Dynamic Sparse Attention

Long-horizon and multi-turn agents typically generate short actions and process long observations from tools and environments. This growing context demands efficient prefill, compact KV-cache storage, and accurate long-context retrieval. To meet these demands, we introduce HySparse2, a hybrid sparse attention architecture with two-level KV sharing. At the outer level, KV Bridging adopts a YOCO-style self-decoder and cross-decoder structure, but bridges only full-attention layers. The self-decoder uses hybrid sliding-window attention (SWA), while the cross-decoder uses hybrid sparse attention. The KV caches for full-attention layers in the cross-decoder are generated from the hidden states of full-attention layers in the self-decoder. At the inner level, HySparse2 retains HySparse's core KV Reuse design with two refinements. First, it replaces block-level sparsity with token-level sparsity for finer long-context retrieval. Second, it removes the separate SWA branch from sparse layers and instead forces a sliding window of recent tokens into the sparse selection. This two-level KV sharing allows all cross-decoder KV caches to be constructed from self-decoder hidden states. Prefill can therefore exit after the self-decoder, skipping all cross-decoder layers. On an 80B-A3B MoE model, HySparse2 outperforms HySparse and Hybrid SWA on long-context retrieval and multi-turn agentic tasks, while substantially reducing prefill computation and KV-cache storage.

Despite their strong performance, large language models (LLMs) are bottlenecked by KV cache memory traffic during long-context inference. Sparse attention is widely used to accelerate LLM inference by computing exact attention over a selected subset of tokens. To recover the contribution of tokens excluded from exact attention, recent methods apply coarse-grained compensation to the omitted attention tail. However, existing methods typically select tokens based on attention mass and only then compensate for the unselected tokens. This decoupled design overlooks their interaction: selection should prioritize tokens that would leave the largest compensation error if omitted. To address this limitation, we introduce CompKV, the first compensation-aware sparse attention framework that divides tokens into blocks and explicitly optimizes selection for the downstream compensation mechanism. Our theoretical analysis shows that the residual left by block-level mean compensation is governed by both block attention mass and within-block logit variation. We approximate this residual using compact block-level statistics, yielding a deployable selection criterion. We further develop an efficient asynchronous implementation. Experiments on RULER and LongBench-Pro show that CompKV performs best among the evaluated sparse baselines while delivering up to a $6.85\times$ self-attention speedup over full attention.

Autoregressive image generation has emerged as a paradigm for multimodal AI systems due to its compatibility with transformer-based LLM serving infrastructures. However, generating thousands of visual tokens per request makes decoding increasingly bottlenecked by KV cache accesses during attention computation. Sparse attention is particularly attractive for this workload because many visual generation applications tolerate moderate quality degradation in exchange for improved performance and efficiency. While sparse attention has been extensively explored for text-based LLM inference, it remains unclear whether its sparsity assumptions generalize effectively to autoregressive image generation. We present the first systematic characterization of attention sparsity in autoregressive image generation across diverse workloads and representative open-source models. Our analysis reveals several distinguishing properties, including a pronounced prefill-decode asymmetry, strong attention concentration on prompt and local tokens, and a unique diagonal attention sparsity pattern arising from the spatial locality of visual tokens. Motivated by these observations, we propose a diagonal-aware sparse attention mechanism that selectively skips KV entries along the diagonal attention direction within a recent window. Implemented on top of a GPU-based serving system using FlexGen, FlashAttention-2, and custom kernels, our approach achieves up to 3.1x throughput and 1.19x latency improvements with less than 2% quality degradation compared to dense inference.

Diffusion Transformers deliver state-of-the-art video generation, but their long spatiotemporal sequences make attention the dominant deployment cost, and a deployable low-bit kernel must be accurate and fast. Accuracy is limited by outliers: a block's quantization scale is set by its largest entries, leaving typical entries confined to a narrow range of representable values. Prior work smooths queries and keys, but value outliers follow no fixed channel or spatiotemporal structure and remain the dominant source of output error. Speed is limited by softmax: low-bit Tensor Cores accelerate only the two matrix multiplications, so the high-precision exponential between them becomes the longest pipeline stage on datacenter GPUs. We propose VC-Attention, a training-free low-bit attention framework that addresses both by pairing Value smoothing with a fused probability Cast. V-Smooth reorders value tokens by lightweight online clustering, so the tokens in a hardware block quantize well together. It quantizes only the residual after subtracting the block mean, and restores that mean from the row sum the online softmax already maintains. ExpCast-FP8 maps log-domain scores directly to E4M3 probability codes with one fused multiply-add, eliminating the FP32 exponential and the format conversion. We implement VC-Attention for B200, B300, H200, RTX PRO 6000, and RTX 5090. Across Wan2.2, LongCat-Video, HunyuanVideo-1.5, and MiniMax-H3, VC-Attention improves fidelity over low-bit baselines, speeds up the attention kernel over BF16 FlashAttention-4 by 1.46-1.59x on datacenter Blackwell and Hopper and by 2.3-3.6x on workstation cards, and generates a clip 1.13-1.19x and 1.36-1.70x faster end to end.

Post-training attention sparsification reduces the quadratic cumulative attention cost of pretrained Transformers by selecting a small set of context units (tokens or blocks) for each query. Existing trainable methods usually use a lightweight selector to score context units, followed by hard Top-K selection that blocks gradients from the language modeling loss. Consequently, these methods commonly distill layer-wise dense attention distributions. Although this encourages the selector to rank context units by dense attention weights in the original model, the ranking is not directly aligned with their impact on predictions under a fixed attention budget (i.e., the number of attended context units per query), potentially wasting the limited budget on less useful units. To address this misalignment, we propose Simple Attention Sparsification (SAS), a gated sparse attention mechanism that optimizes context ranking end-to-end with the language modeling loss. The key idea is to inject the selector's continuous scores into attention logits during training, allowing the loss to update the selector through standard backpropagation. We identify several choices crucial for this simple design to work well in practice: placing the gate inside the attention softmax in log form, using normalized softmax gates to calibrate historical context against the always-retained current block, and preserving continuous selector scores so the model learns relative priorities rather than only hard selections. To support long-sequence training, we implement a memory-efficient Triton kernel that integrates SAS into FlashAttention-style computation. Across reasoning, long-context understanding, and agentic tasks, SAS consistently outperforms trainable sparse attention baselines across attention budgets, with especially large gains under tight budgets, demonstrating more effective context ranking for downstream tasks.

Dynamic sparse attention reduces long-context prefill cost by routing each query chunk to a small set of key chunks at every Transformer layer. The sparse attention kernel avoids most token interactions, but the router still rebuilds a chunk--chunk score matrix layer after layer, even when the selected routes change little. We introduce RouteRelay, a router-agnostic method that reuses only route metadata across depth while continuing to compute attention with the current layer's queries, keys, and values. Anchor layers perform full routing. Intermediate layers rescore the previous top-$k$ route and a compact sentinel set of near-miss and randomly probed chunks. A query row is rerouted only when a sentinel challenges its weakest selected chunk. We give a top-$k$ stability condition, a probabilistic bound on missed challengers, and a row-selective GPU execution design. In a reproducible empirical evaluation, RouteRelay retains at least 99.99% route recall while rerouting 25.0%, 55.4%, and 78.2% of rows under low, moderate, and high cross-layer drift, respectively. Across routing scales, RouteRelay retains 100.0% recall while evaluating 38.4--51.6% of full-routing score pairs as the key-chunk count grows from 128 to 1024. Its unfused CPU execution remains slower than dense matrix multiplication, exposing row compaction and ledger updates as the main kernel-engineering targets.

Post-hoc sparse attention accelerates long-context prefill by routing each query to a small set of token-level interactions. Hard selection, however, assigns zero probability to every omitted chunk: a routing miss cannot be recovered, and a fixed expansion budget spends the same work on easy and ambiguous queries. We introduce Coarse-to-fine Error-aware Dynamic Attention Routing (CEDAR), a coarse-to-fine method that keeps the language model frozen while preserving global coverage. Each semantic chunk contributes a cheap key--value summary to a residual attention path; chunks with high estimated approximation error are then expanded to exact token attention. Exact and summarized contributions are combined in a single softmax normalization, so refinement replaces, rather than duplicates, coarse evidence. We derive an output-error bound governed by within-chunk key/value dispersion and use it to allocate a variable refinement budget. A controlled clustered-attention study shows that residual summaries reduce reconstruction error by more than 98% relative to hard dropping at equal exact-chunk budgets. Experiments on long-context benchmarks demonstrate that CEDAR recovers most of the quality lost by hard sparse routing while maintaining approximately $3\times$ kernel speedup at 128K context.

