility component and an irreducible structural component fixed by the dataset's positive rate. Our experiments show that the threshold-utility gap is governed by the training objective rather than data scale, and yields only to re-normalizing scores over the candidate pool or changing the training objective. Ultimately, model selection for semantic caching is a threshold-utility problem, not a ranking one, and measuring it is the first step to closing the gap.

Long-context LLMs demand accurate inference at low latency, yet decoding becomes primarily constrained by KV cache as context grows. Prior pruning methods are largely context-agnostic: their token selection ignores step-wise relevance and local semantics, which undermines quality. Moreover, their irregular accesses and selection overheads yield only limited wall-clock speedups. To address this, we propose CHESS, an algorithm-system co-design KV-cache management system. Algorithmically, CHESS introduces a context-aware, hierarchical selection policy that dynamically reconstructs a coherent context for the current decoding. System-wise, coarse granularity selection eliminates expensive data movement, fully realizing practical acceleration from theoretical sparsity. Extensive evaluations demonstrate that CHESS surpasses Full-KV quality using only 1\% of the KV cache, delivers low-latency stable inference with up to 4.56times higher throughput, and consistently outperforms other strong baselines. Code is available at https://anonymous.4open.science/r/CHESS-9958/{https://anonymous.4open.science/r/CHESS/}.

Large language models (LLMs) have demonstrated exceptional capabilities in generating text, images, and video content. However, as context length grows, the computational cost of attention increases quadratically with the number of tokens, presenting significant efficiency challenges. This paper presents an analysis of various Key-Value (KV) cache compression strategies, offering a comprehensive taxonomy that categorizes these methods by their underlying principles and implementation techniques. Furthermore, we evaluate their impact on performance and inference latency, providing critical insights into their effectiveness. Our findings highlight the trade-offs involved in KV cache compression and its influence on handling long-context scenarios, paving the way for more efficient LLM implementations.

Expert-Specialized Fine-Tuning (ESFT) adapts Mixture-of-Experts (MoE) large language models to enhance their task-specific performance by selectively tuning the top-activated experts for the task. Serving these fine-tuned models at scale is challenging: deploying merged models in isolation is prohibitively resource-hungry, while existing multi-adapter serving systems with LoRA-style additive updates are incompatible with ESFT's expert-oriented paradigm. We present ExpertWeave, a system that serves multiple ESFT adapters concurrently over a single shared MoE base model, drastically reducing the memory footprint and improving resource utilization. To seamlessly integrate into existing inference pipelines for MoE models with non-intrusive modifications and minimal latency overhead, ExpertWeave introduces a virtual-memory-assisted expert weight manager that co-locates base-model and adapter experts without incurring memory overhead from fragmentation, and a fused kernel for batched rerouting to enable lightweight redirection of tokens to the appropriate experts at runtime. Our evaluations show that ExpertWeave can simultaneously serve multiple adapters of a 16B MoE model on a single accelerator where the baseline runs out of memory, or provides up to 94x more KV cache capacity and achieves up to 18% higher throughput while using comparable resources, all without compromising model accuracy. ExpertWeave maintains low overhead even when scaling to 20 adapters, with a 4-11% latency increase compared with serving the base model alone. Source code will be released soon.

Large language models (LLMs) have excelled in various applications, yet serving them at scale is challenging due to their substantial resource demands and high latency. Our real-world studies reveal that over 70% of user requests to LLMs have semantically similar counterparts, suggesting the potential for knowledge transfer among requests. However, naively caching and reusing past responses leads to a big quality drop. In this paper, we introduce IC-Cache, a caching system that enables live LLM capability augmentation to improve serving efficiency: by leveraging historical request-response pairs from larger models as in-context examples, IC-Cache empowers small LLMs to imitate and even exceed the compositional abilities (e.g., reasoning) of their larger counterparts, enabling selective offloading of requests to reduce cost and latency. Achieving this live augmentation at scale introduces intricate trade-offs between response quality, latency, and system throughput. For a new request, IC-Cache efficiently selects similar, high-utility examples to prepend them to the new request's input. At scale, it adaptively routes requests across LLMs of varying capabilities, accounting for response quality and serving loads. IC-Cache employs a cost-aware cache replay mechanism that refines example quality offline to maximize online cache utility and efficiency. Evaluations on millions of realistic requests demonstrate that IC-Cache improves LLM serving throughput by 1.4-5.9x and reduces latency by 28-71% without hurting response quality.

Key-value (KV) caching is a powerful technique for accelerating large language model inference and generation. Inference workloads are large and diverse, which makes them difficult to cache effectively. Existing cache management strategies adopt the least-recently-used policy for evicting cache blocks. However, LRU leads to multiple unrelated workloads flushing each other's caches. To address this, we integrate adaptive caching that dynamically allocates cache space between recently and frequently occurring KV blocks. Evaluations show that it improves the KV cache hit rate by up to 10.8% and reduces time to first token by up to 12.6% over naive vLLM on synthetic document question answering workloads, and 2.1% and 2.0% respectively on real-world conversation workloads. The method generalizes well to batch inference and demonstrates clear interpretability while effectively accommodating diverse workloads.

Leveraging the paradigm of autoregressive diffusion, text-conditioned interactive video world models aim to simulate temporally coherent environments guided by textual instructions. While enabling low-latency, long-duration generation is pivotal for embodied AI and simulation-based planning, current frameworks primarily rely on sliding-window mechanisms to bound computational complexity. However, this approach inherently sacrifices historical context, undermining the long-range interactive capabilities. Conversely, maintaining a full-history cache remains computationally prohibitive and memory-intensive: the quadratic complexity of attention leads to excessive computational overhead, while the linear growth of the KV cache inevitably leads to GPU memory saturation. To overcome these limitations, we propose WorldAttention, a system-oriented attention architecture that achieves high efficiency through the co-design of specialized attention kernels and hierarchical KV cache management. First, we introduce Hybrid Sparse Attention (HSA), which integrates linear global attention supplemented with head-adaptive sparse attention. Additionally, we design a Hierarchical KV Cache (HKV) that organizes historical KV pairs into semantically indexed pages across multi-tier memory, enabling fine-grained retrieval and controlled GPU residency. These two designs are supported by tailored kernels to effectively translate their theoretical efficiency into real-world performance. Extensive experiments on VBench-Long and InterVBench demonstrate that WorldAttention consistently surpasses prior state-of-the-art methods, achieving subject consistency scores of 0.9472 on VBench-Long and 0.9668 on InterVBench, respectively.

Mixture-of-Experts (MoE) models achieve a favorable trade-off between performance and inference efficiency by activating only a subset of experts. However, the memory overhead of storing all experts remains a major limitation, especially in large-scale MoE models such as DeepSeek-R1(671B). In this study, we investigate domain specialization and expert redundancy in large-scale MoE models and uncover a consistent behavior we term few-shot expert localization, with only a few in-domain demonstrations, the model consistently activates a sparse and stable subset of experts on tasks within the same domain. Building on this observation, we propose a simple yet effective pruning framework, EASY-EP, that leverages a few domain-specific demonstrations to identify and retain only the most relevant experts. EASY-EP comprises two key components: output-aware expert importance assessment and expert-level token contribution estimation. The former evaluates the importance of each expert for the current token by considering the gating scores and L2 norm of the outputs of activated experts, while the latter assesses the contribution of tokens based on representation similarities before and after routed experts. Experiments on DeepSeek-R1 and DeepSeek-V3-0324 show that our method can achieve comparable performances and 2.99times throughput under the same memory budget with full model with only half the experts.

Large language models (LLMs) excel at processing long sequences, boosting demand for key-value (KV) caching. While recent efforts to evict KV cache have alleviated the inference burden, they often fail to allocate resources rationally across layers with different attention patterns. In this paper, we introduce Cascading and Adaptive KV cache Eviction (CAKE), a novel approach that frames KV cache eviction as a "cake-slicing problem." CAKE assesses layer-specific preferences by considering attention dynamics in both spatial and temporal dimensions, allocates rational cache size for layers accordingly, and manages memory constraints in a cascading manner. This approach enables a global view of cache allocation, adaptively distributing resources across diverse attention mechanisms while maintaining memory budgets. CAKE also employs a new eviction indicator that considers the shifting importance of tokens over time, addressing limitations in existing methods that overlook temporal dynamics. Comprehensive experiments on LongBench and NeedleBench show that CAKE maintains model performance with only 3.2% of the KV cache and consistently outperforms current baselines across various models and memory constraints, particularly in low-memory settings. Additionally, CAKE achieves over 10x speedup in decoding latency compared to full cache when processing contexts of 128K tokens with FlashAttention-2. Our code is available at https://github.com/antgroup/cakekv.

