Cache

Long-context large language model (LLM) inference is increasingly constrained by the memory footprint and decoding cost of key-value (KV) caches, limiting sustainable deployment on resource-constrained hardware. Existing KV cache eviction methods typically apply heuristic token scoring over all heads in GQA-based LLMs. These methods ignore the different functionalities of attention heads, leading to the eviction of critical tokens and thus degrading the performance of LLMs. To address this issue, we propose CompressKV, a resource-efficient KV-cache compression framework for GQA-based LLMs. Instead of aggregating attention scores from all heads, CompressKV identifies Semantic Retrieval Heads (SRHs) that capture both the initial and final tokens of a prompt and semantically important mid-context evidence, and uses them to select tokens whose KV pairs should be retained. Furthermore, CompressKV allocates cache budgets across layers according to offline estimates of layer-wise eviction error. Experiments on LongBench and Needle-in-a-Haystack show that CompressKV consistently outperforms existing KV-cache eviction methods across memory budgets. Notably, it preserves over 97% of full-cache performance using only 3% of the KV cache on LongBench question-answering tasks and achieves 90% accuracy with just 0.7% KV storage on Needle-in-a-Haystack. These results demonstrate an improved resource--performance trade-off for long-context LLM inference. Our code is publicly available at: https://github.com/TUDa-HWAI/CompressKV

Existing low-bit KV-cache quantizers often treat each cached key as a flat vector. Under RoPE, however, a key's contribution to a future attention logit decomposes into a position-dependent sum over two-dimensional frequency blocks. This makes key-cache quantization a block-wise bit-allocation problem: high-energy RoPE blocks are more sensitive to quantization error and should receive more bits. We introduce Block-GTQ, a RoPE-aware bit allocator for key-cache quantization built on TurboQuant-MSE(TQ-MSE). For each layer and KV head, Block-GTQ computes a label-free energy score for each RoPE block and greedily allocates integer bit widths by marginal gain. Under matched K/V bit budgets, Block-GTQ better preserves RoPE query-key logits on a ten-model diagnostic panel, cutting per-layer MAE by 32-80% at 2 and 3 b/dim K-only quantization and winning all 367/367 layer comparisons against uniform TQ-MSE. These fidelity gains translate to stronger downstream long-context retrieval, understanding, and reasoning. At K2V2 on Llama-3.1-8B-Instruct, Block-GTQ raises the six-task NIAH average from 70.6 to 97.4, and the LongBench-EN average from 36.87 to 53.31. On AIME 2024/2025 with DeepSeek-R1-Distill-Qwen-7B, without an fp16 recent-key buffer, Block-GTQ at K3V2 scores 51.7/37.5, close to fp16's 54.2/37.9, whereas uniform TQ-MSE collapses to 0.0/0.0. We further implement a packed-cache serving path. On a single H800 GPU with Qwen2.5-3B-Instruct, packed K3V3 achieves 3.24x KV-cache compression with fp16-comparable quality, runs 1.34x faster than fp16 FlashAttention2 at 128K context, reduces peak memory from 56.31 GB to 19.85 GB, and remains feasible at 256K and 512K where fp16 OOMs. Code is available at https://github.com/JIA-Lab-research/blockgtq.

Long-context and agentic LLM workloads push the KV cache past any fixed memory budget, forcing the inference stack to permanently evict tokens at every step of a continuous-inference stream. Existing methods all share the same template, a per-step direct-attention score followed by deterministic top-$K$ selection, which converts a single below-cutoff step into an irreversible verdict and permanently erases any subtly important token that direct attention cannot single out from noise. To address this challenge, we propose Nexus Sampling, a training-free eviction method that pairs Nexus scoring, an iterative walk over direct attention that surfaces bridge tokens, with weighted reservoir sampling, which retains tokens with inclusion probability in place of deterministic top-$K$. Theoretically, we show that Nexus Sampling dominates deterministic top-$K$ in long-run survival of subtly important tokens. Empirically, at 80% KV cache eviction, Nexus Sampling matches dense attention within 1% on LongBench while outperforming top-$K$ baselines on retrieval-heavy tasks, with up to 10x smaller per-sequence cache memory.

Multimodal agents repeatedly re-examine the same video frames, UI screenshots, and rendered artifacts as their context window slides and reasoning iterates, yet every look-back re-encodes from scratch, because prefix caches serve reuse only at a fixed leading position. We show this recompute is avoidable, and identify exactly what naive KV reuse loses: the cross-chunk conditioning a chunk absorbs from its neighbours. This loss is asymmetric. The direct readout of a cached chunk is recovered exactly and for free by the standard state-merge. What remains is a diffuse, low-rank residue concentrated in deep layers, invisible to single-hop retrieval but precisely what multi-hop reasoning binds on. Blind reuse therefore leaves single-hop recall intact while halving multi-hop accuracy; this is the failure mode prior position-independent caches, designed for single-context or single-image reuse, do not address. We repair it with a small, training-free low-rank conditioning patch stored alongside each position-free chunk. Reuse reduces to one operator across MLA, GQA, and MHA: exact RoPE re-rotation to any target position, plus the patch that restores cross-chunk binding. This makes three window operations cheap: reorder (one patch serves every ordering of a cached set), sliding-window survival (surviving chunks relocate via rotation only, zero re-encode), and recall (an evicted chunk is rehydrated by its patch, never re-encoded). A rank-m patch recovers full task accuracy on cross-chunk-binding benchmarks, MM-NIAH across two attention families and two-page doc-QA, at a fraction of the KV footprint, and reconstructs re-prefill KV to within bf16 rounding in a production SGLang kernel across six backbones. The conditioning signal is strongest in redundant vision and video streams, making our solution most impactful where multimodal agents spend their recompute budget.

Long-running LLM agents keep valuable state resident on GPUs: KV caches, request schedulers, communication state, and sometimes online adapters. Losing this state after a GPU or communicator failure can discard minutes to hours of work, yet existing recovery mechanisms either restart the whole serving stack or require application-specific checkpoint logic inside every attention and runtime component. This paper argues that fault tolerance for such workloads needs a GPU-resident execution context: checkpoint hooks must run at device synchronization points, observe binary kernels that frameworks and libraries actually execute, and recover without putting the host CPU on the critical path. We present Concordia, a runtime that uses a device-resident persistent kernel as the substrate for fault-tolerant LLM inference. Concordia interposes on GPU module loading and supports PTX- and SASS-level instrumentation, allowing checkpoint and pause hooks to be inserted below framework code and library boundaries. For each registered LLM state region, Concordia JIT-compiles a specialized delta-checkpoint handler -- for example, a KV-block scanner, adapter-page scanner, or recovery applier -- and hot-swaps it into the persistent kernel's operator table. The persistent kernel consumes a lock-free ring buffer of compute, checkpoint, append-log, and recovery tasks, so the same always-on executor triggers dirty-page detection, stages deltas, and appends committed records to a CPU-visible log in CXL memory or host DRAM.

Recently, end-to-end OCR models, exemplified by DeepSeek OCR, have once again thrust OCR into the spotlight. A widely held view is that employing a large language model (LLM) as the decoder allows the model to leverage the prior distribution of language, leading to improved OCR performance. However, the downside is equally evident: as the output sequence lengthens, the accumulated KV cache drives up memory consumption and progressively slows down generation. This stands in stark contrast to humans, who exhibit no such decline in efficiency during long-horizon copying tasks. In this technical report, we propose Unlimited OCR, a model designed to emulate human parsing working memory. Taking DeepSeek OCR as the baseline, we replace all attention layers in the decoder with our proposed Reference Sliding Window Attention (R-SWA), which reduces attention computation costs while maintaining a constant KV cache throughout the entire decoding process. By combining the high compression rate of DeepSeek OCR's encoder with our constant KV cache design, Unlimited OCR can transcribe dozens of pages of documents in a single forward pass under a standard maximum length of 32K. More importantly, R-SWA is a general-purpose parsing attention mechanism - beyond OCR, it is equally applicable to tasks such as ASR, translation, etc. Codes and model weights are publicly available at http://github.com/baidu/Unlimited-OCR.