The attention prefilling phase of long-context LLM inference scales quadratically, making self-attention a severe computational bottleneck. Traditional sparse attention methods mitigate this through fixed patterns or offline profiling, but lack the flexibility to adapt to input-dependent attention structure. Recent dynamic methods address this by routing heads to sparse patterns in real-time, but rely on indirect routing proxies with overhead and budget allocation mechanisms that overlook the post-softmax mass hierarchy. We present CRISP (Cliff-awaRe Input-adaptive Sparse Prefilling), which identifies and addresses two structural challenges in this dynamic routing paradigm. First, we show that the routing decision can be read directly off the structure of the proxy attention map. We replace the Jensen-Shannon Divergence (JSD) routing with C_struct, a structural proxy that measures mass at Vertical-Slash compatible positions and reproduces JSD's routing decisions while eliminating both the pooled matmul and subsequent KL divergence overhead. Second, we formalize the post-softmax mass cliff and demonstrate theoretically that strictly cumulative coverage thresholds accumulate O(n) background noise at long contexts. CRISP navigates this via a sink-aware threshold grounded in the noise floor. Empirically, across InfiniteBench, RULER and LongBench on two model families, CRISP is the strongest sparse method overall and matches or exceeds exact dense attention on retrieval-heavy benchmarks, recovering up to +28.0 pp on retrieval tasks over baselines and achieving up to a 5.30x attention speedup at 512k tokens, driven primarily by our O(n) noise elimination during selection while preserving structural integrity.

The development of long-context Large Language Models (LLMs) is constrained by the memory bandwidth bottleneck and quadratic complexity of the attention mechanism during decoding. To overcome the inherent trade-offs between the memory overhead of metadata-based metrics and the computational inefficiency of adaptive selection strategies, we present Faster Flash Decoding (FFD), a novel hardware-algorithm co-design framework designed to break the memory wall in long-context decoding. FFD integrates the selector and computer into a fully fused kernel, replacing external metadata indices with content-aware scanning via low-bit quantization. Furthermore, we introduce the top-delta strategy, which dynamically filters blocks to achieve distribution-adaptive sparsity without global synchronization. Offering a training-free and plug-and-play solution, FFD also enables the reuse of scanning results for computation, achieving up to 11.6x kernel-level speedup and scaling to 256K context length, with 2.37x end-to-end throughput improvement. Empirical validation on RULER and LongBench confirms that FFD maintains model accuracy while delivering high-ratio sparsity, with code available at https://github.com/qluoluo/faster-flash-decoding

We describe the architecture and ablations of Qwen3.8-Flash-Next, a sparse mixture-of-experts model with 125B parameters, 6B activated per token, and additional 51B parameters of n-gram embedding tables held off the accelerator. On fourteen pre-training benchmarks the model leads the 397B-A17B predecessor on eight and trails it on the rest by at most 2.6 points, at 1/3 the activated parameters, 1/3 the training tokens, and roughly 1/9 the training FLOPs. Token mixing uses a layer-wise hybrid of Gated DeltaNet (GDN) and global attention, with one full-attention layer in every four; at continued-pretraining time those full-attention layers are replaced by Qwen Sparse Attention (QSA), which scores context at micro-block granularity with a compressed lightweight indexer. The residual stream is widened to four branches and read through an elementwise gate, a design we call the Gated Residual (GR). Capacity is added outside the backbone by a single n-gram embedding layer whose tables are prefetched from host memory. We evaluate every candidate change along three axes: loss together with downstream benchmarks; the cost of the change in training, prefill and decode; and its effect on the optimal hyperparameters and training stability. Loss and downstream accuracy do not always move together: enlarging the n-gram vocabulary lowers loss monotonically while downstream accuracy saturates. The architecture and the Muon optimizer together shift the optimal learning rate and batch size upwards, render batch-size warmup unnecessary, and substantially improve stability under stress tests. Loss, benchmarks, efficiency and stability form one design problem. Solved jointly, they yield a recipe that is simultaneously more efficient, more capable and more stable.

Joint text-to-video-audio generation produces synchronized visual and acoustic content, but the long sampling trajectories and heterogeneous multimodal computation of large models make inference prohibitively expensive. We present TurboT2VA, a distillation and inference framework for accelerating a 19B-parameter joint video-audio model. Large-scale T2VA distillation is challenged by modality-imbalanced optimization, the difficulty of continuous-time consistency training at scale, and the quality--diversity trade-off. TurboT2VA addresses these issues with per-modality normalization and a progressive curriculum comprising discrete consistency warm-up, continuous consistency refinement, and joint consistency--distribution matching. The curriculum first establishes a stable, diverse generation trajectory and only then introduces distribution-level refinement. On LTX-2, four-step distillation reduces generator latency from 50.52s to 2.51s at the standard evaluation resolution of 512$\times$768, achieving a 20.1$\times$ speedup while maintaining strong visual quality, audio fidelity, diversity, and video-audio synchronization. We further develop an architecture-aware inference stack that combines guarded W8A8 and fused operators, padded-text compaction, and modality-aware sparse attention while preserving dense cross-modal and text-conditioning paths. Under the high-resolution deployment setting at 1024$\times$1792, the complete stack reduces generator latency from 318.74s to 5.83s on one NVIDIA H20, achieving a 54.67$\times$ generator-only speedup. Inference code and generation demos are available at https://github.com/thu-ml/TurboDiffusion/tree/main/turbot2va.

Diffusion Transformers (DiTs) incur quadratic self-attention cost over spatiotemporal tokens. Existing training-free sparse attention methods often construct sparse masks from block-level or cluster-level proxy scores, which can obscure fine-grained differences among keys and miss high contribution keys under aggressive sparsity. Moreover, such proxy scores may yield overly concentrated softmax distributions, causing Top-$p$ to retain too few keys for some query clusters. Although a fixed Top-$k$ minimum alleviates this failure mode, a shared value cannot adapt to variations across heads and inputs. To address both limitations, we propose SCOPE, a training-free sparse attention framework that combines 3D-RoPE-aligned key subspace clustering with online per-head Top-$k$ estimation for efficient video-DiT inference. SCOPE partitions post-RoPE keys into temporal, height, and width subspaces, clusters them independently, and aggregates the corresponding centroid scores through lookup tables to obtain per key proxy scores for each query cluster. Building on existing hybrid Top-$p$/fixed Top-$k$ selection, SCOPE derives a head-specific Top-$k$ value online by averaging the initial retained key counts within each head, weighted by query cluster size, and selects additional keys only for query clusters whose initial retained key counts fall below this value. Sparse attention is then computed over the selected original keys and values. Across six model--task configurations, SCOPE consistently outperforms existing training-free baselines in both fidelity and latency, achieving up to a $1.99\times$ end-to-end speedup on 720p HunyuanVideo with $28.46$ dB PSNR relative to dense attention.

Video diffusion transformers are costly to sample: every denoising step applies self-attention over a long 3D token sequence, a quadratic cost that dominates as resolution and duration grow. Sparse attention reduces this cost without retraining, but existing methods pursue aggressive sparsity, where further speedup costs disproportionately more attention fidelity. We target the opposite end of this trade-off: fix near-lossless fidelity by construction, and remove as much computation as this constraint permits. Two observations make this regime practical: roughly 40% of block interactions can be removed while retaining 99% of the attention mass, and the high-mass support remains stable across denoising steps. We propose LoSA, a training-free sparse-attention method that fixes a retained-mass threshold of 99% rather than a sparsity ratio: it measures exact block attention masses at one early dense step, keeps, for each head and query block, the smallest key/value block set meeting the threshold, and reuses the frozen block indices for all remaining steps. On Wan2.1-1.3B, LoSA alone gives a $1.36\times$ speedup with a 0.06-point VBench Overall drop. The benefit is largest under composition: combined with feature caching, LoSA reaches a $3.2\times$ speedup on HunyuanVideo at a 0.02-point drop, versus 0.32 points for the strongest sparse baseline at comparable speed. Across three video diffusion transformers and speedups up to $3.2\times$, LoSA consistently achieves the best training-free speed-quality trade-off.

Long-context LLM inference is bottlenecked by quadratic attention computation and growing KV-cache costs. Existing sparse attention and KV-compression methods typically decide which tokens or heads to preserve from runtime attention scores, observation windows, calibration prompts, or learned gates, making head diagnosis input-dependent and costly to deploy. We propose Autonomy-of-Heads (AoH), a data-free method that identifies retrieval and streaming heads from the spectral geometry of query-key projections. AoH defines the kernel attention operator $M_h = W_K^{h\top}W_Q^h$ and uses its effective-rank as a weight-space measure of head function: concentrated spectra indicate a small number of dominant query-key matching directions and are associated with retrieval heads, whereas diffuse spectra indicate the absence of a dominant global matching direction and are associated with streaming heads. We further derive an efficient $d_\text{head}$-dimensional computation that avoids constructing the full $d_\text{model}\times d_\text{model}$ matrix. We conducted extensive experiments across models demonstrating that at 50% sparsity, AoH retains 96.5% of Full Attention performance on average while reducing prefill and decode latency by up to 41.4% and 66.0%, respectively, and KV-cache memory by 50.0% at 256K tokens.