Large Language Models (LLMs) have gained immense success in revolutionizing various applications, including content generation, search and recommendation, and AI-assisted operation. To reduce high training costs, Mixture-of-Experts (MoE) architecture has become a popular backbone for modern LLMs. However, despite the benefits, serving MoE-based LLMs experience severe memory inefficiency due to sparsely activated experts. Recent studies propose to offload inactive experts from GPU memory to CPU memory to improve the serving efficiency of MoE models. However, they either incur high inference latency or high model memory footprints due to coarse-grained designs. To tame the latency-memory trade-off in MoE serving, we present fMoE, a fine-grained expert offloading system for MoE serving that achieves low inference latency with memory efficiency. We design fMoE to extract fine-grained expert selection patterns from MoE models and semantic hints from input prompts to efficiently guide expert prefetching, caching, and offloading decisions. fMoE is prototyped on top of HuggingFace Transformers and deployed on a six-GPU testbed. Experiments with open-source MoE models and real-world workloads show that fMoE reduces inference latency by 47% and improves expert hit rate by 36% over state-of-the-art solutions.

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.

LLM agents increasingly rely on external skills -- reusable tool specifications -- but real-world tasks often require composing multiple skills, not just selecting one. We formalize this as the Compositional Skill Routing problem: given a complex user query and a large skill library, decompose the query into atomic sub-tasks, retrieve the appropriate skill for each sub-task, and compose an executable plan. We present SkillWeaver, a decompose-retrieve-compose framework combining an LLM task decomposer, a bi-encoder skill retriever with FAISS indexing, and a dependency-aware DAG planner. To support evaluation, we introduce CompSkillBench, a benchmark of 300 compositional queries over 2,209 real MCP server skills spanning 24 functional categories, sourced from the public MCP ecosystem. Our experiments reveal that task decomposition quality is the primary bottleneck: standard LLM decomposition reaches only 34.2% category recall at the step level. To address this, we propose Iterative Skill-Aware Decomposition (SAD), a retrieval-augmented feedback loop that iteratively aligns decomposition with available skills. SAD improves decomposition accuracy from 51.0% to 67.7% (+32.7%, Wilcoxon p < 10^-6) in a single iteration; DA-conditioned analysis confirms that correct granularity is the prerequisite for effective retrieval (CatR@1 rises from 34% to 41% when DA=1). SkillWeaver reduces context window consumption by over 99%, and transfer experiments confirm generalization (+35.6% relative DA gain even when target categories are absent from the retrieval pool).

LLM agents increasingly rely on retrieval buffers to store and reuse past experience, yet the cache management policies governing these buffers remain largely ad-hoc. We formalize this as an online semantic cache replacement problem with switching costs, where items are matched by embedding similarity and hit quality is continuous rather than binary. Through experiments on two datasets from MemoryBench-Full (LoCoMo, DialSim) with 8 replacement policies, we reveal a surprising finding: classic heuristics (LRU, LFU) consistently underperform the naive FIFO baseline on semantic workloads, due to the absence of temporal locality and frequency concentration. We propose SOLAR, a learning-augmented framework that derives modification timing from regret accumulation (achieving sim17\% modification rate) and content selection from Bayesian online learning over implicit retrieval feedback. We prove SOLAR achieves a constant competitive ratio leq 3, independent of cache size and horizon (vs.\ Ω(K) for FIFO), and eviction regret O(KTlog T), matching the Ω(KT) lower bound up to logarithmic factors. Experiments demonstrate 5--75\% relative improvement over FIFO at tight cache sizes, with a clearly characterized phase transition at the working set boundary. Synthetic experiments with 5000-item pools further reveal an inverted-U relationship between pool size and retrieval quality, justifying capacity constraints as a retrieval noise phenomenon rather than a storage limitation.

Large Language Models (LLMs) like ChatGPT and Llama have revolutionized natural language processing and search engine dynamics. However, these models incur exceptionally high computational costs. For instance, GPT-3 consists of 175 billion parameters, where inference demands billions of floating-point operations. Caching is a natural solution to reduce LLM inference costs on repeated queries, which constitute about 31% of the total queries. However, existing caching methods are incapable of finding semantic similarities among LLM queries nor do they operate on contextual queries, leading to unacceptable false hit-and-miss rates. This paper introduces MeanCache, a user-centric semantic cache for LLM-based services that identifies semantically similar queries to determine cache hit or miss. Using MeanCache, the response to a user's semantically similar query can be retrieved from a local cache rather than re-querying the LLM, thus reducing costs, service provider load, and environmental impact. MeanCache leverages Federated Learning (FL) to collaboratively train a query similarity model without violating user privacy. By placing a local cache in each user's device and using FL, MeanCache reduces the latency and costs and enhances model performance, resulting in lower false hit rates. MeanCache also encodes context chains for every cached query, offering a simple yet highly effective mechanism to discern contextual query responses from standalone. Our experiments benchmarked against the state-of-the-art caching method, reveal that MeanCache attains an approximately 17% higher F-score and a 20% increase in precision during semantic cache hit-and-miss decisions while performing even better on contextual queries. It also reduces the storage requirement by 83% and accelerates semantic cache hit-and-miss decisions by 11%.

Long contexts enlarge key--value (KV) caches, increasing memory use and attention costs during generation. Compressing these caches must preserve information for unseen questions. We introduce **Latent Compaction**, which learns a compact document-specific cache through test-time training. We initialize cache entries using reconstruction importance, teacher attention, and key-space diversity. Training first fits attention outputs and log-normalizers to those of the full cache, with reconstruction-selected entries held fixed. It then jointly optimizes the entire cache through end-to-end distillation from the full-cache model. Model weights remain frozen throughout, and training uses only generic prompts about the document, without access to downstream questions or answers. Experiments on RULER, LongBench, and Phonebook Lookup demonstrate substantial gains over budget-matched compression baselines. On a RULER, for example, our method achieves 96.7% accuracy at 32times compression, compared with 40.0% for KVZip+. The method also transfers across model scales and families, with retrieval gains on both Qwen3 and MiniCPM5. These results show that treating the KV cache as a neural network and optimizing it with test-time training can preserve retrieval accuracy under aggressive compression. The code is released at https://www.github.com/THUNLP/latent-compaction

The linear growth of Key-Value (KV) cache remains a bottleneck for multi-turn LLM deployment. Existing KV cache compression methods often fail to account for the structural properties of multi-turn dialogues, relying on heuristic eviction that risks losing critical context. We propose SONIC, a learning-based framework that compresses historical segments into compact and semantically rich Nexus tokens. By integrating dynamic budget training, SONIC allows flexible adaptation to varying memory constraints without retraining. Experiments show that at compression ratios of 80\% and 50\%, SONIC consistently outperforms baselines such as H2O and StreamingLLM on four diverse multi-turn benchmarks. Specifically, on the widely used MTBench101 benchmark, SONIC achieves an average score improvement of 35.55\% over state-of-the-art baselines, validating its effectiveness in sustaining coherent multi-turn dialogues. Furthermore, SONIC enhances deployment efficiency, accelerating the overall inference process by 50.1\% compared to full-context generation.

Large Language Model (LLM) agents can handle increasingly complex tasks, navigating long trajectories of observations and actions. As these trajectories grow, key-value (KV) cache I/O and storage become the dominant cost. Trajectories can also exceed the LLM's maximum context length, requiring control over active cache growth. Prior work uses context engineering to reduce cache size by trimming or summarizing history. However, modifying the trajectory prefix invalidates subsequent cached states and triggers costly re-prefilling of retained context. Task accuracy can also suffer when too little context is retained to limit re-prefill costs. We propose cache engineering as a new scheme. It allows agents to directly control their active KV cache without modifying the trajectory or re-prefilling retained context. Specifically, agents can proactively use drop to exclude selected cache from future attention and repos to adjust retained keys' position encodings. We formalize these cache operations and identify history dependence as a serving challenge: identical token sequences can yield different cache states under different cache operation sequences, making prefix matching insufficient for correct cache reuse. Thus, we co-design the AKV serving engine. It provides a stateless interface compatible with existing serving APIs, while internally managing stateful cache reuse, eviction, and restoration. Experiments show that a tool-response rolling-drop cache policy improves average benchmark scores by 1.7 points and reduces average cost by 27.7%, compared with conventional threshold-triggered summarization. With AKV, it achieves up to 1.2x the output throughput and a 1.5x speedup in P95 request latency. Recursive self-improvement of cache policies further reduces cost by 7.1% over expert design. Our code is at https://github.com/thu-nics/AKV.

Efficient KV cache management in LLMs is crucial for long-context tasks like RAG and summarization. Existing KV cache compression methods enforce a fixed pattern, neglecting task-specific characteristics and reducing the retention of essential information. However, we observe distinct activation patterns across layers in various tasks, highlighting the need for adaptive strategies tailored to each task's unique demands. Based on this insight, we propose DynamicKV, a method that dynamically optimizes token retention by adjusting the number of tokens retained at each layer to adapt to the specific task. DynamicKV establishes global and per-layer maximum KV cache budgets, temporarily retaining the maximum budget for the current layer, and periodically updating the KV cache sizes of all preceding layers during inference. Our method retains only 1.7% of the KV cache size while achieving ~85% of the Full KV cache performance on LongBench. Notably, even under extreme compression (0.9%), DynamicKV surpasses state-of-the-art (SOTA) methods by 11% in the Needle-in-a-Haystack test using Mistral-7B-Instruct-v0.2. The code will be released.