Hash-based data structures such as Bloom filters are widely used in network systems for tasks including caching, anomaly detection, and machine learning pipelines. They typically provide binary indications of whether an element belongs to a set of interest, e.g., the contents of a cache. When uncertainty arises due to hash collisions, a positive indication is returned to avoid false negatives. We argue that the certainty associated with such indications can itself be useful information. This work focuses on Counting Bloom Filters (CBFs), a Bloom-filter variant that maintains counters rather than bits. Besides supporting insertions and deletions, these counters provide additional information that can be used to estimate the certainty of positive membership indications. We show how this certainty signal can be exploited in architectures that combine Bloom Filters with machine learning (ML) models.

The explosive demand for interactive Large Language Model serving has highlighted the management of the Key-Value cache's dynamic memory footprint as a critical area for performance optimization in inference engines. Modern inference systems overwhelmingly rely on time-centric scheduling heuristics, such as Shortest Job First. However, their theoretical optimality is rooted in traditional schedule modeling, failing to capture the highly dynamic, 2D spatio-temporal geometric growth specific to LLM inference mechanisms. To resolve this, we propose the geometry-aware online scheduling by introducing the Smallest Volume First (SVF) algorithm and its highly efficient variant, 1-bit SVF. Theoretically, we provide a rigorous mathematical foundation for our approach. Via a novel volume-certificate proof, we sharpen SVF's worst-case competitive ratio from the prior best of 48 towards \textbf{3} in the high-concurrency regime of LLM serving. Building upon this core breakthrough, we complete a comprehensive theoretical taxonomy analyzing our algorithms across different traffic scenarios and information availability. Practically, we seamlessly integrate our approach as a plug-and-play layer in vLLM. Extensive evaluations on Llama-3.1 models demonstrate comprehensive performance gains: SVF delivers strong reductions in both average and tail latency, while 1-bit SVF, with merely a single bit information, achieves competitive throughput and latency. This work establishes a theoretically sound and empirically proven approach for resolving memory-constrained scheduling in modern LLM deployments. To facilitate future research, our code is available at https://github.com/Aurora-Kl/Geometry-Aware-Online-Scheduling.git.

Traditional architectural design space exploration (DSE) is highly inefficient, typically requiring tens of thousands of simulator evaluations across various optimization methods. This inefficiency arises because conventional methods treat the simulator as a black-box oracle. In contrast, human architects effectively guide exploration by reasoning through physical constraints, performance bottlenecks, data reuse, and workload structures. To bridge this gap, we introduce AgentDSE, a simulator-in-the-loop methodology driven by a general-purpose large language model (LLM) coding agent. AgentDSE automates this architectural-reasoning loop without requiring model fine-tuning, precomputed design databases, or domain-specific optimizer code. Across deep neural network (DNN) accelerator mapping, hardware/software co-design, and CPU cache-hierarchy optimization, AgentDSE achieves competitive or better design quality with up to two orders of magnitude fewer evaluations. AgentDSE also produces inspectable traces that surface architectural hypotheses, performance cliffs, implicit priors, and simulator artifacts, making every search decision traceable rather than buried in optimizer state.

Diffusion LLMs (dLLMs) improve GPU utilization over autoregressive decoding by generating multiple tokens per forward pass, but their KV cache still grows linearly with context, limiting throughput at long contexts. KV cache offloading to host DRAM alleviates this memory pressure, but the limited PCIe bandwidth necessitates recalling only a sparse subset of KV entries. In block dLLMs, the relevant KV entries remain consistent across denoising steps within a block, enabling high-accuracy selection by identifying the top-k entries once and reusing them throughout all denoising steps. This property appears attractive for offloading as it amortizes the selection overhead across the entire block, but it requires exact attention over the full KV cache, which is too expensive under offloading. We present HERALD, a KV offloading system for block dLLMs that resolves this through two opportunities that reduce the required selection compute by a factor of the block size and enable selection to be overlapped with denoising. Across three block dLLMs and five long context tasks, HERALD achieves near-lossless accuracy at 5-10% KV budget and up to 1.59x lower per block latency and 2.47x higher throughput over GPU-only inference, with speedups growing with context length.

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.

Mainstream LLM serving systems reuse prefix work mainly through paged or radix key-value (KV) caches. This is highly effective for high-throughput, high-concurrency serving, but it manages only one positional fragment of execution state: the KV cache. We study the opposite regime: low-latency, small-batch, on-device physical-AI serving, where interactive LLM agents, speech systems, and robot policies repeatedly branch, reset, interrupt, and re-enter under tight responsiveness budgets. We introduce execution-state capsules, a graph-bound checkpoint and restore mechanism for the complete restorable state at a committed boundary. FlashRT is a white-box, backend-facing kernel runtime whose evaluated NVIDIA CUDA backend runs captured graph plans over contiguous static buffers with no block-table indirection. Because the live state is a closed set of named buffers, a capsule can snapshot, restore, fork, or roll back the whole execution boundary, including KV, recurrent state, convolution state, MTP state, and metadata. This moves reuse from token-addressed KV fragments to graph-bound execution-state boundaries. On an RTX 5090, capsule restore is byte-exact at the stored-state level and token-identical under greedy decode. A KV-only ablation diverges, showing that recurrent state is load-bearing. GPU-resident snapshot and restore are sub-millisecond, and TTFT speedup over cold prefill grows from 3.9x at 2k tokens to 27x at 16k tokens. On Jetson AGX Thor and DGX Spark, the same correctness and structural properties hold. Capsules are not a replacement for high-throughput KV-cache serving; they define a complementary latency-first serving point for explicit execution-state reuse.

Streaming VideoLLMs must continuously process incoming video while maintaining low query latency, making both video-ingestion throughput and query-time responsiveness critical for real-time deployment. Existing methods largely focus on accelerating individual modules, such as visual encoding, token pruning, or KV-cache compression, but provide limited insight into whether the resulting system can sustain real-time streaming performance. We formulate streaming VideoLLM inference as a coordinated pipeline spanning visual preprocessing, visual encoding, token dropping, and LLM prefilling/decoding. Building on this formulation, we propose ViCoStream (Video Coordinated Streaming), a stage-wise coordinated streaming framework that combines chunk-wise execution, CUDA-stream overlap, visual token control, bounded visual attention, and query-side retrieval to bound per-chunk computation and memory costs. We further provide a systematic study of bottleneck migration, revealing how chunk size, token retention, attention locality, and retrieval scope shape the throughput-accuracy trade-off. Experiments with Qwen2.5-VL-3B/7B-Instruct across multiple streaming benchmarks show that ViCoStream achieves 134 FPS video throughput and less than 50 ms TTFT on a single A100 GPU while maintaining accuracy close to full-history baselines.

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.

Retrieval-Augmented Generation (RAG) improves factual grounding, but it also lengthens prompts and raises prefill cost. Prefix caching in serving engines such as vLLM reduces this cost only when requests share the same token prefix. In grounded generation, however, adjacent queries may retrieve overlapping evidence in different orders, so set overlap does not become reusable prefix overlap. We present CacheWeaver, a lightweight prompt-layer method for cache-aware evidence ordering. The method keeps a prefix tree over recently served evidence sequences and uses a greedy walk to place the most reusable prefix first, while leaving the serving engine and retrieved evidence set unchanged. Across three vLLM configurations, the method lowers median time-to-first-token (TTFT) by about 20-33 percent relative to retrieval-order prefix caching, without hurting answer quality in our QA tests. The greedy policy reaches 97.5 percent of the median TTFT gain from oracle ordering, indicating that most reusable prefix locality can be recovered by a simple scheduling layer between retrieval and inference.

Large language models (LLMs) outperform earlier architectures on generative inference and long-context tasks, but their large size introduces significant challenges in memory usage, energy cost, and on-device deployment. Since scaling pre-trained language models improves downstream capability \cite{zhao2023survey}, the key-value (KV) cache becomes a dominant inference bottleneck. Recent KV cache compression methods \cite{jo2025fastkv,li2024snapkv,zhou2024dynamickv} reduce this cost by retaining only a subset of attention-relevant tokens. However, while these approaches preserve accuracy on benign workloads, their compression policies either fail to defend against jailbreak attacks \cite{jiang2024robustkv} or degrade safety alignment under aggressive eviction. We propose AnchorKV, a drop-in modification to KV cache compression that biases token retention scores away from directions in key space associated with harmful prompts. AnchorKV constructs an offline safety anchor by adapting a difference-of-means representation engineering approach \cite{arditi2024refusal,zou2023representation} to the layer-specific key projection space used in KV caching. Based on this anchor, a soft penalty token selection rule trades a small amount of utility for substantially improved safety alignment, while reducing to the original compressor when the penalty is zero.