Frame selection is essential for applying Large Multimodal Models (LMMs) to long videos due to severe frame redundancy and limited context windows. Since the appropriate frame budget varies with the downstream LMM, reasoning demands, and latency constraints, a practical selector should serve multiple budgets. However, existing methods typically optimize an isolated frame subset for each predefined budget: when the budget changes, previously selected evidence may be replaced rather than progressively augmented. Ranking frames by a fixed score would allow prefix reuse across budgets, but it ignores the distinct roles of different ranking positions. In this paper, we formulate long-video frame selection as a Matryoshka ranking problem: constructing a single priority sequence whose small prefixes concentrate query-conditioned evidence, while progressively larger prefixes preserve this evidence and add broader temporal context. Efficiently constructing such a ranking is itself challenging, as densely sampling long videos and evaluating frame-query relevance incurs substantial overhead. We therefore introduce Matryoshka Evidence-to-Context (MEC) Frame Selection, a training-free framework that builds a reusable sparse video index, discovers candidates through sparse probing and local zooming, and greedily constructs a position-adaptive ranking: early positions emphasize evidence; later positions progressively favor temporal coverage while preserving visual diversity. A single ranking can thus be truncated to any target budget without rerunning the selector. Across four benchmarks and six frame budgets, MEC improves average accuracy over uniform sampling by 3.77 percentage points, matches strong state-of-the-art selectors, and reduces end-to-end selection latency by 47.37-51.19%.

Long-context large language models (LLMs) are increasingly deployed in real-world applications, yet self-attention remains a major efficiency bottleneck -- especially during decoding -- due to the necessity of repeatedly processing ever-growing key-value (KV) caches. Existing sparse attention reduce computation by attending to fewer KV pairs, but often suffer from substantial accuracy degradation, require additional training, or rely on expensive hashing. In this work, we present BinaryPC, a training-free, data-aware hashing-based sparse attention for long-context LLMs. BinaryPC constructs compact binary hash codes and corresponding hash function by computing binary principal components of data. Unlike Locality-Sensitive Hashing (LSH) with data-independent random projections or learned non-linear hashing methods, BinaryPC constructs binary codes that explicitly preserve the structural information of data without requiring gradient-based training. Comprehensive experiments across multiple model families and long-context benchmarks show that BinaryPC preserves accuracy relative to full attention while achieving superior performance among sparse and hashing-based baselines. On modern GPUs, BinaryPC improves end-to-end decoding throughput by 3.56$\times$ over the FlashAttention kernel. Our code is available at https://github.com/yudaohai666/BPC.

Video diffusion transformers (vDiTs) generate high quality but pay quadratic self-attention cost, making inference prohibitive at video-token scales. The challenge is input-adaptive sparsity: selecting critical Q/K/V tokens with negligible overhead and executing them for end-to-end gains. We present SPADE, a training-free sparse-attention engine of three parts: (i) vDiT-SSR, a specification defining 3D blocking candidates and formalizing dynamic masks via Summarizer/Estimator expressions; (ii) runtime scheme generation using SICS and a head-wise policy; and (iii) an executor with low-overhead index search, flash block-sparse attention, and kernel grouping. Across Hunyuan-Video and Wan 2.1/2.2 for text-to-video and image-to-video generation, SPADE raises sparsity and speed while preserving quality, accelerating attention by 2.26x-3.40x and end-to-end inference by 1.49x-1.80x. Our code is open-sourced at https://github.com/6somehow/DAC-SPADE.

Video Diffusion Transformers (VDiTs) enable high-fidelity generation but incur quadratic cost from dense 3D self-attention. Existing head- and block-level sparse methods share computation budgets across queries, overlooking token-specific attention demand. We observe that retained density varies across queries yet correlates log-linearly with attention entropy, while dominant interactions form query-centered neighborhoods with token-dependent radii. Based on these findings, we propose Token Radius Attention (TRA), a training-free framework that maps query entropy to an analytic token budget and converts it into a temporally decayed radius without explicit key ranking. Fused entropy extraction, warm-up reuse, and block-sparse mask construction further reduce overhead. Across seven Wan2.1, Wan2.2, and HunyuanVideo T2V/I2V configurations, TRA retains only 9-19% of attention interactions and achieves 1.56x-2.05x speedup with competitive generation quality. Code is available at https://github.com/IF-LAB-PKU/Token-Radius-Attention.

Sparse attention is widely deployed in long-context serving stacks, yet no framework audits how discarding blocks changes the influence of specific content on model output. We first establish that the phenomenon is real and causal: Block Sparse Flash Attention (BSFA) route replay across four architectures changes output decisions in 13 of 16 cells, with zero identity-replay label flips. We then introduce a dense-calibrated counterfactual audit using matched probe cards---Gold (carrying the correct answer label), Poison (carrying a target wrong label), and Benign (filler only)---under six-layout position symmetry, isolating the sparsification-specific effect. Two patterns compete. Signal concentration: the selector preserves Gold and Poison blocks far above filler-matched Benign blocks (G$\approx$P$\gg$B across all model--task pairs). Integration loss: discarding blocks severs cross-block attention---confirmed by an ablation where isolating the probe block collapses its influence from 4.48 logits to zero. Compression ratio governs the balance: a full sweep from mild ($c=0.25$) to aggressive ($c=0.75$) compression across four model--task pairs reveals that three of four cells move toward stronger sparse amplification at higher compression, with two exhibiting sign reversals. Three independent arms---BSFA route replay, controlled block-top-$k$, and KV-cache eviction---converge: sparsification changes content influence in ways aggregate accuracy cannot detect. We provide an open measurement framework deployable on any model exposing block identities.

DeepSeek Sparse Attention (DSA) enables efficient long-context modeling through its Lightning Indexer. However, practical deployment remains constrained by the indexer's expensive $O(L^2)$ scoring overhead and the hardware-inefficient, discontinuous memory-access patterns induced by its outputs. To address these system-level bottlenecks, we introduce LongCat Sparse Attention (LSA), a hardware-algorithm co-designed framework comprising three complementary and orthogonal strategies: (1) Streaming-Aware Indexing, which selectively converts scattered KV entries into hardware-aligned contiguous layouts to enable coalesced HBM access; (2) Cross-Layer Indexing, which amortizes indexing overhead by reusing the results produced by a single layer across consecutive layers, supported by cross-layer distillation; and (3) Hierarchical Indexing, which adopts a coarse-to-fine scoring scheme to progressively narrow the candidate set for each query, thereby substantially reducing indexing computation. Extensive scaling experiments, ranging from 69B-A3B to 560B-A27B models, demonstrate that LSA consistently achieves performance on par with full attention across both general-purpose and long-context benchmarks. Moreover, LSA supports native training with context lengths of up to one million tokens and underpins the development of LongCat-2.0 (1.6T-A48B). To facilitate further research, we also introduce and open-source LongCat-Flash-Lite-Sparse (69B-A3B), which integrates LSA into LongCat-Flash-Lite and incorporates an updated long-context training corpus.

The quadratic cost of self-attention makes long-context inference prohibitively expensive, and proxy-based block-sparse attention has become a practical remedy. Existing methods typically rely on a proxy to predict a binary sparse mask and a kernel to consume this mask and perform sparse attention computation. Such an approach is effective under moderate budgets. However, as the budget tightens, the estimated proxy inevitably drops some salient blocks, while the kernel can only apply the sparse mask mechanically, leading to an evident drop in model accuracy. We propose CoSA, a two-stage training-free Sparse Attention under proxy-kernel CO-design, which couples a Kernel-Aware Proxy (KAP) with an Ordered-Skipping Kernel (OSK). In the first stage, the KAP selects blocks under a moderate budget and produces an ordered mask that prescribes the order in which KV pages are visited in the kernel inner loop. In the second stage, the OSK applies this mask and skips more blocks under a tightened budget given online-softmax statistics. Across mainstream LLM backbones and long-context benchmarks, CoSA attains higher accuracy at lower budgets. Impressively, CoSA achieves a 4.93$\times$ attention speedup and reduces end-to-end Time-to-First-Token by 2.53$\times$ under a context length of 128K with negligible performance degradation. Code is available at https://github.com/Tencent/AngelSlim.