Large language models increasingly operate over long contexts, where the KV cache becomes a dominant memory bottleneck: its size grows linearly with sequence length and must be retained throughout decoding, making full GPU caching prohibitively expensive without compression. Existing KV cache compression methods struggle to balance efficiency with faithful context preservation. Token eviction discards information, while semantic grouping fixes compression decisions at prefill time; neither can recover token-level detail from a compressed span once it becomes relevant during generation. As a solution, we propose SeKV, a resolution-adaptive semantic KV cache that organizes context into entropy-guided semantic spans and stores them across a GPU-CPU memory hierarchy without discarding information. Each span keeps a lightweight summary vector on GPU for coarse routing and a low-rank SVD basis on CPU for on-demand token-level reconstruction. A trained zoom-in mechanism selectively expands query-relevant spans during decoding, enabling precise retrieval without materializing the full KV cache on GPU. SeKV enables adaptive token-level reconstruction while keeping the base LLM fully frozen and adding fewer than 0.05% trainable parameters. Across four benchmarks, SeKV improves over the strongest semantic compression baseline by 5.9% on average while reducing GPU memory by 53.3% versus full KV caching at 128K context. Code is available on https://github.com/AmirAbaskohi/SeKV.

Mixture-of-Experts (MoE) architectures in large language models (LLMs) deliver exceptional performance and reduced inference costs compared to dense LLMs. However, their large parameter counts result in prohibitive memory requirements, limiting practical deployment. While existing pruning methods primarily focus on expert-level pruning, this coarse granularity often leads to substantial accuracy degradation. In this work, we introduce HEAPr, a novel pruning algorithm that decomposes experts into smaller, indivisible atomic experts, enabling more precise and flexible atomic expert pruning. To measure the importance of each atomic expert, we leverage second-order information based on principles similar to Optimal Brain Surgeon (OBS) theory. To address the computational and storage challenges posed by second-order information, HEAPr exploits the inherent properties of atomic experts to transform the second-order information from expert parameters into that of atomic expert parameters, and further simplifies it to the second-order information of atomic expert outputs. This approach reduces the space complexity from O(d^4), where d is the model's dimensionality, to O(d^2). HEAPr requires only two forward passes and one backward pass on a small calibration set to compute the importance of atomic experts. Extensive experiments on MoE models, including DeepSeek MoE and Qwen MoE family, demonstrate that HEAPr outperforms existing expert-level pruning methods across a wide range of compression ratios and benchmarks. Specifically, HEAPr achieves nearly lossless compression at compression ratios of 20% ~ 25% in most models, while also reducing FLOPs nearly by 20%. The code can be found at https://github.com/LLIKKE/HEAPr{https://github.com/LLIKKE/HEAPr}.

Existing key-value (KV) cache compression methods typically rely on heuristics, such as uniform cache allocation across layers or static eviction policies, however, they ignore the critical interplays among layer-specific feature patterns and task performance, which can lead to degraded generalization. In this paper, we propose EvolKV, an adaptive framework for layer-wise, task-driven KV cache compression that jointly optimizes the memory efficiency and task performance. By reformulating cache allocation as a multi-objective optimization problem, EvolKV leverages evolutionary search to dynamically configure layer budgets while directly maximizing downstream performance. Extensive experiments on 11 tasks demonstrate that our approach outperforms all baseline methods across a wide range of KV cache budgets on long-context tasks and surpasses heuristic baselines by up to 7 percentage points on GSM8K. Notably, EvolKV achieves superior performance over the full KV cache setting on code completion while utilizing only 1.5% of the original budget, suggesting the untapped potential in learned compression strategies for KV cache budget allocation.

Key-value (KV) caching is widely used to accelerate transformer inference, but its memory cost grows linearly with input length, limiting long-context deployment. Existing token eviction methods reduce memory by discarding less important tokens, which can be viewed as a coarse form of dimensionality reduction that assigns each token either zero or full dimension. We propose MixedDimKV, a mixed-dimension KV cache compression method that allocates dimensions to tokens at a more granular level, and MixedDimKV-H, which further integrates head-level importance information. Experiments on long-context benchmarks show that MixedDimKV outperforms prior KV cache compression methods that do not rely on head-level importance profiling. When equipped with the same head-level importance information, MixedDimKV-H consistently outperforms HeadKV. Notably, our approach achieves comparable performance to full attention on LongBench with only 6.25% of the KV cache. Furthermore, in the Needle-in-a-Haystack test, our solution maintains 100% accuracy at a 50K context length while using as little as 0.26% of the cache.

Large Language Models (LLMs) have been widely deployed in a variety of applications, and the context length is rapidly increasing to handle tasks such as long-document QA and complex logical reasoning. However, long context poses significant challenges for inference efficiency, including high memory costs of key-value (KV) cache and increased latency due to extensive memory accesses. Recent works have proposed compressing KV cache to approximate computation, but these methods either evict tokens permanently, never recalling them for later inference, or recall previous tokens at the granularity of pages divided by textual positions. Both approaches degrade the model accuracy and output quality. To achieve efficient and accurate recallable KV cache compression, we introduce ClusterKV, which recalls tokens at the granularity of semantic clusters. We design and implement efficient algorithms and systems for clustering, selection, indexing and caching. Experiment results show that ClusterKV attains negligible accuracy loss across various tasks with 32k context lengths, using only a 1k to 2k KV cache budget, and achieves up to a 2times speedup in latency and a 2.5times improvement in decoding throughput. Compared to SoTA recallable KV compression methods, ClusterKV demonstrates higher model accuracy and output quality, while maintaining or exceeding inference efficiency.

Mixture-of-Experts (MoE) models have shown remarkable capability in instruction tuning, especially when the number of tasks scales. However, previous methods simply merge all training tasks (e.g. creative writing, coding, and mathematics) and apply fixed sampling weights, without considering the importance of different tasks as the model training state changes. In this way, the most helpful data cannot be effectively distinguished, leading to suboptimal model performance. To reduce the potential redundancies of datasets, we make the first attempt and propose a novel dynamic data mixture for MoE instruction tuning. Specifically, inspired by MoE's token routing preference, we build dataset-level representations and then capture the subtle differences among datasets. Finally, we propose to dynamically adjust the sampling weight of datasets by their inter-redundancies, thus maximizing global performance under a limited training budget. The experimental results on two MoE models demonstrate the effectiveness of our approach on both downstream knowledge \& reasoning tasks and open-ended queries. Code and models are available at https://github.com/Spico197/MoE-SFT .

In prefill-decode (PD) disaggregated LLM serving, each request is assigned to a decode worker after prefill. Existing decode routers balance only load; for mixture-of-experts (MoE) models this is incomplete: equally loaded workers can differ in latency, since each decode step loads the weights of every distinct expert its batch activates. We present ELDR, an expert-locality-aware decode router for PD-disaggregated MoE serving. From a request's prefill expert activations, ELDR builds an expert signature predicting the experts it will activate during generation. Offline, balanced K-means partitions signature space across decode workers; online, locality-band routing sends each request to the least-loaded worker among those best matching its signature. A signature cache, co-indexed with the KV cache at KV-block granularity, keeps signatures exact under prefix caching. Implemented in vLLM and evaluated on deployments of up to 40 GPUs, ELDR reduces median TPOT by 5.9-13.9% over the strongest of four load-balancing baselines across three MoE models and two workloads, with model outputs unchanged.

Long-running LLM applications repeatedly send growing context, making prefix caching critical for reducing prefill cost. Yet prefix-cache behavior under agentic workloads remains poorly understood. We study production traces from two companies and evaluate 14 eviction algorithms across HBM-constrained and large memory-pool settings. Despite a large gap to Belady, sophisticated policies designed for traditional caches provide little benefit over LRU. The reason is structural: prefix reuse is dominated by the regular pacing of active sessions, making recency unusually predictive. Prefix caching nevertheless introduces new challenges, including heavy-tailed session footprints and highly variable miss costs as attention computation grows with sequence length. We introduce the compute-savings ratio and two offline oracles to quantify these effects. Our results show that effective prefix-cache management should retain recency as its foundation while selectively adding quick demotion for one-hit prefixes, compute-aware partial eviction for expensive misses, and capacity-dependent eviction granularity. We will release the traces and simulator to support future research.

Large language models (LLMs) with extended context windows have become increasingly prevalent for tackling complex tasks. However, the substantial Key-Value (KV) cache required for long-context LLMs poses significant deployment challenges. Existing approaches either discard potentially critical information needed for future generations or offer limited efficiency gains due to high computational overhead. In this paper, we introduce Chelsea, a simple yet effective framework for online KV cache clustering. Our approach is based on the observation that key states exhibit high similarity along the sequence dimension. To enable efficient clustering, we divide the sequence into chunks and propose Chunked Soft Matching, which employs an alternating partition strategy within each chunk and identifies clusters based on similarity. Chelsea then merges the KV cache within each cluster into a single centroid. Additionally, we provide a theoretical analysis of the computational complexity and the optimality of the intra-chunk partitioning strategy. Extensive experiments across various models and long-context benchmarks demonstrate that Chelsea achieves up to 80% reduction in KV cache memory usage while maintaining comparable model performance. Moreover, with minimal computational overhead, Chelsea accelerates the decoding stage of inference by up to 3.19times and reduces end-to-end latency by up to 2.72times.