Post-hoc context erasing over the KV cache is challenging because a local edit has a global consequence: once a span has been processed, its influence propagates into the cached states of all subsequent tokens. This issue arises naturally in long-context LLM applications, where stale retrieved facts, incorrect tool observations, retracted user preferences, or harmful prompt injections may be identified only after prefill. Exact erasing must then recompute all tokens after the deleted span, making its computational cost depend on suffix length rather than erased-span length. We introduce KVEraser, a learned KV-cache editing method for efficient localized context erasing. Given a processed context and a span to remove, KVEraser replaces only the KV states of the erased interval with learned steering states while reusing the remaining cache unchanged. To learn a transferable erasing mechanism, we build a two-stage training pipeline: generic span-neighbor pre-training teaches the eraser to suppress the influence of the erased span, while task-specific fine-tuning adapts this capability to downstream scenarios. Experiments show that KVEraser nearly matches full recomputation in post-erasure performance on in-domain tasks across 1K--32K context lengths, while its latency increases by only 24% compared with a 17.6x increase for full recomputation. KVEraser also generalizes to unseen long-document QA tasks with harmful factual distractors, achieving the best performance among approximate baselines with a 3--4x speedup over full recomputation.

Muon is an optimizer that computes updates using the polar factor of the momentum matrix and has shown strong empirical performance across a range of training settings. A key component of Muon is the Newton-Schulz iteration used to compute this polar factor. Although this avoids the cost of an exact singular value decomposition, it remains expensive in practice because it is applied at every optimization step. At the same time, the momentum matrix changes smoothly over training, suggesting strong temporal correlation in the corresponding polar factors. In this paper, we exploit this structure and propose CacheMuon, a temporal preconditioning method that reuses information from previous optimization steps to approximate the polar factor at the current step. This reduces redundant orthogonalization computation across iterations. We analyze CacheMuon as an inexact Muon update, with error controlled by fresh-solver error and cache staleness. Empirically, CacheMuon provides a controllable quality-efficiency frontier: conservative thresholds closely match fresh Muon on language-model and vision training while reducing orthogonalization FLOPs, whereas more aggressive thresholds yield larger arithmetic savings at the cost of modest validation-quality degradation.

This study systematically compares Turbo-Quant and SpectralQuant KV-cache compression, evaluating non-dominated schemes, including WHT rotation with Beta Lloyd-Max and QJL, through a statistical validation methodology that separates systematic codec differences from implementation variance. Key findings reveal that while eigenbasis-based methods fail on heavy-tailed data due to covariance instability, they excel in structured regimes, with the effective semantic dimension ($d_{eff}$) adapting to calibration budgets rather than true data rank. (this is an abstract of the abstract thank you )

Prefix caching reuses prefill only across an exactly shared prefix, so one changed field invalidates the entire downstream cache. Yet overwriting the field's own key/value vectors and reusing the rest leaves the model acting on the old value. The reason, established causally across four model families: at prefill the model has already written the field-conditioned conclusion onto downstream notes; the field's own key/value drives under 1% of the decision. Read as a notebook of memoized conclusions, two capabilities follow. (1) It is editable. A salient erratum amends the notes; and with chain-of-thought, editing the field alone recovers the decision (1.00 at 8B, ~1% compute), while without CoT it is ignored. (2) It is composable. The notes are position-portable, so a precompiled skill can be RoPE-repositioned and spliced into any context, indistinguishable from full recompute (logit cosine 0.90-0.999, twelve models) at O(L) rather than O(L^2) time-to-first-token. A unified edit+compose agent stays decision-identical to recompute at up to 14.9x lower latency. The approach applies to any per-token attention KV cache, validated across scale, quantization, Mixture-of-Experts, and multimodal caches, and extends to several attention variants through small adapters. Because the erratum is append-only, it composes with production prefix caching: in an online vLLM benchmark it keeps the prefix cache-aligned (98.5% hit-rate), cutting p90 time-to-first-token by 53-398x.

Diffusion Transformers with Mixture-of-Experts (DiT-MoE) improve model capacity under sparse activation, but diffusion inference is still bottlenecked by redundant computation across timesteps. Existing caching methods mainly operate at the token level, which becomes suboptimal in DiT-MoE because each token update is internally decomposed into multiple routed expert branches. Our analysis shows that cross-timestep redundancy in DiT-MoE is better characterized at the expert-branch level than at the whole-token level. Based on this observation, we propose MoECa, a fine-grained caching framework that performs branch-level feature reuse across timesteps. MoECa further introduces expert-aware adaptive control and synchronized cache updates across MoE and attention paths to maintain stable intermediate states. Experiments on multiple DiT-MoE models show that MoECa consistently achieves a better speed-quality trade-off than prior caching methods, with up to 2.83$\times$ inference speedup and minimal quality degradation.

In large language model (LLM) serving, each request accumulates persistent graphics processing unit (GPU) memory during service as its key-value cache grows with every generated token. Under high concurrency, aggregate memory usage therefore increases endogenously over time: the service process itself creates future capacity pressure. When memory capacity is exceeded, systems evict active requests, discarding cached state and restarting them later, which wastes computation and reduces throughput. We develop a discrete-time dynamical model of memory-constrained LLM inference that captures admission, memory growth, and eviction under continuous batching. In the saturated-input regime, the system admits both eviction-free fixed points and limit cycles with evictions. For homogeneous workloads, we show that the eviction-free equilibrium is unstable and that, except for a Lebesgue-measure-zero exact-capture set, the system converges to a unique worst-case limit cycle that is asymptotically stable outside this exceptional set, with throughput losses as large as 50%. For heterogeneous workloads, we prove a stability criterion in the two-class common-input setting and explain how the survival-polynomial mechanism generalizes to multiple classes and heterogeneous-input lengths. Under an input-dominated scaling regime, coprime decoding lengths stabilize the eviction-free equilibrium, while non-coprime lengths create synchronized modes that drive instability. These results characterize when workload heterogeneity desynchronizes completions and helps stabilize memory-constrained serving. More broadly, we identify service-induced congestion as a structural instability mechanism and derive scheduling design principles for sustaining high throughput.

In a modern processor, computing is the cheap part. Most of its area and energy go to \emph{addressing} -- moving operands to and from a register file and cache, and running the tags, ports, miss queues, and bypass networks that find a value where it was left. MADAR deletes that machinery by abolishing the address. All state circulates in rings of slots that advance one position per clock; instructions and data ride in the same slots; a value is named by its place in an orbit -- a \rp{} coordinate -- not by an address; a fixed station computes when a circulating instruction sweeps past its operands, on a schedule set at compile time; and a hierarchy of rings of increasing period replaces the cache hierarchy, movement between them scheduled rather than triggered by a miss. No prior circulating-store, dataflow, or statically scheduled machine combines all four of these. We define the execution model, validate it in a cycle-accurate register-transfer-level implementation, show it \emph{compilable} -- a constructive scheduler emits programs cross-checked against the implementation -- and price it with a first-order energy model. The payoff is clearest for AI acceleration: the multiply-accumulate at the heart of every matmul and convolution compiles to a streaming form whose energy per operation stays flat as the reduction grows, and the operand reuse that makes matrix multiplication efficient is carried by the ring-period hierarchy -- the memory hierarchy doing by rotation what a cache does by tags. MADAR is a new design point for any computation whose data movement is known before the program runs.

KV cache eviction methods typically use a single retention-rule family throughout a model, making eviction-method identity a model-level design choice. Yet Transformer layers differ substantially in their attention behavior, representations, and sensitivity to compression, suggesting that a uniform rule may overlook useful layer-wise structure. This raises a basic question: should eviction methods themselves vary across layers? We investigate this question by composing existing eviction methods across Transformer layers and systematically exploring the resulting layer-wise design space. Using simple offline profiles, we construct fixed routes and study how their quality varies with method placement and cache budget. On LongBench, heterogeneous routing improves performance on a majority of tasks over homogeneous policies at the same cache budget. Even when method counts are held fixed, the profile-guided placement ranks second among 100 evaluated assignments, demonstrating that routing quality depends strongly on where methods are placed. Moreover, the same fixed route outperforms the best of nine standalone baselines across all five tested cache budgets. Together, these results establish layer-wise method composition as an exploitable, placement-sensitive design dimension for KV cache compression.