Token-level sparse attention, as implemented by DeepSeek Sparse Attention (DSA) in production systems, makes the downstream attention efficient but shifts the bottleneck to the indexer that feeds it. To select the top-k tokens for each query, the indexer must still score every preceding token, incurring a cost of O(L^2) per layer for a sequence of length L. We observe that this per-query scan is largely redundant: nearby queries select highly overlapping top-k tokens, and the indexer scores are long-tailed along the key axis. We exploit these properties in PIVOT, Proxy Indexing Via One full-prefix Traversal, a training-free, drop-in replacement for the DSA indexer that shares one prefix scan across a group of nearby queries. PIVOT aggregates a group into a single proxy query, performs one shared full-prefix scan to obtain a candidate set, and then selects a top-k for each query from that set. Two variants trade speed for fidelity: PIVOT-Reuse shares the proxy top-k across the group for maximum speed, whereas PIVOT-Refine re-scores the candidate set with the indexer of each query and then selects an individual top-k, matching the dense indexer at a small additional cost. A single algorithm covers both inference phases, differing only in how groups are formed: fixed-size groups of consecutive queries in prefill, and the queries decoded together in one multi-token prediction (MTP) step in decode. On DeepSeek-V3.2 and GLM-5.1 across LongBench and RULER, PIVOT matches the accuracy of the dense DSA indexer while accelerating it by up to 4x and reducing end-to-end latency by up to 1.6x at long context.

Diffusion transformers are essential for high-fidelity video generation, but long token sequences make attention a dominant inference bottleneck. Training-free dynamic sparse attention alleviates this bottleneck by computing only selected key-value blocks, yet existing methods struggle to sparsify attention both efficiently and accurately for two reasons: (1) Rigid, unpredictable, and costly routing: selecting a fixed fraction of top-ranked blocks by proxy score imposes fixed budgets, whereas retaining blocks to reach a target cumulative proxy probability mass yields dynamic but potentially imbalanced budgets; both incur non-negligible overhead from computing and materializing proxy scores. (2) Lossy keep-or-drop sparsification: unselected blocks are discarded entirely, degrading accuracy under aggressive sparsity. These limitations motivate cheaper dynamic-budget routing while limiting accuracy degradation. In this paper, we introduce training-free Sol-Attn (Sparsifying online attention), which unifies dynamic routing, sparse computation, and approximation correction in a single online-softmax pass, achieving a better accuracy-efficiency trade-off in sparse attention. The core of Sol-Attn is on-the-fly block thresholding with proxy-score reuse, which selects critical blocks by comparing block proxy scores against a threshold during online softmax. This design enables dynamic yet controllable block budgets without materializing the proxy map, while directly reusing the proxy scores of unselected blocks to approximate their contribution. Experiments across image and video generation tasks show that Sol-Attn advances the quality-efficiency frontier of training-free sparse attention, delivering 2.1 times and 2.3 times end-to-end speedups for video generation and editing, respectively, while preserving visual quality.

Full self-attention in large language models scales as O(N^2), which limits long-context document analysis to 65,536 tokens and requires costly GPU clusters. The Reduced Interaction Sampling (RIS) inference engine addresses this constraint as a model-agnostic architecture. Without modifying weights, RIS reduces self-attention complexity to O(N log N) using sparse stochastic geometry that fits within commodity memory limits. We validate RIS on Qwen2-1.5B-Instruct across two regimes. In controlled evaluations at 32,768 tokens (where native dense attention serves as the upper bound), RIS-Stochastic at 1% density and 70 ensemble seeds achieves 75.00% accuracy, outperforming the native dense baseline (71.88%), while RIS-Stochastic at 5% density and 10 seeds matches it (71.88%). This demonstrates that sparse attention acts as a regularizer: low density (1%) over multiple seeds filters out sequence-level noise, whereas higher density (5%) reintroduces distractor noise. Under the tightest budget, RIS-Structural reaches 68.75% accuracy at 1% density with just 10 seeds, recovering 75% of the contextual gap relative to the zero-context floor (59.38%). At 65,536 tokens, where dense attention triggers out-of-memory faults, RIS yields retrieval gains of up to 14.06 percentage points over the zero-context floor (51.56%), which is confirmed as marginally significant under McNemar's paired test (p = 0.078 < 0.10). All evaluations run on commodity, unaccelerated CPU servers (16-128 GB of RAM), demonstrating that long-context LLM inference is feasible on standard academic hardware without GPU acceleration.

Data-adaptive sparse attention masks substantially outperform fixed patterns (e.g., BigBird and Longformer) and can even exceed dense attention on long sequences. Existing adaptive approaches---including SBM-Transformer, Dynamic Mask Attention, and NSA---typically require additional learnable parameters, custom gradient estimators, or specialized CUDA kernels. We show that classical data compression provides an effective masking signal with \textbf{no additional parameters}. By computing per-block gzip compression ratios, we identify non-redundant content blocks and route long-range attention selectively through them. Intuitively, blocks that gzip cannot compress contain information not predictable from local repetition, making them natural long-range attention targets. Because the compression profile is input-dependent, the resulting sparse mask adapts dynamically to content without learned parameters, auxiliary losses, or custom kernels. On PG-19 byte-level language modeling at 92M parameters with 8K context, our method achieves 1.71 bits-per-byte (BPB), outperforming dense attention (2.89), BigBird (2.34), Longformer (3.21), and a reimplemented SBM-Transformer (3.38)---the only learned-mask baseline---by up to 1.67 BPB while adding no parameters. The advantage grows with sequence length, with the gap over BigBird widening from 0.05 BPB at 4K context to 0.63 BPB at 8K, while convergence is 3.3$\times$ faster.

Time series forecasting (TSF) is vital to many applications, yet existing models often struggle to capture the heterogeneous long-range global patterns and short-range local variations in multivariate time series. While some approaches partially model these dependencies, they often do not jointly exploit temporal and feature-wise information. To address this challenge, we propose HyBDM, a multi-scale hybrid model that decomposes temporal dynamics into global patterns and local variations, which are modeled by two specialized experts. The Global Patterns Expert employs an enhanced BiConv-Mamba module that integrates bidirectional convolutions, an M-SSM layer, a forgetting mechanism, and a GDD-MLP module for cross-channel modeling. The Local Variations Expert uses a Local Window Transformer (LWT) to perform efficient locality-aware attention with reduced computational complexity. In addition, a Multi-Scale Patcher and a Long-Short Router enable multi-resolution representations and adaptive fusion of the two experts. Experiments on six benchmark datasets show that HyBDM outperforms state-of-the-art methods in both forecasting accuracy and computational efficiency, demonstrating its effectiveness in bridging global-local dependencies for multivariate TSF.

Video Diffusion Transformers process long spatio-temporal sequences, making self-attention the main bottleneck in high-resolution video generation. Training-free sparse attention reduces this cost, but adaptive Top-$p$ routing creates uneven per-head workloads under multi-GPU sequence parallelism. The resulting workload heterogeneity turns sparse attention into a rank-level straggler problem. We present \method{}, a training-free sparse-attention system that improves the distributed execution efficiency of adaptive sparse attention under multi-GPU sequence parallelism. \method{} uses Top-$p$ routing, a Top-$k$ safety floor, and video-aware block organization as the sparse-routing frontend, then repairs the materialized mask at runtime. Runtime Load Balancing migrates a small number of heavy heads via P2P communication to shorten the current critical path. Slack-Aware Sparse Augmentation fills residual non-critical-rank slack with additional high-value blocks, while overlap hides scheduling and migration overhead behind existing computation. On step-distilled Wan2.2 I2V, \method{} reduces average load imbalance from 1.34 to 1.08 and delivers a $4.41\times$ attention speedup over FlashAttention, while achieving a $2.02$--$2.11\times$ DiT inference speedup with competitive video quality.

Recent advances in AI-generated content have driven widespread adoption of Diffusion Transformers (DiTs) for high-resolution, long-duration content generation. While parallelization techniques accelerate diffusion inference, they face significant scalability challenges due to excessive communication overhead in multi-node environments. We observe that sequence partitions in Context Parallelism (CP) exhibit distinct heterogeneity: spatially proximate partitions contribute more significantly to attention computation results. By mapping this heterogeneous pattern to hierarchical communication topology, we can access high-contribution partitions with reduced communication cost. This insight motivates our novel selective attention state mechanism that strategically balances partial attention computation and historical result reuse across denoising steps. We present DiTango, an efficient parallel framework for DiT generation. DiTango features an anchor-guided state selection planner that optimizes computation-reuse decisions for each partition, complemented by a runtime that orchestrates efficient state-centric operations. This design achieves superior system efficiency while preserving generation quality. Experimental evaluation on popular diffusion models demonstrates that DiTango achieves up to 1.9x end-to-end and 3.2x attention speedup with near-linear scaling in multi-node settings, while maintaining generation quality comparable to state-of-the-art approaches.