Large Language Models (LLMs) have achieved remarkable advancements, but their monolithic nature presents challenges in terms of scalability, cost, and customization. This paper introduces the Composition of Experts (CoE), a modular compound AI system leveraging multiple expert LLMs. CoE leverages a router to dynamically select the most appropriate expert for a given input, enabling efficient utilization of resources and improved performance. We formulate the general problem of training a CoE and discuss inherent complexities associated with it. We propose a two-step routing approach to address these complexities that first uses a router to classify the input into distinct categories followed by a category-to-expert mapping to obtain desired experts. CoE offers a flexible and cost-effective solution to build compound AI systems. Our empirical evaluation demonstrates the effectiveness of CoE in achieving superior performance with reduced computational overhead. Given that CoE comprises of many expert LLMs it has unique system requirements for cost-effective serving. We present an efficient implementation of CoE leveraging SambaNova SN40L RDUs unique three-tiered memory architecture. CoEs obtained using open weight LLMs Qwen/Qwen2-7B-Instruct, google/gemma-2-9b-it, google/gemma-2-27b-it, meta-llama/Llama-3.1-70B-Instruct and Qwen/Qwen2-72B-Instruct achieve a score of 59.4 with merely 31 billion average active parameters on Arena-Hard and a score of 9.06 with 54 billion average active parameters on MT-Bench.

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.

Recent optimization studies on Retrieval-Augmented Generation (RAG) have exploited chunk-level KV cache reuse to avoid processing long retrieved contexts for higher efficiency, while significant information redundancy and noise still remain in the coarse-grained chunks. This paper optimizes the Pareto frontier under low prefill latency constraints while maximizing accuracy by proposing CoinRAG (Contextualized Information Nugget KV Cache Reuse for Long-Context RAG). The name metaphorically reflects our core mechanism: much like assembling small tokens (or "coins") to accumulate a larger value, CoinRAG compositionally reuses offline-computed, fine-grained nugget caches to form a learned contextual representation efficiently in a more semantically relevant but compact manner. Specifically, instead of full-chunk encoding, CoinRAG identifies query-relevant semantic units within retrieved chunks through two-stage retrieval and seamlessly assembles their sliced KV representations with a chunk-level context. Extensive evaluations on LongBench multi-hop question answering tasks demonstrate that CoinRAG significantly reduces operational costs and outperforms the other baselines with a new Pareto frontier and an average 5.3% relative improvement in answer quality (F1) under a standard fast prefill latency budget.

Diffusion-based large language models (dLLMs), despite their promising performance, still suffer from inferior inference efficiency. This is because dLLMs rely on bidirectional attention and cannot directly benefit from the standard key-value (KV) cache as autoregressive models (ARMs) do. To tackle this issue, we introduce Dual aDaptive Cache (d^2Cache), which is a training-free approximate KV cache framework for accelerating dLLM inference. d^2Cache features a two-stage fine-grained selection strategy to identify tokens and adaptively update their KV states at each decoding step, while caching the KV states of the remaining tokens for reuse. Furthermore, d^2Cache naturally offers a more reliable decoding alternative, which can enable quasi left-to-right generation and mitigate premature overconfidence in tokens at the end of the sequence. Extensive experimental results on two representative dLLMs (\ie, LLaDA and Dream) demonstrate that d^2Cache not only achieves substantial inference speedups, but also yields consistent improvements in generation quality. The code is available at https://github.com/Kamichanw/d2Cache.

Modern Mixture-of-Experts (MoE) architectures allocate expert capacity through a rigid per-layer rule: each transformer layer owns a separate expert set. This convention couples depth scaling with linear expert-parameter growth and assumes that every layer needs isolated expert capacity. However, recent analyses and our routing probe challenge this allocation rule: replacing a deeper layer's learned top-k router with uniform random routing drops downstream accuracy by only 1.0-1.6 points across multiple production MoE models. Motivated by this redundancy, we propose UniPool, an MoE architecture that treats expert capacity as a global architectural budget by replacing per-layer expert ownership with a single shared pool accessed by independent per-layer routers. To enable stable and balanced training under sharing, we introduce a pool-level auxiliary loss that balances expert utilization across the entire pool, and adopt NormRouter to provide sparse and scale-stable routing into the shared expert pool. Across five LLaMA-architecture model scales (182M, 469M, 650M, 830M, and 978M parameters) trained on 30B tokens from the Pile, UniPool consistently improves validation loss and perplexity over the matched vanilla MoE baselines. Across these scales, UniPool reduces validation loss by up to 0.0386 relative to vanilla MoE. Beyond raw loss improvement, our results identify pool size as an explicit depth-scaling hyperparameter: reduced-pool UniPool variants using only 41.6%-66.7% of the vanilla expert-parameter budget match or outperform layer-wise MoE at the tested scales. This shows that, under a shared-pool design, expert parameters need not grow linearly with depth; they can grow sublinearly while remaining more efficient and effective than vanilla MoE. Further analysis shows that UniPool's benefits compose with finer-grained expert decomposition.

Sparse Mixture-of-Experts (MoE) models offer a powerful way to scale model size without increasing compute, as per-token FLOPs depend only on k active experts rather than the total pool of E experts. Yet, this asymmetry creates an MoE efficiency paradox in practice: adding more experts balloons memory and communication costs, making actual training inefficient. We argue that this bottleneck arises in part because current MoE training allocates too many experts from the beginning, even though early-stage data may not fully utilize such capacity. Motivated by this, we propose EMO, a simple progressive training framework that treats MoE capacity as expandable memory and grows the expert pool over the course of training. EMO explicitly models sparsity in scaling law to derive stage-wise compute-optimal token budgets for progressive expansion. Empirical results show that EMO matches the performance of a fixed-expert setup in large-scale experiments while improving wall-clock efficiency. It offers a surprisingly simple yet effective path to scalable MoE training, preserving the benefits of large expert pools while reducing both training time and GPU cost.

Recent advances in Large Language Model (LLM)-based agents have been propelled by Retrieval-Augmented Generation (RAG), which grants the models access to vast external knowledge bases. Despite RAG's success in improving agent performance, agent-level cache management, particularly constructing, maintaining, and updating a compact, relevant corpus dynamically tailored to each agent's need, remains underexplored. Therefore, we introduce ARC (Agent RAG Cache Mechanism), a novel, annotation-free caching framework that dynamically manages small, high-value corpora for each agent. By synthesizing historical query distribution patterns with the intrinsic geometry of cached items in the embedding space, ARC automatically maintains a high-relevance cache. With comprehensive experiments on three retrieval datasets, our experimental results demonstrate that ARC reduces storage requirements to 0.015% of the original corpus while offering up to 79.8% has-answer rate and reducing average retrieval latency by 80%. Our results demonstrate that ARC can drastically enhance efficiency and effectiveness in RAG-powered LLM agents.

Similarity search is a key operation in multimedia retrieval systems and recommender systems, and it will play an important role also for future machine learning and augmented reality applications. When these systems need to serve large objects with tight delay constraints, edge servers close to the end-user can operate as similarity caches to speed up the retrieval. In this paper we present AÇAI, a new similarity caching policy which improves on the state of the art by using (i) an (approximate) index for the whole catalog to decide which objects to serve locally and which to retrieve from the remote server, and (ii) a mirror ascent algorithm to update the set of local objects with strong guarantees even when the request process does not exhibit any statistical regularity.

Mixture-of-experts (MoE) architectures enable scaling large language models (LLMs) to vast parameter counts without a proportional rise in computational costs. However, the significant memory demands of large MoE models hinder their deployment across various computational environments, from cloud servers to consumer devices. This study first demonstrates pronounced task-specific specialization in expert activation patterns within MoE layers. Building on this, we introduce PreMoe, a novel framework that enables efficient deployment of massive MoE models in memory-constrained environments. PreMoe features two main components: probabilistic expert pruning (PEP) and task-adaptive expert retrieval (TAER). PEP employs a new metric, the task-conditioned expected selection score (TCESS), derived from router logits to quantify expert importance for specific tasks, thereby identifying a minimal set of critical experts. TAER leverages these task-specific expert importance profiles for efficient inference. It pre-computes and stores compact expert patterns for diverse tasks. When a user query is received, TAER rapidly identifies the most relevant stored task pattern and reconstructs the model by loading only the small subset of experts crucial for that task. This approach dramatically reduces the memory footprint across all deployment scenarios. DeepSeek-R1 671B maintains 97.2\% accuracy on MATH500 when pruned to 8/128 configuration (50\% expert reduction), and still achieves 72.0\% with aggressive 8/32 pruning (87.5\% expert reduction). Pangu-Ultra-MoE 718B achieves 97.15\% on MATH500 and 81.3\% on AIME24 with 8/128 pruning, while even more aggressive pruning to 4/64 (390GB memory) preserves 96.95\% accuracy on MATH500. We make our code publicly available at https://github.com/JarvisPei/PreMoe.