We present CacheRL, a system for training small agent foundation models that achieves 92 percent process accuracy on multi-step tool-calling tasks, approaching GPT-5's 94 percent while requiring 100 times less compute. Our approach addresses three challenges in practical agent training: transferring tool-calling knowledge from large models at scale, enabling reinforcement learning without costly live tool execution, and learning robustly from noisy cached environments. CacheRL introduces three key innovations. First, a hybrid thinking trajectory pipeline augments agent trajectories with LLM-generated reasoning traces, producing training examples that teach models not only what tools to call but also why. Second, the CacheAgentLoop eliminates live execution costs through a three-tier fuzzy cache while preserving trajectory fidelity using token-level masking. Third, a cache-tier-aware reward dynamically adjusts answer-quality weights to avoid penalizing models for cache-induced limitations. Through iterative supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO), CacheRL improves Qwen3-4B-Thinking's validation reward from 0.43 to 0.78. On public agentic tool-calling benchmarks, our model achieves competitive performance against frontier models such as GPT-5. Ablation studies show that removing knowledge transfer reduces performance by 41 percent, while cache-aware rewards contribute a 17 percent improvement. Interestingly, reinforcement learning improves training stability but yields limited gains beyond strong supervised fine-tuning, suggesting that data quality and reward design play a more important role than complex optimization methods in building practical small agent models.

Multi-agent systems communicate mostly through text, paying a lossy and expensive decode and re-encode cost. KV-cache communication is a promising alternative, yet most prior work is homogeneous, using duplicate copies of the same model, and avoids the central challenge of cross-model latent alignment; existing heterogeneous methods are also restrictive, typically assuming shared input and using transferred caches mainly for steering. We study a more fundamental question: can heterogeneous agents be aligned well enough to perform real "mind reading" and transfer both what one agent sees and how it thinks? Our information-structure analysis reveals a duality: context-aware transfer is driven by sparse reasoning signals, while context-unaware transfer, where the receiver sees no input, requires dense contextual knowledge preservation. Motivated by this, we propose dense alignment for heterogeneous KV-cache communication via a lightweight cross-model cache transformation and two-phase training: reconstruction followed by generation. Across all six directions of {Qwen3-4B, 8B, 14B} and six in-domain and out-of-domain benchmarks, our method outperforms prior heterogeneous baselines, matches or exceeds text communication in context-aware settings at roughly 2 to 3 times lower compute, and remains effective in context-unaware transfer where prior methods collapse.

Step-level caching accelerates diffusion models by exploiting temporal redundancy across denoising steps. Existing methods make per-step cache decisions using threshold-based heuristics, without directly optimizing for final output quality. As a result, their inference latency varies across inputs and is difficult to control at deployment. In this work, we propose BudCache, which inverts this formulation: rather than letting per-step error thresholds dictate the runtime cost, we fix the compute budget in advance and search for the cache policy that best preserves the final output. To tackle the combinatorial complexity of step selection, we combine Simulated Annealing with deterministic Hill Climbing. This offline search identifies high-quality cache policies within minutes and introduces no online search or thresholding overhead during inference. When the compute budget is very tight, we further introduce cache-aware schedule alignment, which adapts the time discretization to the selected cache policy to reduce cache-induced trajectory mismatch. Experiments on FLUX.1-dev and Wan2.1 show that BudCache achieves better generation quality than heuristic caching baselines under the same inference budgets. Code is available at https://github.com/Westlake-AGI-Lab/BudCache

Right now, across the world, AI agents are repeating the same absurd act: to read one document, they each recompute it from scratch. Every agent re-runs prefill, the most compute-intensive step a large model takes, over identical text, only to rebuild a key-value (KV) cache identical to the one the agent before it just built. The same answer, computed a million times. We make a proposal that is almost offensively simple: compute it once. Let a publisher precompute a document's KV cache, and let every other agent buy the right to load it and skip prefill. It works, and it is token-exact: loading a precomputed KV and continuing matches prefilling from scratch (24/24 greedy tokens, and at the logits level), with no accuracy cost. On Qwen3-4B, reuse is 9-50x cheaper in compute than prefill, and the gap widens with length (prefill's attention scales with L^2), so a single reuse already pays it back. Then the part that matters: where the KV lives. Shipping it fails, because KV is nearly incompressible, so per-load egress costs more than the prefill it saves. Hosting it provider-side, exactly as production prompt-caching works, removes egress entirely. The size of the prize is set by our measured compute saving: serving one hot 3774-token document to 80M agents costs ~$1.5M to re-prefill but only ~$0.03M of reuse compute (49.7x less). The 0.1x cache-read tariff APIs charge passes a 10x discount to users while sitting inside this measured envelope, so the 10x is a floor that the measured ~50x compute saving clears, and the gap to the physical ~50x is provider margin: millions of dollars per popular document. We frame the resulting agent-native prefill CDN and leave lossless KV compression and a cross-party payment layer as the open problems.

Retrieval-augmented and agentic workloads repeatedly prefill recurring predictable structured inputs (which we call "spans") such as documents and code files. Yet, prefix caching in engines such as vLLM cannot reuse their KV entries unless they share identical prefixes with another request, while Position-Independent Caching (PIC) implementations within production-grade inference servers typically either require substantial server code changes or keep KV state outside the server, incurring host-to-device transfer overhead. We present Minimalistic PIC (MiniPIC): a minimal, flexible and fast vLLM design built from two ingredients: positional-encoding-free KV cache and user-controlled cache-reuse primitives. MiniPIC stores unrotated K vectors in the KV cache, applies RoPE to K tiles inside attention using per-request logical positions, and exposes three user-facing and token-level primitives: block-aligned padding, span separator (SSep), and prompt depend (PDep), that modify hashing behavior and effective block-level causal attention structure. With fewer than 100 lines of core-engine changes plus a custom attention backend, these primitives are sufficient to realize multiple PIC methods, including Block-Attention, EPIC, and Prompt Cache, within the same running vLLM instance, while natively integrating with KV cache CPU offload implementations. On 2WikiMultihopQA, MiniPIC with interleaved scheduling improves prefill throughput by 49% over baseline vLLM, reduces cached-span time-to-first-token by up to two orders of magnitude, preserves the linear prefill scaling of uncached spans, and incurs only 5.7% worst-case overhead.

Code-writing large language models (CodeLLMs) generate executable code policies for embodied agents by translating natural language goals and environmental constraints into structured control programs. However, policy generation in open-domain embodied environments suffers from two fundamental limitations: (i) delayed decoding caused by repetitive prefill computation over long prompts, and (ii) limited robustness due to fully generative decoding, which often produces API mismatches, missing safety guards, and unstable control logic. To address these limitations, we present FCGraft, a Functional Cache Grafting framework. FCGraft maintains a library of function-level validated code skeletons and their associated prompt-level Transformer key-value (KV) caches, and synthesizes new policies by retrieving relevant functions and grafting their KV caches when a new task is provided. Given retrieved function caches, FCGraft performs cache grafting via stitching, which composes cached function segments into a composite policy, and patching, which locally adapts only the necessary code regions to satisfy task-specific parameters and constraints with minimal additional decoding. By eliminating redundant prefill computation, this approach reduces generation latency, while reusing validated control structures improves robustness over prompt-level caching methods RAGCache, achieving 18.31% higher task success rate and 2.3x faster policy synthesis.