Indexer-TopK, the operation to compute the scores and select the top-k candidates, is widely used by sparse attention kernels in large language models and vector retrieval in recommendation systems and vector databases. However, existing GPU-based Indexer-TopK kernels like DeepSeek Sparse Attention (DSA) remain inefficient due to excessive global memory traffic, costly synchronization, and prohibitive memory overhead. In this work, we exploit the curse of dimensionality in high-dimensional spaces, where distances between high-dimensional vectors tend to concentrate within a narrow range, to design LITETOPK, a novel and efficient fused Indexer-TopK kernel. LITETOPK first samples a small subset of data to estimate query-data score ranges, then uses these estimates to partition candidate results into bins online. This organization allows the LITETOPK kernel to maintain a tight approximate threshold, write back only promising candidates, reduce unnecessary I/O, substantially lower memory overhead, and still preserve exact Top-k correctness. Experimental results show that LITETOPK accelerates the prefill stage of GLM 5.2 by 1.2x in real-world deployment scenarios while incurring lower memory overhead.

The growing adoption of large language model-based agents within operating system workflows has increased the importance of energy-efficient inference on laptop-class systems-on-chip (SoCs). While cloud offloading remains common, it introduces reliability and privacy concerns that are particularly problematic for agentic workloads. Recent laptop SoCs, therefore, incorporate neural processing engines (NPUs) optimized for energy efficiency; however, effectively mapping attention mechanisms onto NPUs remains challenging due to architectural diversity and explicit data-movement programming models. In this work, we present STEEL, the first open-source implementation of FlashAttention targeting XDNA-like NPUs. STEEL introduces a dataflow formulation of prefill attention, enabling efficient exploitation of spatial parallelism and on-chip memory. Furthermore, STEEL addresses the load imbalance induced by the causal mask by leveraging a sparsity-aware pipeline placement onto the NPU array, reducing synchronization overhead and improving utilization. We evaluate STEEL on the AMD Ryzen AI 9 HX 370 SoC and compare its performance against optimized CPU and GPU implementations. Experimental results show that STEEL reduces energy consumption by an average of 9.17x and 1.75x relative to CPU and GPU baselines, respectively. On XDNA 1, STEEL achieves an average 9.6x latency reduction over the prior state of the art, and delivers a 22.8x speedup on average compared to a layer-by-layer attention implementation on XDNA 2.

Block sparse attention is a hardware friendly way to alleviate the key-value (KV) cache read bottleneck in large language models (LLMs). However, it is not prevalent among leading open-weight LLMs, which rely instead on dense attention or fine-grained selection, thereby motivating our analysis. We study DeepSeek's Native Sparse Attention (NSA) as a representative method, whose three-branch design lets us isolate block selection, the most challenging and consequential stage. We formalize selection and reduce it to ranking blocks by a single quantity, the attention mass: the sum of a block's attention scores. We show that if selection retrieves the blocks with the largest attention mass, block sparse attention can match the quality of dense attention. However, computing the exact attention mass requires reading every key, so the problem of block selection ultimately reduces to approximating this mass from a compact summary instead of the full keys. Via a cumulant expansion, we show why existing methods falter: their selection strategies attempt to estimate the attention mass, but are confined to a first-order approximation. Therefore, we propose COBS (Cumulant Order Block Sparse Attention), an attention method that builds on NSA, incorporating a novel selector that stores a compressed second-order statistic per block. On the 32k RULER long-context retrieval benchmark, COBS raises the NSA baseline's mean score from 0.2999 to 0.8195, approaching dense attention at 0.9040 and closing about 86% of the gap, while using only 1.21x the KV cache read traffic of the NSA baseline and 15.15x less read traffic than dense. The same model preserves short-context behavior and attains lower position-wise negative log-likelihood (NLL) than dense attention in our comparison.

Feed-forward 3D reconstruction (F3R) transformers have recently achieved remarkable success. However, scaling them to long image sequences remains challenging, as the quadratic complexity of cross-view global attention quickly becomes the dominant computational bottleneck. While recent efforts attempt to improve efficiency through compressed or sparse attention, they fail to fully exploit the inherent sparsity and dynamic behavior of global attention. In this work, we present a comprehensive analysis of global attention across multiple F3R transformers and reveal that attention patterns are highly heterogeneous, dynamic, and extremely sparse across layers and attention heads. Motivated by these findings, we propose SAF3R, a training-free dynamic sparse attention framework tailored to F3R transformers. SAF3R integrates tailored sparse attention mechanisms with offline head profiling and an efficient online adaptation strategy to match input-dependent attention behaviors. Extensive experiments demonstrate that SAF3R achieves high sparsity ratios while preserving camera pose estimation and 3D reconstruction quality, translating into substantial end-to-end speedup on F3R transformers compared to existing methods. Code is available at https://github.com/jndeng/SAF3R

Video Diffusion Transformers (VDiTs) have demonstrated significant capabilities in high-fidelity video generation. However, their ability to produce long-duration videos is fundamentally constrained by the quadratic complexity of the self-attention mechanism. Recent clustering-based sparse attention methods improve the quality-speed trade-off by grouping semantically similar tokens, but their practical efficiency remains limited by two bottlenecks: substantial clustering overhead and low CTA utilization caused by irregular cluster-induced blocks. We propose HyperVAttention (HVA), a training-free sparse attention framework that addresses both bottlenecks jointly. To reduce clustering overhead, we introduce 3D local-window clustering, which exploits the spatio-temporal locality of video tokens to restrict centroid search to fixed local neighborhoods, and implement it with a custom Triton kernel for efficient execution. We further propose a hybrid clustering strategy that performs full clustering only at anchor steps and updates only subset tokens at intermediate steps, leveraging the temporal stability of cluster assignments across denoising steps. To improve CTA utilization, we present hardware-aware cluster merging that minimizes CTA-aligned execution cost through parallel agglomerative merging, improving block density and approximation fidelity by utilizing idle tile capacity. Together, these components reduce clustering overhead, avoid redundant updates, and better align sparse attention with the fixed tile structure of modern GPU kernels. Experiments on Text-to-Video generation show that HVA establishes a new Pareto frontier for training-free sparse attention in video diffusion, reducing end-to-end latency by up to $2.13\times$ while improving fidelity over existing training-free sparse attention baselines.

Scaling modern large language models (LLMs) to long contexts is limited by the quadratic computation cost, and poor length extrapolation of dense attention. Chunk-wise sparse attention offers a promising alternative, but all existing methods fall short of full attention because of their inaccurate chunk selection. We propose Hierarchical Landmark Sparse (HiLS) Attention, a chunk-wise sparse attention mechanism that learns chunk selection end-to-end under the language-modeling (LM) loss. HiLS factorizes attention hierarchically: each query performs attention independently with each retrieved chunk to extract chunk-specific information, and the resulting outputs are fused according to chunk retrieval scores. By incorporating retrieval scores into the forward attention computation, HiLS optimizes them directly with the LM loss, enabling end-to-end retrieval learning and native sparse training. Experimental results show that HiLS-Attention achieves performance comparable to, and in some cases better than, full attention at in-domain context lengths. Meanwhile, HiLS-Attention extrapolates more than $64\times$ the training context length with 90% retrieval accuracy, far beyond full attention. Moreover, existing full-attention models can be converted to HiLS-Attention with lightweight continued pretraining, preserving in-domain performance while acquiring ultra-long-context extrapolation. Together with its sparse KV access and computation, HiLS-Attention breaks the usual efficiency-performance trade-off, enabling long-context LLMs that are both more efficient and more effective on general long-context tasks than their full-attention counterparts.

Language models (LMs) raise an intriguing alternative to vector-based retrieval: conditioning on an in-context corpus and directly generating a relevant answer. However, prior work has largely focused on proprietary systems or the smaller-scale reranking task, leaving corpus-scale in-context retrieval largely unexplored. In this work, we present the first systematic study of in-context retrieval on two scales practical retrievers demand: million-token corpora and length-generalization far beyond training-time sizes. We first introduce BlockSearch, a 0.6B LM retriever whose architectural and training modifications improve over prior LM baselines and length-generalize up to 10 times beyond its training regime. Nevertheless, retrieval still collapses under more extreme extrapolation. We trace this failure to an attention dilution effect: as the corpus grows, irrelevant documents dominate the softmax denominator, reducing the normalized mass on the gold document even when its pre-softmax score stays high. Motivated by this analysis, we introduce length-aware adjustments to the attention softmax and document-level sparse attention. With these modifications, at the million-token scale, our model matches dense retrieval on widely studied benchmarks (e.g, MS MARCO and NQ), while outperforming the concurrent model MSA despite being 7 times smaller. Furthermore, it significantly outperforms dense retrieval on tasks requiring entirely different notions of similarity, such as LIMIT, achieving a 3 times higher score. Together, our results position in-context retrieval a promising alternative to classical retrieval while emphasizing attention control under extreme context growth as a new challenge.