Pretraining large language models (LLMs) typically requires centralized clusters with thousands of high-memory GPUs (e.g., H100/A100). Recent decentralized training methods reduce communication overhead by employing federated optimization; however, they still need to train the entire model on each node, remaining constrained by GPU memory limitations. In this work, we propose SParse Expert Synchronization (SPES), a memory-efficient decentralized framework for pretraining mixture-of-experts (MoE) LLMs. SPES trains only a subset of experts per node, substantially lowering the memory footprint. Each node updates its local experts and periodically synchronizes with other nodes, eliminating full-parameter transmission while ensuring efficient knowledge sharing. To accelerate convergence, we introduce an expert-merging warm-up strategy, where experts exchange knowledge early in training, to rapidly establish foundational capabilities. With SPES, we train a 2B-parameter MoE LLM using 16 standalone 48GB GPUs over internet connections, which achieves competitive performance with centrally trained LLMs under similar computational budgets. We further demonstrate scalability by training a 7B model from scratch and a 9B model upcycled from a dense checkpoint, both of which match prior centralized baselines. Our code is available at https://github.com/zjr2000/SPES.

Large Language Models (LLMs), despite their recent impressive accomplishments, are notably cost-prohibitive to deploy, particularly for applications involving long-content generation, such as dialogue systems and story writing. Often, a large amount of transient state information, referred to as the KV cache, is stored in GPU memory in addition to model parameters, scaling linearly with the sequence length and batch size. In this paper, we introduce a novel approach for implementing the KV cache which significantly reduces its memory footprint. Our approach is based on the noteworthy observation that a small portion of tokens contributes most of the value when computing attention scores. We call these tokens Heavy Hitters (H_2). Through a comprehensive investigation, we find that (i) the emergence of H_2 is natural and strongly correlates with the frequent co-occurrence of tokens in the text, and (ii) removing them results in significant performance degradation. Based on these insights, we propose Heavy Hitter Oracle (H_2O), a KV cache eviction policy that dynamically retains a balance of recent and H_2 tokens. We formulate the KV cache eviction as a dynamic submodular problem and prove (under mild assumptions) a theoretical guarantee for our novel eviction algorithm which could help guide future work. We validate the accuracy of our algorithm with OPT, LLaMA, and GPT-NeoX across a wide range of tasks. Our implementation of H_2O with 20% heavy hitters improves the throughput over three leading inference systems DeepSpeed Zero-Inference, Hugging Face Accelerate, and FlexGen by up to 29times, 29times, and 3times on OPT-6.7B and OPT-30B. With the same batch size, H2O can reduce the latency by up to 1.9times. The code is available at https://github.com/FMInference/H2O.

We propose ListT5, a novel reranking approach based on Fusion-in-Decoder (FiD) that handles multiple candidate passages at both train and inference time. We also introduce an efficient inference framework for listwise ranking based on m-ary tournament sort with output caching. We evaluate and compare our model on the BEIR benchmark for zero-shot retrieval task, demonstrating that ListT5 (1) outperforms the state-of-the-art RankT5 baseline with a notable +1.3 gain in the average NDCG@10 score, (2) has an efficiency comparable to pointwise ranking models and surpasses the efficiency of previous listwise ranking models, and (3) overcomes the lost-in-the-middle problem of previous listwise rerankers. Our code, model checkpoints, and the evaluation framework are fully open-sourced at https://github.com/soyoung97/ListT5.

Large Reasoning Models (LRMs) achieve superior problem-solving through extended Chain-of-Thought (CoT) generation, but the resulting key-value (KV) cache grows linearly with sequence length and creates severe memory bottlenecks, often exceeding GPU capacity for long reasoning traces. Existing KV cache compression methods rely on recent queries to estimate future token importance, implicitly assuming these serve as reliable proxies for future attention patterns. We demonstrate that this assumption fails in long-horizon reasoning: certain decoding steps generate Thought Revisiting Tokens (TRT) that re-attend to distant previous context, such as task-solving plans formulated early in the trace. Through systematic analysis, we discover that queries corresponding to the TRT cluster into a small number of similarity groups in the embedding space. Based on this insight, we propose BeaconKV, a training-free KV cache compression method that maintains beacon queries, compact representatives for each global query cluster, to anticipate which KV pairs will be revisited without storing the entire query history. Across four open-source LRMs and diverse reasoning benchmarks, BeaconKV generally outperforms existing compression methods, achieving up to 5.8times memory reduction while nearly preserving full cache accuracy and improving throughput by over 4.3times.

Large Language Models (LLMs) have revolutionized the field of natural language processing, achieving unprecedented performance across a variety of applications by leveraging increased model sizes and sequence lengths. However, the associated rise in computational and memory costs poses significant challenges, particularly in managing long sequences due to the quadratic complexity of the transformer attention mechanism. This paper focuses on the long-context scenario, addressing the inefficiencies in KV cache memory consumption during inference. Unlike existing approaches that optimize the memory based on the sequence lengths, we uncover that the channel dimension of the KV cache exhibits significant redundancy, characterized by unbalanced magnitude distribution and low-rank structure in attention weights. Based on these observations, we propose ThinK, a novel query-dependent KV cache pruning method designed to minimize attention weight loss while selectively pruning the least significant channels. Our approach not only maintains or enhances model accuracy but also achieves a reduction in memory costs by over 20% compared with vanilla KV cache eviction methods. Extensive evaluations on the LLaMA3 and Mistral models across various long-sequence datasets confirm the efficacy of ThinK, setting a new precedent for efficient LLM deployment without compromising performance. We also outline the potential of extending our method to value cache pruning, demonstrating ThinK's versatility and broad applicability in reducing both memory and computational overheads.

Retrieval Augmented Generation faces a trade-off: concatenating documents in a long prompt enables multi-document reasoning but creates prefill bottlenecks, while encoding document KV caches separately offers speed but breaks cross-document interaction. We propose Parallel Context-of-Experts Decoding (Pced), a training-free framework that shifts evidence aggregation from the attention mechanism to the decoding. Pced treats retrieved documents as isolated "experts", synchronizing their predictions via a novel retrieval-aware contrastive decoding rule that weighs expert logits against the model prior. This approach recovers cross-document reasoning capabilities without constructing a shared attention across documents.

Large Language Models (LLMs) exhibit enhanced reasoning capabilities by employing Chain-of-Thought (CoT). However, the extended reasoning sequences introduce significant GPU memory overhead due to increased key-value (KV) cache size, particularly in tasks requiring long reasoning sequences, such as mathematics and programming. Existing KV cache compression methods mitigate memory bottlenecks but struggle in long reasoning tasks. In this paper, we analyze attention patterns in reasoning tasks and reveal a Token Importance Recurrence phenomenon: a large proportion of tokens receive renewed attention after multiple decoding steps, which is failed to capture by existing works and may lead to unpredictable eviction on such periodically critical tokens. To address this, we propose LazyEviction, a lagged KV eviction framework designed to maintain reasoning performance while reducing KV memory. LazyEviction is an Observation Window-based Lagged Eviction Mechanism retaining latent recurring tokens by performing lagged evictions across decoding steps, which contains two key components: (1) Recurrence Interval Tracking for capturing temporal variations in token importance, and (2) an Maximum Recurrence Interval-Centric Eviction Policy that prioritizes eviction based on tokens' recurrence patterns. Extensive experiments demonstrate that LazyEviction reduces KV cache size by 50% while maintaining comparable accuracy on mathematics reasoning datasets, outperforming state-of-the-art methods. Our findings highlight the importance of preserving recurring tokens, which are critical for maintaining knowledge continuity in multi-step reasoning tasks.

Streaming video understanding requires processing unbounded video streams with limited memory and computation, posing two key challenges. First, continuously constructing new and evicting old key-value(KV) caches is required for unbounded streams. Secondly, due to the high cost of collecting and training on unbounded streams, models must learn from short sequences while generalizing to long streams. Existing streaming VideoVLLMs fail to scale to unbounded video streams or focus on cache reuse strategies, leaving the impact of cache construction underexplored. In this paper, we propose Decoupled Streaming Cache(DSCache), a training-free cache construction mechanism that adapts pretrained offline models to streaming settings. DSCache maintains a cumulative past KV cache while constructing a separate instant cache on-demand, decoupled from past caches to preserve the informativeness of recent inputs. To enable position extrapolation beyond the training length, DSCache further incorporates a position-agnostic encoding strategy, ensuring KV caches to support unseen positions and preventing position overflow. Experiments on Streaming Video QA benchmarks demonstrate DSCache's state-of-the-art performance, with an average 2.5% accuracy gains over prior methods.