Multimodal Large Language Models (MLLMs) achieve strong vision-language reasoning, but long visual contexts enlarge the KV cache and increase decoding latency. Existing compression methods rely on observation window attention for stable token-importance estimation, yet this aggregation can dilute sparse visual evidence and discard answer-critical tokens under aggressive compression. Therefore, we identify last-query attention as a complementary source for recovering such evidence, but its answer-irrelevant signals can mislead retention. We propose BACON, a plug-and-play method that calibrates observation window attention with last-query evidence and suppresses isolated noise via intra-layer coherence and inter-layer persistence. Across diverse benchmarks, models, budgets, and compression methods, BACON improves multimodal KV compression by 7.5% on average under the most aggressive budget, with gains up to 30.9%. Our project page is available at https://ryu1ion.github.io/official_BACON/

How far can we reduce the number of physical keys if we endow an ambiguous keyboard with modern language models? Fewer keys increase hardware design freedom in constrained settings such as assistive devices and mobile form factors. This paper systematically evaluates text entry systems using 2-5 physical keys combined with language-model-based disambiguation. On a 300-sentence English corpus (100 sentences each for Business / Conversational / Technical), we compare key counts (2-5), letter-to-key mappings (layout-based / frequency-based / intentionally worst-case), and decoders (Trie-only, GPT-2 beam search, GPT-4o selection). We find that 3 keys + GPT-4o achieves character error rate (CER) 9.46% and word error rate (WER) 12.20%, reducing CER by 59% relative to 2 keys (CER 23.3%). At 3 keys, the key-stream entropy is 1.54 bits/char; while increasing to 5 keys improves accuracy (CER 5.4%), the marginal gains diminish. Mapping choice has a small impact under standard designs (ΔCER < 0.5 pp), and even an intentionally worst mapping degrades CER by only +0.5 pp, whereas Technical sentences yield roughly twice the error rate of Business. These results suggest that, in our evaluated offline setting under a strong LM prior, 3 keys are a practical minimum for general English.

Long chain-of-thought (CoT) trajectories in large language model (LLM) reasoning cause severe inference bottlenecks due to rapid key-value (KV) cache growth. Current decoding-time compression methods mitigate this issue via token eviction, but typically assume a uniform budget distribution across all layers and heads. In contrast, existing non-uniform budget allocation methods are predominantly designed for the static prompt prefill phase, and they do not capture the stepwise context demands of autoregressive reasoning. To bridge this gap, we propose ReasonAlloc, a training-free framework that recasts decoding-time KV compression as a hierarchical budget allocation problem. ReasonAlloc operates at two complementary levels: an offline layer-wise preallocation strategy captures an architecture-driven demand pattern which we call ``\textit{Reasoning Wave}'', while an online head-wise strategy reallocates resources during decoding to information-rich heads based on real-time utility. Evaluations on mathematical reasoning benchmarks (MATH-500, AIME~2024) using DeepSeek-R1-Distill-Llama-8B, DeepSeek-R1-Distill-Qwen-14B, and AceReason-14B show that ReasonAlloc outperforms uniform-budget R-KV, SnapKV, and Pyramid-RKV (a baseline enforcing a static, monotonically decreasing layer budget), with the largest gains at small budgets (128-512 tokens). ReasonAlloc is plug-and-play with existing token-eviction policies and introduces negligible inference-time overhead.

Federated learning (FL) is often subject to aggregation variance if clients do not consistently participate in training rounds. While reusing stale model updates from inactive clients is a common technique to reduce this variance, we find that with skewed client participation, the resulting update staleness can become severe enough to destabilize training. To remedy this, we propose FedSteer, a novel method that constructs a gradient subspace from a cache of recent client gradients to serve as a low-dimensional representation of the current optimization landscape. FedSteer projects an active client's true gradient onto this subspace to find a set of optimal coordinates. For an inactive client, FedSteer reuses these coordinates with the now-evolved subspace drifted by other active clients. This process effectively "steers" outdated gradients toward the current global objective. This is complemented by a selective caching strategy that identifies a representative client subset to form the subspace, reducing server memory. Experiments demonstrate that FedSteer significantly outperforms baselines, preventing performance collapse in challenging scenarios while delivering accuracy gains of over 7% in others.

Long-context language model inference is bottlenecked by memory, as the KV cache grows with context length. Recent techniques to compress the KV cache fall short: they either degrade model quality substantially or require considerable time and compute to compress a single long prompt. Furthermore, many methods require the input to fit within the target model's context window, and are generally incompatible with modern production inference engines. Encoder-decoder compressors, which map a long token sequence to a shorter sequence of latent embeddings consumed by a decoder, are an appealing alternative in principle. However, existing approaches are not competitive with KV cache compression on the accuracy-efficiency frontier. In this work, we revisit encoder-decoder compression and close this gap. We first perform an architecture search, pre-training many variants from scratch to determine how best to design and train encoder-decoder compressors. Guided by our findings, we continually pre-train a family of 0.6B-encoder, 4B-decoder models on over 350B tokens each, at compression ratios of 1:4, 1:8, and 1:16. We introduce Latent Context Language Models (LCLMs), a family of compressors that improve the Pareto frontier across general-task performance, compression speed, and peak memory usage. We demonstrate that LCLMs serve as efficient backbones for long-horizon agents, letting the agent skim through a compressed long context and adaptively expand relevant segments on demand.

Multi-turn LLM agents interleave model calls with external tool invocations, shifting serving from stateless request processing to stateful program execution. Serving these workloads requires scheduling, KV-cache management, and routing policies that use program-level context, including turn dependencies, tool-induced gaps, and reusable KV state. Evaluating such policies directly on real systems is costly, since each design point may require dedicated accelerator time across arrival rates, model scales, serving-instance counts, and memory hierarchies. Simulation offers a scalable alternative, but existing LLM serving simulators target stateless request-level workloads and therefore omit the core dynamics of agent serving: multi-turn program execution, cross-turn cache locality, and KV-cache residency during tool gaps. We present AGENTSERVESIM, a hardware-aware simulator for multi-turn LLM agent serving. AGENTSERVESIM evaluates serving policies at program granularity through composable modules: a Program Orchestrator preserves program identity and turn order, a Tool Simulator materializes tool-induced gaps, a Session-Aware Router maintains program-to-instance affinity for cache-aware dispatch, and a KV Residency Model tracks policy-defined KV placement across HBM, host DRAM/CXL, and eviction. Across real serving deployments and hardware configurations, AGENTSERVESIM reproduces real-system behavior within 6% error across key performance metrics while running entirely on commodity CPUs. These results show that AGENTSERVESIM enables controlled, repeatable exploration of agent-serving policies without requiring exhaustive deployment on costly accelerators.

Existing sparse attention and KV cache compression methods for long-context LLM inference typically apply fixed sparsity patterns or uniform budgets across all attention heads, overlooking the substantial variation in attention behavior among heads and contexts. We observe two distinct entropy patterns among attention heads: Rigid Heads, whose entropy stays near zero across input segments, and Dynamic Heads, whose entropy fluctuates significantly. Crucially, the distribution of these types is context-dependent and cannot be predetermined offline. We therefore propose EntropyInfer, a training-free framework that uses attention entropy to adaptively allocate compute at the granularity of individual heads and segments during prefilling. For decoding, we introduce a latent KV cache compression scheme that leverages generated output tokens, rather than prefill tokens alone, to identify and retain the most critical cache entries. Extensive experiments on Llama, Qwen and openPangu model series show that EntropyInfer consistently outperforms baselines including SnapKV, AdaKV, and CritiPrefill, achieving up to 2.39$\times$ end-to-end speedup beyond 100k tokens with minimal quality degradation compared to full attention. The code is released in https://github.com/SHA-4096/EntropyInfer.

We introduce RKSC (Reasoning-Aware KV Cache Sharing), a training-free inference framework that eliminates two structural redundancies in multi-branch LLM reasoning pipelines. ASKS (Attention-Similarity KV Sharing) computes the prefix KV cache once and broadcasts it to all semantically similar branches via hidden-state cosine similarity, strictly generalising the token-exact prefix caching used by vLLM and SGLang. CGEE (Confidence-Gated Early Exit) applies two complementary exit mechanisms: (1) it skips the verification forward pass entirely when generation confidence is decisive across branches, and (2) it terminates the verification pass at an intermediate layer when per-layer entropy stabilises, using lightweight hooks on the transformer backbone. RSBCM (Reasoning-Selective Block Cache Manager) prevents unbounded cache growth via attention-weighted depth-priority eviction. Across five model families (7B-10B), four benchmarks, and 1,000 evaluated problems, RKSC achieves a mean speedup of 3.008x over the No-KV baseline (peak 3.990x), a 1.66x mean improvement over vLLM-equivalent prefix caching, with a CGEE-induced error rate of only 0.37% (6 errors out of 1,616 verify calls). No fine-tuning or architecture changes are required. Code is available at https://github.com/AnirudhSekar/RKSC.