Long-context Large Language Model inference is severely bottlenecked by the massive Key-Value (KV) cache, yet existing sparse attention methods often suffer from static fixed-budget (Top-k) retrieval or rely on proxy scores that are computationally expensive and biased. To address these limitations, we propose RaBitQCache, a novel sparse attention framework that utilizes randomized rotated binary quantization and high-throughput binary-INT4 arithmetic to efficiently estimate attention weights. Our proxy score serves as an unbiased estimator with a proven error bound, enabling adaptive Top-p retrieval that dynamically adjusts the token budget based on actual attention sparsity. We further implement a hardware-aware system with asynchronous pipelining and lazy updates to mask overhead. Evaluations demonstrate that RaBitQCache significantly accelerates inference and reduces memory I/O while preserving generation quality compared to state-of-the-art baselines. Code is available at https://github.com/Sakuraaa0/RaBitQCache.git.

Dynamic sparse attention (DSA) accelerates long-context LLM decoding by attending to only the top-K KV blocks relevant to each query, but it introduces a serialized selection-to-attention dependency that emerges as a new latency bottleneck. We present PRR, a speculate-reuse-repair runtime that exploits temporal locality in DSA selections to predict likely blocks, speculate the attention over them while selection is in flight, and incrementally repair missed blocks once the true selected set is known. PRR uses a lightweight EMA-based predictor, a profiling-guided speculation budget that keeps speculative work off the critical path, and a FlashAttention-based repair kernel that folds missed blocks into the partial attention state using online-softmax statistics. Across long-context benchmarks and representative DSA methods, PRR reduces per-token decoding latency by up to 40% while preserving downstream task accuracy. Github: https://github.com/Tianyu9748/Incremental_FlashAttention

The quadratic computational cost of traditional attention mechanisms poses a major bottleneck to the scalability and practical deployment of large language models (LLMs), particularly in long-context scenarios. To improve efficiency, existing approaches often enforce rigid structural constraints such as local attention windows. However, these strategies typically lead to substantial performance degradation on tasks requiring precise long-range recall. In this work, we propose MATCH, a scalable and efficient framework that augments sparsified attention mechanisms with dynamically integrated in-context information through an efficient retrieval system. Empirical results show that MATCH significantly improves the performance of sparse-attention models on both synthetic and real-world natural-language tasks. These findings highlight the versatility of MATCH as a general approach for enhancing in-context retrieval capabilities while maintaining the efficiency benefits of sparse attention architectures.

We study sparse self-attention in which each query attends to a dense local window plus a set of Fibonacci-spaced offsets, with a per-layer scalar alpha that compresses or expands the spacing. Across 21 language models trained under one matched recipe (60M parameters, 512 hidden, 16 layers, 426M tokens), we compare four ways of setting alpha across depth: fixed, per-layer learned, a static linear stagger, and a coprime (anti-gridding) reassignment of that stagger, together with a reach-matched power-of-2 control. Three results stand out. First, a static per-layer stagger improves perplexity over both fixed and learned alpha, and the gain is base-agnostic: applying the same stagger to a power-of-2 base lifts it above fixed Fibonacci and to parity with learned Fibonacci attention. Second, learning per layer is inert: it does not beat the static schedule and costs roughly five times the inference latency. Third, and most consequential, all sparse variants extrapolate to four times their training length with little or no degradation, whereas a recipe-matched dense baseline collapses (perplexity rises by 201% at 4x length); we attribute this to fixed-offset attention only ever querying relative positions seen during training. We also report two honest negatives: at training length the best sparse model has about 26% higher perplexity than the dense baseline, and the staggering gain is uniform across context positions rather than concentrated at long range.

Neural scaling laws for transformer language models predict smooth improvements in pretraining loss with increasing parameters, but downstream capabilities such as in-context learning are known to emerge abruptly past a certain model scale. In this paper, we show that emergent capabilities arise stochastically throughout training, with larger models acquiring them earlier on average. We demonstrate that the emergence of capabilities such as pattern completion and indirect object identification corresponds to the abrupt learning of task-relevant attention patterns. To isolate this phenomenon, we train transformer models on synthetic linear map and cellular automata datasets, and we show that the difficulty of learning attention patterns depends on context length and pattern sparsity. Moreover, scaling the number of attention heads improves learning efficiency on our synthetic tasks, while increasing the head dimension yields diminishing returns past a minimum capacity. We additionally investigate architectures with alternative attention mechanisms, showing that MLP-Mixer outperforms a transformer on linear map tasks with complex attention patterns. Our findings provide a mechanistic insight into emergence, showing that downstream capabilities arise abruptly due to the intrinsic difficulty of learning sparse attention patterns in transformer models.

Light field (LF) image denoising is challenging due to the high-dimensional structure of LF data. While noise is independent across sub-aperture images, scene content exhibits strong cross-view correlations. We introduce VSANet, a view-aware sparse attention network for LF denoising. Specifically, we propose a view-aware sparse attention (VSA) block that represents the 4D LF feature map as a unified spatial-angular token space and performs cross-view aggregation via locality-sensitive hashing-based sparse attention. This enables global feature interactions with linear complexity, effectively exploiting LF correlations across views and spatial locations. In addition, we design a feature refinement (FR) block to emphasize informative features in spatial, angular, and epipolar subspaces. The VSA and FR blocks are integrated within a sequential attention refinement module, forming the core of VSANet. Experiments demonstrate VSANet outperforms stateof-the-art LF denoising methods.

While Diffusion Transformers (DiTs) have revolutionized high-fidelity video generation, their reliance on 3D full attention creates a quadratic computational bottleneck. Existing sparse methods face a dilemma: dynamic pruning suffers from prohibitive runtime overhead and memory fragmentation, while static heuristics fail to capture fine-grained dependencies. In this work, we propose ScalingAttention, a training-free framework grounded in a key inductive bias: while individual activations are input-dependent, the high-mass attention regions for each head rapidly converge to a stable, prompt-agnostic Intrinsic Sparse Topology. This topology is weight-encoded, scale-invariant, and efficient to extract. ScalingAttention decouples topology discovery from sparsity control via: (1) WEST (Weight-Encoded Sparse Topology), which extracts a robust block-sparse prior mask offline to eliminate runtime search; (2) FAST (Fidelity-Aware Sensitivity Tuning), which adaptively tunes head-wise sparsity based on diffusion fidelity requirements. To ensure practical acceleration, we co-design a hardware-aligned bit-wise block-sparse kernel. Experiments on Wan2.1 show up to 1.90X end-to-end speedup with superior fidelity, establishing a new Pareto frontier over state-of-the-art baselines.

Long contexts have become standard in pretrained LLMs, yet they remain expensive to run: prefill compute grows quadratically with sequence length, and every decode step re-reads a key-value cache that grows linearly with it. Sparse attention cuts these costs by attending only to a relevant subset of past tokens, but selecting that subset is itself expensive. We present SpotAttention, a lightweight selector that attaches to a frozen pretrained transformer and learns by KL distillation to estimate its attention distribution. The selector picks the top-K keys each query attends to, and because its estimate is a calibrated distribution, a dual top-p rule reads the per-query, per-layer budget directly from it. Across Qwen3 (dense, 4B-32B) and Qwen3.5 (hybrid linear/full attention, 4B-9B), SpotAttention matches dense accuracy at contexts up to 128K tokens, eight times the training length. Decode at L=128K runs 3.9x faster than FlashAttention and 1.8x faster than Twilight, the strongest training-free baseline. Quantizing the selector's K-cache to INT4 or FP4 microscale shrinks it 3.5x at no accuracy cost.

Sparse attention accelerates Diffusion Transformers (DiTs) for video generation by computing only the important tokens while skipping the rest. The token selection strategy is key to balancing sparsity and accuracy. We formulate the token filtering process as a dual-goal optimization problem: maximizing sparsity and minimizing accuracy degradation. Existing algorithms cannot fulfill both objectives simultaneously. For example, Top-p only considers the accuracy constraint, while Top-k maintains a fixed computational budget but loosens the accuracy constraint. This paper demonstrates that maintaining a fixed recall rate is sufficient for ensuring accuracy, whereas a fixed threshold is suboptimal for reducing computational cost. Therefore, we propose a dynamic thresholding scheme to improve sparsity while maintaining the same level of accuracy. Furthermore, our algorithm is deeply integrated with Flash Attention (FA), eliminating the need for any additional masking computation overhead. Experimental results on Wan 2.2 validate that, compared to the BLASST algorithm which is also integrated with FA, our dynamic thresholding strategy enhances sparsity from 61.42% to 82% with a VBench metric drop of less than 5%. This results in an approximate 15% in attention computation and a $1.61\times$ increase in computational efficiency, which is 1.18x higher than that of BLASST.