We present LeMiCa, a training-free and efficient acceleration framework for diffusion-based video generation. While existing caching strategies primarily focus on reducing local heuristic errors, they often overlook the accumulation of global errors, leading to noticeable content degradation between accelerated and original videos. To address this issue, we formulate cache scheduling as a directed graph with error-weighted edges and introduce a Lexicographic Minimax Path Optimization strategy that explicitly bounds the worst-case path error. This approach substantially improves the consistency of global content and style across generated frames. Extensive experiments on multiple text-to-video benchmarks demonstrate that LeMiCa delivers dual improvements in both inference speed and generation quality. Notably, our method achieves a 2.9x speedup on the Latte model and reaches an LPIPS score of 0.05 on Open-Sora, outperforming prior caching techniques. Importantly, these gains come with minimal perceptual quality degradation, making LeMiCa a robust and generalizable paradigm for accelerating diffusion-based video generation. We believe this approach can serve as a strong foundation for future research on efficient and reliable video synthesis. Our code is available at :https://github.com/UnicomAI/LeMiCa

Personal AI agents incur substantial cost via repeated LLM calls. We show existing caching methods fail: GPTCache achieves 37.9% accuracy on real benchmarks; APC achieves 0-12%. The root cause is optimizing for the wrong property -- cache effectiveness requires key consistency and precision, not classification accuracy. We observe cache-key evaluation reduces to clustering evaluation and apply V-measure decomposition to separate these on n=8,682 points across MASSIVE, BANKING77, CLINC150, and NyayaBench v2, our new 8,514-entry multilingual agentic dataset (528 intents, 20 W5H2 classes, 63 languages). We introduce W5H2, a structured intent decomposition framework. Using SetFit with 8 examples per class, W5H2 achieves 91.1%+/-1.7% on MASSIVE in ~2ms -- vs 37.9% for GPTCache and 68.8% for a 20B-parameter LLM at 3,447ms. On NyayaBench v2 (20 classes), SetFit achieves 55.3%, with cross-lingual transfer across 30 languages. Our five-tier cascade handles 85% of interactions locally, projecting 97.5% cost reduction. We provide risk-controlled selective prediction guarantees via RCPS with nine bound families.

Mixture of Experts (MoE) models have emerged as the de facto architecture for scaling up language models without significantly increasing the computational cost. Recent MoE models demonstrate a clear trend towards high expert granularity (smaller expert intermediate dimension) and higher sparsity (constant number of activated experts with higher number of total experts), which improve model quality per FLOP. However, fine-grained MoEs suffer from increased activation memory footprint and reduced hardware efficiency due to higher IO costs, while sparser MoEs suffer from wasted computations due to padding in Grouped GEMM kernels. In response, we propose a memory-efficient algorithm to compute the forward and backward passes of MoEs with minimal activation caching for the backward pass. We also design GPU kernels that overlap memory IO with computation benefiting all MoE architectures. Finally, we propose a novel "token rounding" method that minimizes the wasted compute due to padding in Grouped GEMM kernels. As a result, our method SonicMoE reduces activation memory by 45% and achieves a 1.86x compute throughput improvement on Hopper GPUs compared to ScatterMoE's BF16 MoE kernel for a fine-grained 7B MoE. Concretely, SonicMoE on 64 H100s achieves a training throughput of 213 billion tokens per day comparable to ScatterMoE's 225 billion tokens per day on 96 H100s for a 7B MoE model training with FSDP-2 using the lm-engine codebase. Under high MoE sparsity settings, our tile-aware token rounding algorithm yields an additional 1.16x speedup on kernel execution time compared to vanilla top-K routing while maintaining similar downstream performance. We open-source all our kernels to enable faster MoE model training.

Information retrieval methods often rely on a single embedding model trained on large, general-domain datasets like MSMARCO. While this approach can produce a retriever with reasonable overall performance, models trained on domain-specific data often yield better results within their respective domains. While prior work in information retrieval has tackled this through multi-task training, the topic of combining multiple domain-specific expert retrievers remains unexplored, despite its popularity in language model generation. In this work, we introduce RouterRetriever, a retrieval model that leverages multiple domain-specific experts along with a routing mechanism to select the most appropriate expert for each query. It is lightweight and allows easy addition or removal of experts without additional training. Evaluation on the BEIR benchmark demonstrates that RouterRetriever outperforms both MSMARCO-trained (+2.1 absolute nDCG@10) and multi-task trained (+3.2) models. This is achieved by employing our routing mechanism, which surpasses other routing techniques (+1.8 on average) commonly used in language modeling. Furthermore, the benefit generalizes well to other datasets, even in the absence of a specific expert on the dataset. To our knowledge, RouterRetriever is the first work to demonstrate the advantages of using multiple domain-specific expert embedding models with effective routing over a single, general-purpose embedding model in retrieval tasks.

Mixture-of-Experts (MoE) models enable scalable performance but face severe memory constraints on edge devices. Existing offloading strategies struggle with I/O bottlenecks due to the dynamic, low-information nature of autoregressive expert activation. In this paper, we propose to repurpose Speculative Decoding (SD) not merely as a compute accelerator, but as an informative lookahead sensor for memory management, supported by our theoretical and empirical analyses. Hence, we introduce MoE-SpAc, an MoE inference framework that integrates a Speculative Utility Estimator to track expert demand, a Heterogeneous Workload Balancer to dynamically partition computation via online integer optimization, and an Asynchronous Execution Engine to unify the prefetching and eviction in the same utility space. Extensive experiments on seven benchmarks demonstrate that MoE-SpAc achieves a 42% improvement in TPS over the SOTA SD-based baseline, and an average 4.04x speedup over all standard baselines. Code is available at https://github.com/lshAlgorithm/MoE-SpAc .

Recently, large vision-language models (LVLMs) have rapidly gained popularity for their strong generation and reasoning capabilities given diverse multimodal inputs. However, these models incur significant computational and memory overhead during inference, which greatly hinders the efficient deployment in practical scenarios. The extensive key-value (KV) cache, necessitated by the lengthy input and output sequences, notably contributes to the high inference cost. Based on this, recent works have investigated ways to reduce the KV cache size for higher efficiency. Although effective, they generally overlook the distinct importance distributions of KV vectors across layers and maintain the same cache size for each layer during the next token prediction. This results in the significant contextual information loss for certain layers, leading to notable performance decline. To address this, we present PrefixKV. It reframes the challenge of determining KV cache sizes for all layers into the task of searching for the optimal global prefix configuration. With an adaptive layer-wise KV retention recipe based on binary search, the maximum contextual information can thus be preserved in each layer, facilitating the generation. Extensive experiments demonstrate that our method achieves the state-of-the-art performance compared with others. It exhibits superior inference efficiency and generation quality trade-offs, showing promising potential for practical applications. Code is available at https://github.com/THU-MIG/PrefixKV.

Sparse Mixture-of-Experts (MoE) models scale capacity by routing each token to a small subset of experts. However, their routers exhibit a fundamental trade-off: strong load balancing can suppress expert specialization, while aggressive diversity often causes routing collapse. We propose Hi-MoE, a grouped MoE framework that decomposes routing control into two coupled levels: (i) inter-group balancing that enforces fair traffic across expert groups, and (ii) intra-group specialization that promotes complementary expert behaviors while preventing within-group collapse. Our analysis provides a principled explanation of how our hierarchical objectives reshape the router, thereby promoting stable specialization and mitigating collapse. We observe consistent improvements over recent sparse-routing and grouped-MoE baselines across NLP and vision benchmarks, and confirm robustness via scaling studies (model size, expert count) and targeted ablations. In large-scale pre-training on 58B tokens, Hi-MoE-7B achieves a 5.6% perplexity reduction and a 40% improvement in expert balance over OLMoE-7B across diverse evaluation domains.

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}.

Transformer-based Large Language Models rely critically on the KV cache to efficiently handle extended contexts during the decode phase. Yet, the size of the KV cache grows proportionally with the input length, burdening both memory bandwidth and capacity as decoding progresses. To address this challenge, we present RocketKV, a training-free KV cache compression strategy containing two consecutive stages. In the first stage, it performs coarse-grain permanent KV cache eviction on the input sequence tokens. In the second stage, it adopts a hybrid sparse attention method to conduct fine-grain top-k sparse attention, approximating the attention scores by leveraging both head and sequence dimensionality reductions. We show that RocketKV provides a compression ratio of up to 400times, end-to-end speedup of up to 3.7times as well as peak memory reduction of up to 32.6% in the decode phase on an NVIDIA A100 GPU compared to the full KV cache baseline, while achieving negligible accuracy loss on a variety of long-context tasks. We also propose a variant of RocketKV for multi-turn scenarios, which consistently outperforms other existing methods and achieves accuracy nearly on par with an oracle top-k attention scheme. The source code is available here: https://github.com/NVlabs/RocketKV.

Mixture-of-Experts (MoE) models are typically pre-trained with explicit load-balancing constraints to ensure statistically balanced expert routing. Despite this, we observe that even well-trained MoE models exhibit significantly imbalanced routing. This behavior is arguably natural-and even desirable - as imbalanced routing allows models to concentrate domain-specific knowledge within a subset of experts. Expert parallelism (EP) is designed to scale MoE models by distributing experts across multiple devices, but with a less-discussed assumption of balanced routing. Under extreme imbalance, EP can funnel a disproportionate number of tokens to a small number of experts, leading to compute- and memory-bound failures on overloaded devices during post-training or inference, where explicit load balancing is often inapplicable. We propose Least-Loaded Expert Parallelism (LLEP), a novel EP algorithm that dynamically reroutes excess tokens and associated expert parameters from overloaded devices to underutilized ones. This ensures that all devices complete their workloads within the minimum collective latency while respecting memory constraints. Across different model scales, LLEP achieves up to 5x speedup and 4x reduction in peak memory usage compared to standard EP. This enables faster and higher-throughput post-training and inference, with ~1.9x faster for gpt-oss-120b. We support our method with extensive theoretical analysis and comprehensive empirical evaluations, including ablation studies. These results illuminate key trade-offs and enable a principled framework for hardware-specific hyper-parameter tuning to achieve optimal performance.