Linear attention reformulates sequence modeling as recurrent state evolution, enabling efficient linear-time inference. Under the key-value associative paradigm, existing approaches restrict the role of the query to the readout operation, decoupling it from state evolution. We show that query-conditioned state readout induces a structured value prediction over accumulated memory that complements key-based retrieval. Based on this insight, we propose Q-Delta, a query-aware delta rule that integrates mixed key-query prediction errors into state evolution, enabling jointly corrective dynamics while preserving delta-rule efficiency. We establish stability guarantees for the resulting dynamics and derive a hardware-efficient chunkwise-parallel formulation with a custom Triton implementation. Empirical results demonstrate stable optimization, competitive throughput, and consistent improvements over strong baselines on language modeling and long-context retrieval tasks.

Prefill-decode (PD) disaggregation decouples prompt processing from token generation, but it also turns the key-value (KV) cache into a network payload. Existing PD-side KV reduction methods are mostly binary: selected tokens are transmitted at full precision and the rest are not transmitted. This paper argues that binary selection leaves a useful design space unused. SpectrumKV assigns a precision level to each token instead: attention sinks and other high-importance tokens are protected at FP16, medium-importance tokens are sent at INT8, and low-importance tokens are sent at INT4 when the model can tolerate it. The main practical complication is that INT4 tolerance is model-dependent. Qwen2.5-7B catastrophically fails under INT4 KV quantization, while Mistral-7B and Gemma-2-9B remain stable. SpectrumKV therefore runs a lightweight deployment-time probe: three aggressive NIAH trials under a 3-tier policy. Models that pass use FP16+INT8+INT4; models that fail fall back to FP16+INT8. Across Qwen2.5-7B-Instruct, Mistral-7B-Instruct-v0.3, and Gemma-2-9B-it, SpectrumKV improves quality at the same transfer budget. At a 50% normalized KV budget on WikiText-2, SpectrumKV changes perplexity by +1.97%,-0.06%, and-0.44%, respectively, compared with PDTrim's +25.85%, +22.07%, and +35.63%. On NIAH retrieval at 4096 tokens, the adaptive policy reaches 52.6% on Qwen at the aggressive b=0.3 budget versus 26.3% for PDTrim, and reaches 100% by b=0.5; Mistral and Gemma preserve retrieval under the 3-tier policy. End-to-end GPU timing of the transfer path shows 50-62% TTFT reductions at b=0.5. These results suggest that PD KV transfer should be treated as a precision-allocation problem, not only as token pruning.

Tensor networks provide efficient representations for compressing large neural networks. By carefully designing shapes and topologies, they can significantly reduce memory and computational costs. However, identifying implicit low-rank structures in large foundation models remains challenging due to their enormous scale and un-structured weight distributions. We propose an adaptive tensorization method that discovers inherent low-rank structure in a target tensor by index ordering. Experiments on weight and KV-cache compression demonstrate improved reconstruction quality compared to baselines.

Although video virtual try-on (VVT) has achieved significant progress, existing methods still exhibit two fundamental limitations: first, they are restricted to single-garment transfer, rendering simultaneous multi-object try-on highly impractical; second, their heavy reliance on explicit external priors (e.g., garment masks) inevitably destroys crucial physical dynamics and degrades visual quality. To bridge this gap, this paper proposes the novel Try-On Anything task, which aims to simultaneously transfer diverse wearable objects onto a person in a video in a single inference pass. To support and standardize this paradigm, we introduce TryAny-Bench, a comprehensive benchmark encompassing a paired video dataset alongside a tailored evaluation protocol. Furthermore, we present OmniTryOn, an external-prior-free generative framework designed to tackle this task. Specifically, OmniTryOn employs a First Frame Wearable Cache strategy, which directly provides diverse wearable objects for the generation process through the initial video frame. To maintain consistency, we propose the Spatiotemporally Consistent RoPE (STC-RoPE), which inherently establishes robust spatiotemporal anchors to strictly preserve complex human motions and background dynamics. Optimized by the proposed Gradual Try-On (GTO) training strategy, our model progressively masters robust multi-object synthesis. Extensive experiments on TryAny-Bench demonstrate that OmniTryOn significantly outperforms existing specialized video virtual try-on models and general video editing baselines, establishing a powerful new standard for the Try-On Anything task. Our dataset, code, and models are available at https://github.com/xcltql666/OminTryOn.

Low-rank projection has emerged as a promising approach for compressing the KV cache by exploiting hidden-dimension redundancy. However, prior methods rely on fixed or heuristic rank selection and struggle to achieve aggressive compression with minimal accuracy degradation. We propose STAR-KV, an adaptive low-rank KV cache compression framework with fine-grained rank control. STAR-KV encompasses 1) a differentiable thresholding mechanism that enables optimal rank selection at both attention-head and block levels, 2) a hybrid decomposition strategy that applies different low-rank factorizations according to the sensitivity of key and value projections, and 3) a low-rank-aware mixed precision quantization that leverages data statistics for near lossless low-bit quantization. Evaluated across multiple LLMs and benchmarks, STAR-KV achieves up to 75% KV cache compression and up to 20x overall KV cache reduction when combined with quantization. Enabled by custom Triton-based GPU kernels, STAR-KV delivers up to 6.9x speedup for the attention module and 3.1x end-to-end generation throughput. Our code is publicly available at: https://github.com/PriyanshBhatnagar/STAR-KV.

Visual Autoregressive (VAR) models adopt a next-scale prediction paradigm, offering high-quality generation with substantially fewer decoding steps. However, existing VAR models suffer from significant attention complexity and severe memory overhead due to the accumulation of key-value (KV) caches across scales. In this paper, we tackle this challenge by introducing KV cache compression into the next-scale paradigm. We begin with an in-depth analysis of VAR attention and observe that attention heads can be stably divided into two functionally distinct categories: Contextual Heads focus on maintaining semantic consistency, while Structural Heads preserve spatial coherence. Their functional divergence makes existing one-size-fits-all compression methods perform poorly on VAR models. We further find that the two head types differ markedly in their reliance on historical scales, and that this reliance shifts across layers and generation steps, arguing for an adaptive cache budget allocation. To address these challenges, we propose HACK++, a training-free Head-Aware key-value Compression frameworK for VAR models. From a one-time offline calibration, HACK++ classifies head types and derives head-specific priors. At inference, it decouples attention from cache compression under independent budgets, bounding the current-scale attention cost while compressing the accumulated cache far more aggressively, via pattern-specific strategies and a reliance-aware budget allocation. Extensive experiments on multiple VAR models across text-to-image, class-conditional, and unified understanding-and-generation tasks validate the effectiveness and generalizability of HACK++. For example, on Infinity-2B/8B, HACK++ maintains near-lossless generation with only a 30% attention budget and a 10% cache budget, and remains robust even under a 1% cache budget.

Multi-turn LLM agents fan short queries into long trajectories of tool calls, search results, and intermediate reasoning. Both KV memory and KV read bandwidth grow by orders of magnitude across a single trajectory, making the key-value (KV) cache, not parameter compute, the dominant serving bottleneck for long-horizon agents. We introduce IntentKV, learned KV pruning that keeps the base LLM frozen. IntentKV maintains a session-level QueryMemory of cross-turn intent, scores live history tokens with a memory-attention rule, and adds a zero-initialized residual head with cross-attention over current-query K-vectors. To stay composable with prefix caches, eviction is a slot-map redirection: dropped positions route to a sentinel dead slot while surviving K/V rows, RoPE phases, and slot identities stay in place. IntentKV matches the no-pruning full-cache baseline with almost no accuracy drop under tight KV budgets: at an 8k KV budget, mean peak request tokens drop 23.9% on Qwen3-8B and 30.7% on Qwen2.5-14B. On the 100 longest BCP queries that all methods complete on Qwen2.5-14B, IntentKV-8k further cuts worst-case peak request tokens from 92.3k to 20.5k, a 77.8% reduction, and worst-case raw KV reads from 411M to 31M, a 92.6% reduction.

The KV cache is the memory bottleneck of long-horizon language model deployment. Practically, a deployable compactor must be lightweight enough to call during inference, expressive enough to preserve context under constraint, and reusable across a trajectory. Existing compaction methods satisfy only part of this requirement: selection methods are lightweight but subset-bound, while synthesis methods are expressive but rely on per-context optimization. Here we introduce Still, a small per-layer Perceiver trained once against a frozen base model that produces compact keys and values in a single forward pass. On Qwen and Gemma models, Still occupies the favorable side of the speed--quality frontier across compression ratios from $8\times$ to $200\times$ and context lengths from $8$k to $128$k. On the long-context RULER grid, Still exceeds the strongest baseline by 8--22 points. The same compact cache also supports free-form summarization, preserving most of the full-context gain on HELMET and winning a pairwise LongBench summarization comparison against KV-Distill. Because compaction is a forward pass, Still can be applied iteratively, entering a long-horizon regime unavailable to per-context methods. We show that amortization makes long-context cache compaction tractable, and synthesis makes its compact state useful at extreme compression.