Remote photoplethysmography (rPPG) transformers achieve low heart-rate error on benchmarks, yet their decisions remain opaque--a growing concern as rPPG moves toward clinical heart rate estimation. Existing rPPG XAI is dominated by qualitative heatmap inspection without quantitative faithfulness metrics or physiology-grounded validation, leaving a gap between visual plausibility and auditable evidence. We address this gap. First, we adapt four attribution methods (raw attention, rollout, flow, Beyond Intuition) to RhythmFormer's bi-level routing attention with top-$k$ selection. Second, we introduce a skin coverage metric quantifying how much attribution mass falls on skin regions. Third, we adapt the SaCo faithfulness coefficient from its original classification setting to rPPG regression by using the MAE between original and perturbed predicted rPPG waveforms as the perturbation impact. Applying these tools, we quantify a multi-hop leakage effect under sparse top-$k$ routing: attention rollout and flow almost completely restores the connections that individual refined-attention layers explicitly set to zero. Beyond Intuition mitigates this via its value-projection-weighted rollout and gradient-supported mask, attaining the highest median refined skin coverage ($0.83$ vs. $0.57$ for vanilla rollout) and faithfulness ($F=0.92$) among the evaluated methods on UBFC-rPPG. Validation across diverse datasets and model variants is needed. A case study on a low-SaCo outlier further shows all four methods recovering consistently once an artefactual region is replaced, suggesting consistent SaCo behavior across attribution families in this illustrative case. Together, these metrics move XAI for rPPG toward auditable numerical evidence about spatial alignment and perturbation faithfulness, i.e. trustworthy rPPG XAI.

Ultra-long-context capability is becoming indispensable for frontier LLMs: agentic workflows, repository-scale code reasoning, and persistent memory all require the model to jointly attend over hundreds of thousands to millions of tokens, yet the quadratic cost of softmax attention makes this untenable at deployment scale. We introduce MiniMax Sparse Attention (MSA), a blockwise sparse attention built upon Grouped Query Attention (GQA). A lightweight Index Branch scores key-value blocks and independently selects a Top-k subset for each GQA group, enabling group-specific sparse retrieval while maintaining efficient block-level execution; the Main Branch then performs exact block-sparse attention over only the selected blocks. Designed around a principle of simplicity and scalability, MSA is deliberately streamlined, making it straightforward to deploy efficiently across a broad range of GPUs. To translate sparsity into practical speedups, we co-design MSA with a GPU execution path that uses exp-free Top-k selection and KV-outer sparse attention to improve tensor-core utilization under block-granular access. On a 109B-parameter model with native multimodal training, MSA performs on par with GQA while reducing per-token attention compute by 28.4x at 1M context. Paired with our co-designed kernel, MSA achieves 14.2x prefill and 7.6x decoding wall-clock speedups on H800. Our inference kernel is available at: https://github.com/MiniMax-AI/MSA. A production-grade natively multimodal model powered by MSA has been publicly released at: https://huggingface.co/MiniMaxAI/MiniMax-M3.

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.

Conventional LLMs keep the full KV cache loaded during decoding, causing a severe GPU memory bottleneck for ultra-long context serving. In this report, we propose Lookahead Sparse Attention (LSA), a novel inference paradigm powered by a Neural Memory Indexer built upon the DeepSeek-V4 architecture. Rather than passively attending to all historical tokens, LSA proactively predicts future context demands and preserves only the query-critical KV chunks in the GPU memory. Crucially, we instantiate this architecture via a backbone-free decoupled training strategy. By formulating the indexer as a standard dual-encoder architecture, we train it independently using standard retrieval training frameworks without ever loading the massive backbone model into GPU memory. We demonstrate that this "less is more" paradigm significantly maximizes serving efficiency while acting as an effective attention denoiser in tasks that rely on long-term global memory. Across primary long-context evaluation suites (e.g., LongBench-v2, LongMemEval, and RULER), FM-DS-V4 compresses the average physical KV cache footprint down to merely 13.5% of the full-context baseline, while consistently preserving or slightly elevating downstream accuracy (+0.6% absolute margin on average). Crucially, at extreme 500K scales, FlashMemory suppresses the physical KV cache overhead by over 90% without destabilizing the backbone's core reasoning capacities.

Long-context prefill remains expensive because full/GQA layers still score the historical sequence, even in hybrid models with local, sparse, linear, or recurrent components. We study how much dense attention is needed to preserve task-level behavior under explicit support granularity and top-k budgets. We introduce an attention-mass top-k oracle for existing GQA checkpoints: for each layer and query position, it computes dense attention, selects head-averaged token support, and recomputes attention only on that support. The oracle is a diagnostic reference, not a deployable accelerator, and separates sparse-budget feasibility from indexer error and runtime realization effects. On Qwen-family retrieval-heavy evaluations, the longest per-query oracle rows stay within 1 point of dense, and a Qwen3.5-9B RULER-style sweep from 4K to 100K stays within 0.48 points. Guided by the oracle, we derive a head-collapsed auxiliary indexer trained by KL distillation from dense attention-mass distributions while keeping the backbone frozen. With separately distilled Qwen3.5-0.8B and Qwen3.5-9B indexers, the reported 16K/32K validation macro gaps are +2.04 and +1.13 points, treated as quality preservation rather than improvement; fused selection-block-shared support can introduce a larger realization gap. Preliminary single-card TTFT measurements show distilled-indexer sparse serving speedups of 1.71x for Qwen3.5-0.8B on NPU and 1.93x for Qwen3.5-9B on GPU against its dense FlashAttention-2 baseline. Additional random-init stress rows reach 3.44x, indicating sparse-runtime headroom but not validated output quality. This first release separates oracle feasibility, distilled-indexer quality, and runtime headroom, leaving a fully matched quality-latency frontier to future work.

Long-context inference in modern LLMs is increasingly constrained by decoding efficiency, especially in reasoning-heavy settings where models generate long intermediate chains of thought. Existing sparse attention methods often face a practical efficiency-quality trade-off. Structured block sparse methods typically provide stronger acceleration but incur noticeable quality loss, while token sparse methods are usually more accurate yet deliver limited end-to-end speedup because top-k routing over the full cache remains expensive. In this work, we propose cross-layer sparse attention (CLSA), which is built on top of KV-sharing architectures such as YOCO. The core idea is to share not only the KV cache across cross-decoder layers, but also the routing index. A single indexer computes token-level top-k selection once and reuses the resulting index across layers, thereby preserving the fine-grained selectivity of token sparse attention while amortizing the routing overhead. The resulting architecture improves all major inference bottlenecks jointly, including pre-filling, KV-cache storage, and long-context decoding. Experiments across short-context and long-context benchmarks show that CLSA is both accurate and efficient, achieving up to 7.6x decoding speedup and 17.1x overall throughput improvement at 128K context. These results suggest a more complete architectural solution for long-context LLMs that jointly advances model quality and inference efficiency.

Sparse attention is becoming increasingly important for serving large language models (LLMs) as generation lengths continue to grow. However, deploying and evaluating new sparse attention algorithms at scale remains highly engineering-intensive, slowing both human researchers and AI agents in exploring the sparse attention design. To address this challenge, we present Vortex, a system that combines a Python-embedded frontend language atop a page-centric tensor abstraction for expressing a broad range of sparse attention algorithms, with an efficient backend tightly integrated into modern LLM serving stacks. Vortex enables rapid prototyping, deployment, and evaluation of sparse attention algorithms, effectively translating their theoretical efficiency gains into real-world throughput improvements. As a result, Vortex substantially accelerates the design and iteration of sparse attention algorithms. First, AI agents use Vortex to automatically generate and refine diverse algorithms, the best reaching up to $3.46\times$ higher throughput than full attention while preserving accuracy. Second, Vortex extends sparse attention to emerging architectures and very large models that are otherwise hard to experiment with, reaching up to $4.7\times$ higher throughput on the MLA-based GLM-4.7-Flash and $1.37\times$ on the 229B-parameter MiniMax-M2.7 on NVIDIA B200 GPUs.

Sparse attention reduces compute and memory bandwidth for long-context LLM inference. However, two key challenges remain: (1) KV cache capacity still grows with sequence length, and offloading to CPU memory introduces a PCIe transfer bottleneck; (2) the sparse selection step itself retains $O(T^2)$ complexity and can dominate attention cost at long contexts. We propose SparDA, a decoupled sparse attention architecture that introduces a fourth per-layer projection, the Forecast, alongside Query, Key, and Value. The Forecast predicts the KV blocks needed by the next layer, enabling lookahead selection that overlaps CPU-to-GPU prefetch with current-layer execution. Because Forecast is decoupled from the attention query, our GQA implementation uses one Forecast head per GQA group, reducing selection overhead versus the original multi-head selector. SparDA adds $<$0.5% parameters and trains only the Forecast projections by matching the original selector's attention distribution. On two sparse-pretrained 8B models, SparDA matches or slightly improves accuracy and delivers up to 1.25$\times$ prefill speedup and 1.7$\times$ decode speedup over the sparse-attention offload baseline. By enabling larger feasible batch sizes on a single GPU, SparDA further reaches up to 5.3$\times$ higher decode throughput than the non-offload sparse baseline. Our source code is available at https://github.com/NVlabs/SparDA.