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.

In recent years, Mixture-of-Experts (MoE) has emerged as an effective approach for enhancing the capacity of deep neural network (DNN) with sub-linear computational costs. However, storing all experts on GPUs incurs significant memory overhead, increasing the monetary cost of MoE-based inference. To address this, we propose eMoE, a memory efficient inference system for MoE-based large language models (LLMs) by leveraging our observations from experiment measurements. eMoE reduces memory usage by predicting and loading only the required experts based on recurrent patterns in expert routing. To reduce loading latency while maintaining accuracy, as we found using the same experts for subsequent prompts has minimal impact on perplexity, eMoE invokes the expert predictor every few prompts rather than for each prompt. In addition, it skips predictions for tasks less sensitive to routing accuracy. Finally, it has task-aware scheduling to minimize inference latency by considering Service Level Objectives (SLOs), task-specific output lengths, and expert loading latencies. Experimental results show that compared to existing systems, eMoE reduces memory consumption by up to 80% while maintaining accuracy and reduces inference latency by up to 17%. It also enables processing prompts 40x longer, batches 4.5x larger, and achieves 1.5x higher throughput.

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.

Extending the context length (i.e., the maximum supported sequence length) of LLMs is of paramount significance. To facilitate long context training of LLMs, sequence parallelism has emerged as an essential technique, which scatters each input sequence across multiple devices and necessitates communication to process the sequence. In essence, existing sequence parallelism methods assume homogeneous sequence lengths (i.e., all input sequences are equal in length) and therefore leverages a single, static scattering strategy for all input sequences. However, in reality, the sequence lengths in LLM training corpora exhibit substantial variability, often following a long-tail distribution, which leads to workload heterogeneity. In this paper, we show that employing a single, static strategy results in inefficiency and resource under-utilization, highlighting the need for adaptive approaches to handle the heterogeneous workloads across sequences. To address this, we propose a heterogeneity-adaptive sequence parallelism method. For each training step, our approach captures the variability in sequence lengths and assigns the optimal combination of scattering strategies based on workload characteristics. We model this problem as a linear programming optimization and design an efficient and effective solver to find the optimal solution. Furthermore, we implement our method in a high-performance system that supports adaptive parallelization in distributed LLM training. Experimental results demonstrate that our system outperforms state-of-the-art training frameworks by up to 1.98x.

Multi-LoRA agent systems enable efficient role specialization by sharing a common backbone model. However, each agent repeatedly processes the growing shared trajectory and constructs its own KV cache, introducing substantial memory and computation redundancy in long-horizon tasks. Existing KV cache sharing methods reduce this repeated prefill, but they either require additional training or architectural constraints or retain substantial model computation. Moreover, direct cache reuse causes the current agent to rely on cache states generated by the previous agent's adapter, weakening the role-specific behavior encoded by its own LoRA. We present PReCache, a training-free KV cache sharing framework with two designs, namely PreLRShared and ReBaseShared, that share the base cache computed using the pretrained weights and precompute a compact agent-specific low-rank (LR) cache. To remove repeated prefill, PreLRShared precomputes each agent's LR cache when the shared context is first processed, allowing the current agent to use its own LR cache without reprocessing context processed by previous agents. To improve sharing accuracy, ReBaseShared reconstructs the shared base cache from adapter-free hidden states, reducing the remaining error caused by the previous agent's adapted representation. To minimize its reconstruction cost, we propose two inference schemes tailored to single-stream inference and concurrent serving, performing the same reconstruction after each agent's turn or alongside its execution, respectively. Across multiple models and agent benchmarks, PreLRShared achieves up to a 3.1x TTFT speedup and a 2.3x improvement in per-request throughput over inference without KV cache sharing. ReBaseShared best preserves accuracy overall among the evaluated cache-sharing methods, with an average drop of only 1.1 points relative to inference without cache sharing.

Diffusion Language Models (DLMs) have been seen as a promising competitor for autoregressive language models. However, diffusion language models have long been constrained by slow inference. A core challenge is that their non-autoregressive architecture and bidirectional attention preclude the key-value cache that accelerates decoding. We address this bottleneck by proposing a KV-cache-like mechanism, delayed KV-Cache, for the denoising process of DLMs. Our approach is motivated by the observation that different tokens have distinct representation dynamics throughout the diffusion process. Accordingly, we propose a delayed and conditioned caching strategy for key and value states. We design two complementary variants to cache key and value step-by-step: (1) dKV-Cache-Decode, which provides almost lossless acceleration, and even improves performance on long sequences, suggesting that existing DLMs may under-utilise contextual information during inference. (2) dKV-Cache-Greedy, which has aggressive caching with reduced lifespan, achieving higher speed-ups with quadratic time complexity at the cost of some performance degradation. dKV-Cache, in final, achieves from 2-10x speedup in inference, largely narrowing the gap between ARs and DLMs. We evaluate our dKV-Cache on several benchmarks, delivering acceleration across general language understanding, mathematical, and code-generation benchmarks. Experiments demonstrate that cache can also be used in DLMs, even in a training-free manner from current DLMs.

Attention mechanisms have revolutionized sequence learning but suffer from quadratic computational complexity. This paper introduces Lattice, a novel recurrent neural network (RNN) mechanism that leverages the inherent low-rank structure of K-V matrices to efficiently compress the cache into a fixed number of memory slots, achieving sub-quadratic complexity. We formulate this compression as an online optimization problem and derive a dynamic memory update rule based on a single gradient descent step. The resulting recurrence features a state- and input-dependent gating mechanism, offering an interpretable memory update process. The core innovation is the orthogonal update: each memory slot is updated exclusively with information orthogonal to its current state hence incorporation of only novel, non-redundant data, which minimizes the interference with previously stored information. The experimental results show that Lattice achieves the best perplexity compared to all baselines across diverse context lengths, with performance improvement becoming more pronounced as the context length increases.

Context-heavy agents place unusual pressure on the key-value (KV) cache: long prefixes are reused across many short turns, while concurrency determines whether the serving system can keep GPUs utilized. We study 4-bit KV-cache compression for this setting, using TurboQuant-style rotation and codebook quantization as a quality anchor and vLLM FP8 KV caching as the deployment anchor. We report three contributions. First, we frame 4-bit KV caching around multi-round agent workloads where task quality, cache residency, and serving throughput must be measured jointly. Second, we describe the practical design choices needed to make the 4-bit path robust, including asymmetric K/V treatment, Walsh-Hadamard rotation, QJL removal, and block-scale variants. Third, we present serving optimizations on AMD GPUs, including optimized decode-attention kernels and UltraQuant, an FP4 approximation path that uses FP8 queries, FP4 KV tensors, UE8M0 group scales, and native scaled-MFMA support on CDNA4. On a long-context, multi-turn agentic workload, UltraQuant cuts P50 time-to-first-token by 3.47x in the cache-pressured late rounds (2.3x across all rounds) and raises output throughput by 1.63x over the FP8 KV baseline.

Large multimodal Mixture-of-Experts (MoEs) effectively scale the model size to boost performance while maintaining fixed active parameters. However, previous works primarily utilized full-precision experts during sparse up-cycling. Despite they show superior performance on end tasks, the large amount of experts introduces higher memory footprint, which poses significant challenges for the deployment on edge devices. In this work, we propose MoTE, a scalable and memory-efficient approach to train Mixture-of-Ternary-Experts models from dense checkpoint. Instead of training fewer high-precision experts, we propose to train more low-precision experts during up-cycling. Specifically, we use the pre-trained FFN as a shared expert and train ternary routed experts with parameters in {-1, 0, 1}. Extensive experiments show that our approach has promising scaling trend along model size. MoTE achieves comparable performance to full-precision baseline MoE-LLaVA while offering lower memory footprint. Furthermore, our approach is compatible with post-training quantization methods and the advantage further amplifies when memory-constraint goes lower. Given the same amount of expert memory footprint of 3.4GB and combined with post-training quantization, MoTE outperforms MoE-LLaVA by a gain of 4.3% average accuracy on end tasks, demonstrating its effectiveness and potential for memory-constrained devices.

KV cache has become a de facto technique for the inference of large language models (LLMs), where tensors of shape (layer number, head number, sequence length, feature dimension) are introduced to cache historical information for self-attention. As the size of the model and data grows, the KV cache can quickly become a bottleneck within the system in both storage and memory transfer. To address this, prior studies usually focus on the first three axes of the cache tensors for compression. This paper supplements them, focusing on the feature dimension axis, by utilizing low-rank projection matrices to transform the cache features into spaces with reduced dimensions. We begin by investigating the canonical orthogonal projection method for data compression through principal component analysis (PCA). We observe the issue with PCA projection where significant performance degradation is observed at low compression rates. To bridge the gap, we propose to directly tune the orthogonal projection matrices with a distillation objective using an elaborate Matryoshka training strategy. After training, we adaptively search for the optimal compression rates for various layers and heads given varying compression budgets. Compared to previous works, our method can easily embrace pre-trained LLMs and hold a smooth tradeoff between performance and compression rate. We empirically witness the high data efficiency of our training procedure and find that our method can sustain over 90% performance with an average KV cache compression rate of 60% (and up to 75% in certain extreme scenarios) for popular LLMs like LLaMA2-7B-base and Mistral-7B-v0.3-base.