The attention mechanism is a cornerstone of modern transformer architectures. However, its expressive power comes at the cost of quadratic runtime and linear space usage. In particular, the classical transformer architecture explicitly stores all previously seen input elements (tokens) in order to generate the next one. The problem of implementing a transformer in limited space, known as KV cache compression, has received much interest over the past few years, spurring the development of powerful heuristics. Recent works of Haris et al, COLT'25 and Kochetkova et al, NeurIPS'25, formalized KV cache compression as the streaming attention approximation problem, providing both upper bounds (based on discrepancy theory) and information theoretic lower bounds. However, those papers left open a significant gap between the upper and lower bounds. For example, the space usage of their algorithms increases with the precision parameter, but the lower bound does not get stronger. In this work, we revisit the streaming attention approximation problem and provide nearly tight bounds on its space complexity. On the algorithmic side, we achieve the result through a surprisingly tight interplay between three distinct methods for kernel density estimation: discrepancy-based coreset constructions (e.g., Charikar-Kapralov-Waingarten'24), the polynomial method (e.g., Greengard-Rokhlin'87, Alman-Song'23), and space partitioning (e.g., Andoni-Laarhoven-Razenshteyn-Waingarten'17, Charikar-Kapralov-Nouri-Siminelakis'20). On the lower bound side, our main technical contribution is a new technique for using the INDEX problem with a large amount of side information that we hope will prove useful in other high dimensional geometric estimation problems.

Disaggregated serving alleviates memory bottlenecks in Large Language Model (LLM) inference but creates a severe communication bottleneck: transmitting high-dimensional Key-Value (KV) caches often dominates time-to-first-token (TTFT). Moreover, reusing caches across heterogeneous models (e.g., base and fine-tuned variants) causes semantic misalignment that accumulates over layers, degrading generation quality. We propose Semantic Cache Distillation (SCD), a loss-constrained framework that replaces raw KV transmission with compact semantic codes. SCD addresses these challenges via two mechanisms: (1) Reuse, which reconstructs most layers from low-rank subspaces to minimize transfer cost, and (2) Patch, which predicts normalized inputs at sparse transition layers to truncate error propagation. Empirically, SCD delivers up to 2.65 $\times$ TTFT speedup over the oracle consumer prefill and dominates quantization and selective recomputation baselines on the quality--latency Pareto frontier in bandwidth-constrained regimes, while keeping generation quality within 5% F1 of the oracle.

Multi-turn LLM serving accumulates dialogue history whose Key-Value (KV) cache grows with every turn and every user, quickly exceeding the model weights themselves and making memory -- not compute -- the binding constraint on throughput. Non-uniform KV compression, which allocates heterogeneous budgets across attention heads, preserves accuracy far better than uniform schemes, yet remains impractical: modern serving stacks assume identical KV lengths across heads, so heterogeneity traps freed memory as page fragmentation, spends up to 25% of prefill time reclaiming scattered pages, and skews GPU workloads that inflate decode latency by up to $1.7\times$ or burn 15--20% of each decode step on re-planning. We observe that this heterogeneity need not be discovered at runtime: head-wise retention follows a two-level structural regularity -- an input-invariant head ranking with narrowly bounded per-head ratios -- that can be calibrated offline from as few as 50 samples. Building on this insight, we present Tangram, a serving framework that statically resolves what prior systems handle dynamically: Budget Reservation fixes each head's post-compression footprint at scheduling time, eliminating page reclamation; Ragged Paging clusters similar-budget heads into independent page tables, turning fragmentation into reclaimable memory; and Ahead-of-Time Load Balancing precomputes balanced GPU partitions with zero runtime planning. Implemented on vLLM, Tangram serves as a drop-in substrate for existing non-uniform compression methods, matching their accuracy while improving end-to-end throughput by up to $2.6\times$ over the full-KV baseline. Our implementation is publicly available at https://github.com/aiha-lab/TANGRAM.

As the input length of large language model (LLM) serving continues to grow, the KV cache has become a dominant bottleneck in AI infrastructure. It limits GPU memory capacity, serving concurrency, cache reuse, and distributed scalability. Multiple important problems, including position-independent KV cache, prefix KV cache compression, hot/cold KV cache separation, and distributed KV cache management, all depend on how the KV cache is represented and managed. However, existing serving systems largely rely on a monolithic KV cache abstraction, where the KV cache is treated as a homogeneous sequence of token-level memory blocks and managed with similar policies across attention heads and serving scenarios. We observe that KV cache utility is highly structured across KV heads: different heads exhibit different functional roles, attention distances, and runtime importance. Therefore, a full KV cache is not always necessary for every head, token range, or serving scenario. We present RedKnot, a head-aware KV cache management system for LLM serving. RedKnot breaks the conventional monolithic KV cache abstraction by decomposing the KV cache along KV heads, whose importance and effective attention ranges vary significantly across serving scenarios. This head-level decomposition turns the KV cache from a monolithic tensor abstraction into a structured memory object, enabling RedKnot to uniformly support position-independent KV reuse, prefix KV compression, hot/cold KV separation, and distributed KV placement while preserving output fidelity and improving resource efficiency, without requiring model retraining or fine-tuning. RedKnot establishes a new foundation for AI infrastructure by transforming the KV cache from a monolithic, passive runtime artifact into a dynamic, model-aware runtime substrate for scalable LLM serving.

Modern diffusion models generate high-quality images and videos, but their iterative denoising process makes inference expensive. Feature caching accelerates sampling by reusing or predicting intermediate activations across neighboring denoising steps, exploiting the redundancy of computations along the reverse trajectory. In this work, we focus on the caching schedule: selecting which denoising steps should be fully recomputed. Existing schedules are either fixed (e.g. uniform) or chosen adaptively from per-step error heuristics; in both cases, the actual compute cost is a side-effect of hand-tuned thresholds rather than a quantity the user can specify. We propose ReCache, which inverts this: given a target budget k, it learns the recomputation schedule that maximizes generation quality, turning compute into a directly controllable input. ReCache trains via policy gradients, sidestepping backpropagation through full diffusion inference, and uses no labelled data. Generations from uncached inference serve as matching targets, paired with a reward for generation quality. ReCache is compatible with any caching mechanism, including feature reuse and feature forecasting; for each mechanism, a single trained policy adapts across computational budgets at inference time. ReCache consistently outperforms scheduling baselines: under a $\times5.04$ FLOPs reduction on FLUX, it reduces LPIPS by 31% (from 0.456 to 0.316) compared to DiCache; on Wan 2.1 at a $\sim \times2.6$ speedup, it drops LPIPS by 65% (from 0.480 to 0.169) and boosts the VBench score by 7% (5.6 points, from 70.4 to 76.0) over uniform HiCache. Code is available at https://github.com/thecrazymage/ReCache.

Retrieval-augmented generation (RAG) improves large language model (LLM) answer quality by grounding generation in external evidence, but processing retrieved contexts makes the prefill stage a dominant serving cost. RAG cache fusion reduces this cost by reusing precomputed key-value (KV) caches for retrieved chunks and selectively recomputing tokens under the current prompt. Existing selectors, however, face a dilemma between quality and efficiency: fast query-agnostic or final-layer query-to-context selectors can miss request-relevant evidence, whereas full-view query-aware selectors require broad context and layer visibility before recomputation and therefore stall the layer-wise cache-fusion pipeline. We present QCFuse, a compressed-view query-aware selector for RAG cache fusion. QCFuse uses chunk-anchor query probing to condition user-query states on compact per-chunk anchors and critical-layer profiling to identify recomputation tokens without all-layer inspection. We implement QCFuse in SGLang and evaluate it on four open-weight LLMs across six datasets. QCFuse reaches full-prefill-level quality. At matched quality, QCFuse achieves an average prefill-time speedup of 1.7x over full prefill and 1.5x over ProphetKV, the strongest quality-preserving baseline.