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.

Frontier LLMs increasingly decide what a query attends to with a sparse-attention indexer that picks a few KV-cache blocks per query: attention's unit is now a small, reusable chunk. Agentic workloads hammer it: many sub-agents query one large codebase, reusing the same blocks. When that corpus outgrows one GPU it is partitioned across instances, so a query and the blocks it selects often sit on different GPUs: answering it means attention across instances. The reflex of prior cross-instance KV systems is to move the cache: pull the selected blocks to the requester. Multi-head Latent Attention inverts the arithmetic, compressing each token's key and value into one narrow vector, so a routed query row is only ~1 KB, smaller than the chunk it attends; routing the query is then often cheaper than moving the cache. Which primitive wins, over which fabric and request shape, is uncharted, least of all on device-initiated RDMA that makes per-request cross-node transfers cheap. We characterize cross-instance MLA attention on a real multi-node H100 cluster, distilling two reusable artifacts: a topology-aware cost model (probe / transfer / compute / return / merge) and a closed-form route/fetch/local predicate, whose constants we measure on real IBGDA, where the model tracks batched round-trips to within ~7%. At decode it routes the query, trading the cost of moving the cache (a ~3 ms re-adaptation splice for a contiguous chunk, or a scattered gather under selection) for a tens-of-microsecond round trip, and picks the fabric by probe latency, not peak bandwidth. We instantiate the cost model and predicate for MLA, but neither is MLA-specific: they apply wherever compression or sparse selection shrinks attention to small chunks (DeepSeek-V3.2, V4, and GLM-5.1 today). Extending them to a new architecture requires measuring just two coefficients: the routed payload and fetch's move-the-cache cost.

Interactive video world models generate video chunk by chunk in response to user-controlled camera movements, enabling applications such as real-time game simulation, virtual scene navigation, and embodied AI training. However, scaling to long interactive trajectories is prohibitively expensive due to growing context memory, quadratic attention complexity, and repeated denoising steps. We present Light Interaction, a training-free inference acceleration framework for interactive video world models. Our key insight is that interaction naturally enables trajectory-dependent adaptive computation: retrieved spatial memory can be discarded during novel exploration, temporal context can be adjusted according to local latent dynamics, and early-step model outputs can be reused when the camera revisits familiar regions. Based on this insight, Light Interaction combines adaptive context management, denoising cache acceleration, and hardware-software co-designed 3D block sparse attention with fused Triton kernels. Evaluated on HY-WorldPlay and Matrix-Game-3.0, Light Interaction achieves up to 2.59x speedup without model retraining while maintaining competitive visual quality.

Dense self-attention is the compute and quality bottleneck of long-video diffusion inference: cost grows quadratically with the sequence length, and beyond the training horizon the model converges to near-static output, that is, "frozen" repetitive video. State of the art approaches are either too costly, e.g., they require retraining, or fail to satisfy both performance and quality objectives in a scalable manner. To this end, we introduce Long Video Sparse Attention (LVSA), a training-free model-agnostic block-sparse attention for video diffusion transformers that combines a structured window pattern with rotating global anchors, thus removing the fixed-grid bias which causes long-range temporal artifacts. LVSA, combined with a FlashInfer kernel, reduces compute up to 3.17x on Wan 2.1 1.3B at a 6x horizon, 2.98x on Wan 2.1 14B at a 6x horizon, and 3.33x on HunyuanVideo 1.5 at a 1.5x horizon, compared to dense attention. Beyond reducing compute, LVSA enables HunyuanVideo 1.5 generation at a 2x horizon, which is otherwise out-of-memory on a single GPU. Moreover, LVSA provides speedups up to 2.41x compared to RIFLEx and 3.27x compared to UltraViCo on Wan 2.1 1.3B. To demonstrate applicability across diverse platforms, we apply LVSA on NPUs and achieve speedups up to 2.71x on Wan 2.2 A14B and 3.24x on Wan 2.1 1.3B compared to dense attention. To evaluate quality in a fair way, we introduce VQeval, a tool properly scoring loopy video failures, which instead are rewarded in state of the art evaluators like VBench-Long. LVSA is quality-neutral for generation at training horizon length and quality-positive at extended lengths.

Scaling Diffusion Transformers to generate high-resolution, long videos is constrained by the quadratic cost of self-attention, and existing sparse attention methods degrade under high sparsity. We show empirically that generation quality is determined not by the sparsity ratio itself, but by how well the sparse mask aligns with the tile-wise geometry of full attention. Based on this insight, we propose Veda, a distilled sparse attention framework that formulates tile selection as an explicit reconstruction problem from full attention. Veda integrates statistics-aware tile scoring with head-aware tiling to reduce estimation error and structural mismatch, enabling aggressive sparsity. A hardware-efficient tile-skipping kernel converts theoretical sparsity into practical wall-clock speedups. Experiments on large video diffusion models, including Waver and Wan2.1, demonstrate substantial acceleration with no noticeable degradation in generation quality. To generate 720P 10-second videos on Waver-T2V-12B, Veda achieves a 5.1$\times$ end-to-end speedup and a 10.5$\times$ self-attention speedup, reducing attention overhead from 92% to 50%. Notably, the gains increase with sequence length, indicating that Veda scales favorably with spatiotemporal resolution across models.

Long-context inference in large language models (LLMs) is bottlenecked by the linear growth of the self-attention key-value (KV) cache. Top-k sparse attention alleviates this by loading only a small fraction of the KV cache, but accurately and cheaply estimating cache importance, for both training-free use and sparsity-aware training, remains challenging. This paper proposes UNIQUE, a universal top-k sparse attention framework that addresses both requirements and stays consistently effective across LLM modalities. UNIQUE operates at the granularity of KV pages and estimates per-page importance with a simple yet accurate score combining the mean of the page's keys as a representative vector with their standard deviation as an offset term. To further close the train-inference gap, this paper introduces a soft-mask sparsity-aware training scheme that uses the top-k score boundary as a per-query threshold and a sigmoid soft mask around it, requiring neither auxiliary losses nor architectural changes. Experiments on text and speech LLMs show that UNIQUE preserves task performance on long-context benchmarks such as LongBench Pro and on long-form speech recognition, while delivering up to 11.4x attention-kernel speedup over FlashInfer dense attention and at least 5.3x end-to-end decoding speedup over a vLLM-based dense model.

We introduce JetViT, a novel family of hybrid-architecture Vision Transformer (ViT) models that match the accuracy of state-of-the-art full-attention vision foundation models while achieving substantially higher inference efficiency on high-resolution images. At the core of our approach is Post-Training Attention Search, a post-training acceleration framework that converts pre-trained full-attention ViTs into efficient hybrid-attention variants by identifying and replacing redundant full-attention blocks with linear or window-attention blocks. By inheriting the MLP and attention weights from the base model, Post-Training Attention Search efficiently explores the architectural design space through three key steps: (1) optimizing the linear-attention block design; (2) finding the best combination of linear-attention and window-attention blocks; and (3) identifying and preserving critical full-attention blocks. We evaluate JetViT on two representative high-resolution vision foundation models, DINOv3 and DepthAnythingV2. On the NVIDIA H100 GPU, JetViT achieves up to 1.79x higher throughput and up to 44.81% lower latency without sacrificing accuracy. We will release our code and accelerated ViT models soon.

The quadratic complexity of self-attention in Transformer models remains a significant bottleneck for processing long sequences and deploying large language models efficiently. For this approach, there has been significant research into Sparse Attention, and Deepseek Sparse Attention has combined various methods of creating segments of tokens to reduce the time complexity. This paper introduces a novel approach, Grammatically-Guided Sparse Attention, which constrains attention computations based on the grammatical roles of tokens. By leveraging Parts-of-Speech (POS) tags, attention masks are dynamically generated that enforce linguistically coherent connections between tokens, reducing the computational graph without sacrificing essential linguistic dependencies. Two masking strategies are proposed and evaluated: a hard mask that strictly allows only predefined grammatical interactions, and a soft mask that biases attention towards these interactions. The experiments, conducted on the SST-2 sentiment classification task using a DistilBERT-like architecture, demonstrate that Grammatically-Guided Sparse Attention maintains comparable accuracy to full attention while significantly reducing the theoretical computational overhead. Preliminary results show accuracy values of 0.8200 for hard masking and 0.8165 for soft masking, closely matching the 0.8200 of full attention, providing a path towards more efficient, interpretable, and linguistically-informed Transformer architectures.