In order to make the foundation model more efficient and effective, our idea is combining sequence transformation and state transformation. First, we prove the availability of rotary position embedding in the state space duality algorithm, which reduces the perplexity of the hybrid quadratic causal self-attention and state space duality by more than 4%, to ensure that the combining sequence transformation unifies position encoding. Second, we propose dynamic mask attention, which maintains 100% accuracy in the more challenging multi-query associative recall task, improving by more than 150% compared to quadratic causal self-attention and state space duality, to ensure that the combining sequence transformation selectively filters relevant information. Third, we design cross domain mixture of experts, which makes the computational speed of expert retrieval with more than 1024 experts 8 to 10 times faster than the mixture of experts, to ensure that the combining state transformation quickly retrieval mixture. Finally, we summarize these matrix algorithms that can form the foundation model: Wonderful Matrices, which can be a competitor to popular model architectures.

Rotary Position Embedding (RoPE) enables each attention head to capture multi-frequency information along the sequence dimension and is widely applied in foundation models. However, the nonlinearity introduced by RoPE complicates optimization of the key state in the Key-Value (KV) cache for RoPE-based attention. Existing KV cache compression methods typically store key state before rotation and apply the transformation during decoding, introducing additional computational overhead. This paper introduces EliteKV, a flexible modification framework for RoPE-based models supporting variable KV cache compression ratios. EliteKV first identifies the intrinsic frequency preference of each head using RoPElite, selectively restoring linearity to certain dimensions of key within attention computation. Building on this, joint low-rank compression of key and value enables partial cache sharing. Experimental results show that with minimal uptraining on only 0.6% of the original training data, RoPE-based models achieve a 75% reduction in KV cache size while preserving performance within a negligible margin. Furthermore, EliteKV consistently performs well across models of different scales within the same family.

Mixture-of-Experts (MoE) based Large Language Models (LLMs) have demonstrated impressive performance and computational efficiency. However, their deployment is often constrained by substantial memory demands, primarily due to the need to load numerous expert modules. While existing expert compression techniques like pruning or merging attempt to mitigate this, they often suffer from irreversible knowledge loss or high training overhead. In this paper, we propose a novel expert compression paradigm termed expert replacing, which replaces redundant experts with parameter-efficient modules and recovers their capabilities with low training costs. We find that even a straightforward baseline of this paradigm yields promising performance. Building on this foundation, we introduce LightMoE, a framework that enhances the paradigm by introducing adaptive expert selection, hierarchical expert construction, and an annealed recovery strategy. Experimental results show that LightMoE matches the performance of LoRA fine-tuning at a 30% compression ratio. Even under a more aggressive 50% compression rate, it outperforms existing methods and achieves average performance improvements of 5.6% across five diverse tasks. These findings demonstrate that LightMoE strikes a superior balance among memory efficiency, training efficiency, and model performance.

Diffusion Large Language Models (dLLMs) enable breakthroughs in reasoning and parallel decoding but suffer from prohibitive quadratic computational complexity and memory overhead during inference. Current caching techniques accelerate decoding by storing full-layer states, yet impose substantial memory usage that limit long-context applications. Our analysis of attention patterns in dLLMs reveals persistent cross-layer sparsity, with pivotal tokens remaining salient across decoding steps and low-relevance tokens staying unimportant, motivating selective cache eviction. We propose Sparse-dLLM, the first training-free framework integrating dynamic cache eviction with sparse attention via delayed bidirectional sparse caching. By leveraging the stability of token saliency over steps, it retains critical tokens and dynamically evicts unimportant prefix/suffix entries using an attention-guided strategy. Extensive experiments on LLaDA and Dream series demonstrate Sparse-dLLM achieves up to 10times higher throughput than vanilla dLLMs, with comparable performance and similar peak memory costs, outperforming previous methods in efficiency and effectiveness.

Key-value (KV) caching has become the de-facto to accelerate generation speed for large language models (LLMs) inference. However, the growing cache demand with increasing sequence length has transformed LLM inference to be a memory bound problem, significantly constraining the system throughput. Existing methods rely on dropping unimportant tokens or quantizing all entries uniformly. Such methods, however, often incur high approximation errors to represent the compressed matrices. The autoregressive decoding process further compounds the error of each step, resulting in critical deviation in model generation and deterioration of performance. To tackle this challenge, we propose GEAR, an efficient KV cache compression framework that achieves near-lossless high-ratio compression. GEAR first applies quantization to majority of entries of similar magnitudes to ultra-low precision. It then employs a low rank matrix to approximate the quantization error, and a sparse matrix to remedy individual errors from outlier entries. By adeptly integrating three techniques, GEAR is able to fully exploit their synergistic potentials. Our experiments demonstrate that compared to alternatives, GEAR achieves near-lossless 4-bit KV cache compression with up to 2.38x throughput improvement, while reducing peak-memory size up to 2.29x. Our code is publicly available at https://github.com/HaoKang-Timmy/GEAR.

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.

Transformer-based large language models (LLMs) have already achieved remarkable results on long-text tasks, but the limited GPU memory (VRAM) resources struggle to accommodate the linearly growing demand for key-value (KV) cache as the sequence length increases, which has become a bottleneck for the application of LLMs on long sequences. Existing KV cache compression methods include eviction, merging, or quantization of the KV cache to reduce its size. However, compression results in irreversible information forgetting, potentially affecting the accuracy of subsequent decoding. In this paper, we propose SpeCache, which takes full advantage of the large and easily expandable CPU memory to offload the complete KV cache, and dynamically fetches KV pairs back in each decoding step based on their importance measured by low-bit KV cache copy in VRAM. To avoid inference latency caused by CPU-GPU communication, SpeCache speculatively predicts the KV pairs that the next token might attend to, allowing us to prefetch them before the next decoding step which enables parallelization of prefetching and computation. Experiments on LongBench and Needle-in-a-Haystack benchmarks verify that SpeCache effectively reduces VRAM usage while avoiding information forgetting for long sequences without re-training, even with a 10x high KV cache compression ratio.

Large Language Models (LLMs) show great capabilities in a wide range of applications, but serving them efficiently becomes increasingly challenging as requests (prompts) become more complex. Context caching improves serving performance by reusing Key-Value (KV) vectors, the intermediate representations of tokens that are repeated across requests. However, existing context caching requires exact prefix matches across requests, limiting reuse cases in settings such as few-shot learning and retrieval-augmented generation, where immutable content (e.g., documents) remains unchanged across requests but is preceded by varying prefixes. Position-Independent Caching (PIC) addresses this issue by enabling modular reuse of the KV vectors regardless of prefixes. We formalize PIC and advance prior work by introducing EPIC, a serving system incorporating our new LegoLink algorithm, which mitigates the inappropriate "attention sink" effect at every document beginning, to maintain accuracy with minimal computation. Experiments show that EPIC achieves up to 8x improvements in Time-To-First-Token (TTFT) and 7x throughput gains over existing systems, with negligible or no accuracy loss.

Hybrid LLMs interleave full-attention layers with linear-attention layers to reduce long-context inference cost, but this structure complicates prefix caching. Full-attention KV caches are token-addressable, whereas linear-attention layers maintain recurrent states that cannot be rolled back to arbitrary prefix boundaries. Existing systems materialize recurrent-state checkpoints, restricting prefix reuse to checkpoint-aligned positions. We present SuffixReplay, the first prefix caching system that lets hybrid LLMs reuse cached prefixes at every cache-supported page boundary without materializing recurrent-state checkpoints. Our key insight is to just let linear states forget the distant past. Modern linear-attention mechanisms use recurrent decay and gating to attenuate the influence of old inputs. Therefore, instead of checkpointing every prefix boundary, SuffixReplay approximates the state at a matched boundary by replaying only a recent suffix of the layer's input hidden states, which we retain as anchors. At the algorithmic level, SuffixReplay combines layer-wise and token-wise anchor sparsity with a bounded replay budget to control storage, computation, and quality. At the system level, it uses an independently managed anchor sidecar and a pipelined replay path to overlap anchor movement and state reconstruction with the native serving pipeline. We evaluate SuffixReplay on three hybrid LLMs: OLMo-Hybrid-7B, Qwen3.5-4B, and Qwen3.6-27B-FP8. Across these models, SuffixReplay retains 91.4-100% of full-prefill quality on average across LongBench and RULER, while using only 0.36-0.51x the amortized per-token storage of SGLang's default 8192-token checkpoint cache. Integrated into SGLang, SuffixReplay reduces median TTFT by 15-70% on branching workloads, sustains 2.3-4.3x SGLang's throughput when the working set exceeds HBM, and matches SGLang on high-hit continuation traffic.