Parametric retrieval augmentation encodes document information into lightweight, document-specific modules such as LoRA adapters, reducing the need to include all evidence as input context. However, it remains unclear how this parameter-side memory interacts with context-side memory stored in the KV cache. We study this interaction in document-level question answering by progressively evicting document key-value states and measuring when a document LoRA contributes beyond the retained context. We find that document LoRA adds little when the KV cache is largely intact, but becomes increasingly useful under aggressive compression, recovering 13-21 ROUGE-L points when no document context remains. The gain is largest when the base model encodes the document, and the adapter is applied only during answer generation, suggesting that document LoRA is better understood as decoding-time parametric memory than as a document encoder. Finally, QA-style supervision produces substantially stronger adapters than raw-context next-token-prediction. These results position document LoRA as a complementary memory channel whose value emerges precisely when context-side evidence is scarce.

Self-attention selects information freely across the sequence, but across depth, Transformers merely add each layer's output to the residual stream, so later layers cannot selectively reuse earlier-layer representations. Recent cross-layer methods improve this flow but operate on hidden states outside attention, adding state beyond the key-value cache at inference--a cost that becomes increasingly salient as modern LLMs compress the cache with grouped-query and multi-head latent attention. We introduce Depth-Attention, which performs this selection inside the attention module itself: before a layer attends over the sequence, its query attends over the keys of earlier layers at the same token position and mixes their values into the value that self-attention then reads. Because Depth-Attention reuses the standard attention queries, keys, and value-cache slots, storing depth-mixed values in place of the original values, it adds no parameters and introduces no persistent inference state beyond the standard key-value cache--the same cache size as a vanilla decoder and less than hidden-state-based cross-layer methods. On Qwen3-style decoders at 1.5B and 3B parameters, Depth-Attention attains the lowest perplexity and the highest average downstream accuracy, improving over the vanilla Transformer by up to 2.3 accuracy points and surpassing strong cross-layer baselines in perplexity and average accuracy, while adding under 0.01% extra arithmetic FLOPs and no additional persistent inference state. The gains hold from 360M to 3B parameters and extend to looped Transformers.

Large Language Models can reason over long contexts, yet prefilling millions of tokens is wasteful as much of the content remains static across queries. Cartridges address this by distilling document collections into reusable key-value (KV) caches that eliminate prefilling while preserving accuracy. A critical limitation of this approach is that cartridges are monolithic and non-compositional: encoding an entire collection into a single KV block does not scale, and naively mixing cartridges trained in isolation collapses performance to near chance. We introduce Cartridges at Scale (CAS), a training framework for scalable multi-cartridge learning with dynamic distractor mixing and a memory-efficient budget manager that rotates hundreds of per-document cartridges between GPU and persistent storage. Our approach scales to collections exceeding a million tokens, improving over a monolithic cartridge by 10-31 points at comparable token budgets. Oracle cartridge accuracy falls within 2-6 points of full in-context learning even at high compression. When paired with retrieval for cartridge selection, CAS matches or exceeds conventional RAG accuracy while consuming 3-4x fewer prompt tokens.

Reasoning models improve accuracy through extended chains of thought, but their long outputs create a memory and compute bottleneck. KV cache eviction methods reduce this cost by evicting unimportant key-value pairs from the cache, yet they often yield worse accuracy than selection-based sparse attention alternatives, which keep the full KV cache. We identify key factors crucial to KV cache eviction accuracy. First, a small fraction of value states have abnormally large magnitudes, and evicting them causes catastrophic failure where models enter repetitive reasoning loops. Second, introducing stochasticity during eviction improves accuracy by increasing cache diversity. Based on these findings, we propose Value-aware Stochastic KV Cache Eviction (VaSE), a training-free recipe that protects large-magnitude value states and promotes diverse eviction decisions. Across six reasoning tasks, Qwen3 models using VaSE with 4x KV cache compression yield higher average accuracies than SOTA selection method at the same sparsity, while outperforming the strongest eviction method by more than 4%. Overall, VaSE bridges the gap between efficiency and accuracy, supporting FlashAttention2 and enabling a static memory footprint for reasoning models.

Disaggregated LLM inference forces the KV cache to traverse the datacenter network before decoding begins, so transfer time enters directly into the Time to First Token (TTFT) budget. Current schedulers route on compute load and prefix-cache locality alone, ignoring the topological distance and dynamic congestion between prefill and decode instances. We close this gap with a thin operator-to-scheduler interface, the network cost oracle, and we prove that ignoring the network term renders cache-aware-only scheduling arbitrarily suboptimal as context length grows. NetKV, the O(|D|) per-request greedy that consumes this oracle, has tier rankings that are provably robust to stale telemetry. On a 64-GPU four-tier fat-tree simulator driven by Mooncake traces, NetKV reduces mean TTFT by up to 21.2% over round-robin and 17.6% over a tuned cache+load-aware scheduler, lifts SLO attainment by up to 20.1 percentage points, and keeps the Time Between Tokens overhead below 0.5 ms in every condition tested, with no changes to the transport, inference engine, or hardware.

Test-time scaling is a powerful approach to obtain better reasoning in large language models, but it becomes memory-bottlenecked during long-horizon decoding, as the KV-cache grows. KV-cache quantization can help improve this, but current methods are evaluated under prefill-like settings and errors behave differently under autoregressive decoding. We show that in the latter regime, quantization errors accumulate across timesteps, driven primarily by incorrect token scales. We introduce KVarN, a calibration-free KV-cache quantizer that applies a Hadamard rotation followed by a dual-scaling variance normalization across both axes of the K and V matrices. We find that this combination fixes outlying token-scale errors and substantially reduces error accumulation over existing baselines. KVarN establishes a new state-of-theart for KV-cache quantization on generative benchmarks, including MATH500, AIME24 and HumanEval, at 2-bit precision. A vLLM implementation of the KVarN method is available at https://github.com/huawei-csl/KVarN

Vision-Language Models (VLMs) inherit the auto-regressive generation paradigm and cache the keys and values (KV) of all previous tokens to accelerate inference, resulting in memory consumption that scales linearly with context length. This issue is particularly pronounced in VLMs due to substantial redundancy in the visual modality. Although KV cache eviction approaches can effectively reduce inference memory, they often incur significant performance degradation in VLMs, as most are designed for language models and overlook the inherent gap between text and vision. By systematically analyzing the modality gap in VLMs in this work, we argue that the importance of visual information should be grounded in textual guidance and accordingly propose a Text-Grounded KV Eviction method for VLMs (TGV-KV). TGV-KV comprises three submodules: (1) Text-Vision Budgeting (TVB) assigns budget to each layer based on the mutual information interaction. (2) Text-Weighted Ranking (TWR) assesses the priority of text and ranks vision importance based on weighted text-image attention. (3) Text-Prioritised Retention (TPR) policy strategically preserves text KV to avoid acute information loss. We evaluate TGV-KV across five models with different sizes and architectures, showing that TGV-KV preserves 99.2% full-KV accuracy on the VizWiz-VQA task with LLaVA-NeXT and boosts end-to-end throughput by 52.6% with an extreme retention budget of 5%. Code is available at https://github.com/Danielement321/TGV-KV.

Large Language Model (LLM) inference relies on key-value (KV) caches to avoid redundant attention computation. While approximate KV cache retention techniques reduce memory usage by sacrificing model accuracy, lossless approaches instead evict KV cache blocks from GPU memory and reconstruct them on demand to preserve exact outputs. Existing lossless KV cache management systems primarily base eviction decisions on access frequency or positional heuristics, without considering how different KV cache blocks affect the execution efficiency of GPU attention kernels. In this paper, we propose AsymCache, a computation-latency-aware KV cache management system for LLM inference that explicitly aligns cache residency decisions with GPU attention kernel performance, including three key components: Multi-Segment Attention (MSA) for efficient non-contiguous KV context processing, a cache eviction policy that jointly optimizes hit rate and position-aware recomputation cost, and an adaptive chunking scheduler for high hardware utilization. Experiments show that AsymCache reduces TTFT by up to 1.90-2.03x and time-per-output-token (TPOT) by 1.62-1.71x over latest baselines, confirming the effectiveness of the method in common workloads and validating its design goal of balancing computational efficiency with cache hit rate. Moreover, the low-level design of AsymCache allows seamless integration into agent serving systems such as Continuum, where it further reduces average job latency by up to 18.1%.