paperAug 11, 2026▲ 28
Zunhai Su, Bohan Sun, Xialie Zhuang et al.
We present the first systematic study of Massive activations (MAs) in layer-interleaved HLA LLMs and uncover two architecture-aligned morphologies: MAs consistently spike immediately before full attention layers, forming pre-attention spikes (PAS), and can persist through intervening linear attention layers, giving rise to inter-spike plateaus (ISP). As full attention becomes denser, successive PAS become increasingly connected through ISP, ultimately recovering the stable MA morphology of full attention LLMs. We establish the recurrence of this organization across five linear attention architectures, six hybridization configurations, five data domains, and representative open-source hybrid models spanning 1.2B to 397B total parameters. Controlled pretraining of GDN-based hybrids at scales up to 1.3B shows that both morphologies emerge early and respond asymmetrically to output gating: full attention output gating strongly attenuates their absolute magnitudes without eliminating their layerwise organization, whereas removing GDN gates yields comparatively modest amplification. Mechanistically, our systematic-outlier analysis supports a shared lifecycle account governed by the timing of MA cancellation. PAS follows a localized write-sink-cancel process, while the extended persistence of ISP is consistent with delayed cancellation. At the full attention limit, this account recovers the stable MA morphology characteristic of full attention LLMs. Our code is available at https://github.com/StartluxLabs/Massive-Activations-HLA.
paperJul 22, 2026▲ 34
Junsong Chen, Jincheng Yu, Yitong Li et al.
We introduce SANA-Video 2.0, a hybrid video diffusion transformer instantiated at 5B and 14B scales under a unified architecture. Designed to generate high-quality video up to 720p on a single GPU, SANA-Video 2.0 matches full-softmax video DiTs in quality while retaining the favorable long-sequence scaling of linear attention. To avoid quadratic attention throughout, Hybrid Linear-Softmax Attention combines gated linear attention for O(N)-dominated mixing with periodic gated-softmax anchors at a 3:1 ratio, restoring the full-rank token interactions that pure linear attention lacks. To propagate these refreshed representations across depth, Block Attention Residuals (AttnRes) route completed block summaries into later linear layers, enabling anchor-feature reuse and boosting deep-layer effective rank by ~12%. Through from-scratch training, SANA-Video 2.0 learns the complete hybrid directly rather than linearizing pretrained models, with reduced-resolution proxy studies establishing 25% softmax as the optimal quality-efficiency trade-off. With 40-step sampling, SANA-Video 2.0 achieves a VBench score of 84.30 in 13.2s at 480p on a single H100, remaining competitive with far larger softmax video DiTs at a fraction of the latency. Its compiled DiT forward pass is 3.2x faster than a matched full-softmax baseline at 720p/60s, a gap that expands with video duration. Furthermore, full-stack Sol-Engine optimization (kernel fusion, caching, and sparse attention) accelerates this hardware-friendly backbone by a further 3.58x, bringing the 5B pipeline to 13.06s at 720p/5s and making it 120x faster than Wan 2.2-A14B on one H100. Overall, our hybrid design recovers softmax-level expressiveness at substantially reduced cost, unlocking scalable long, high resolution video generation.
quant-ph · cs.LGAug 11, 2026new
Eric A. F. Reinhardt, Adam J. Hauser
The attention mechanism forms the foundation of many modern AI models such as the Transformer. In one subclass of problems where attention is used, inputs and outputs are bound to the probability simplex so that all outputs sum to one. In this setting, softmax attention admits an exact, component-by-component quantum realization. Attention scores are Hadamard-test statistics on block-encoded projections of amplitude-encoded inputs. The exponential softmax is the interior of a cosine-squared family generated by Born-rule measurement under an exact bijection, whose boundary expresses sparse attention with exact zeros at finite parameter values. The softmax temperature is a repetition count where post-selected measurement rounds realize discretized inverse temperature exactly. Value aggregation is a deterministic column-loading channel that dilates the column-stochastic value matrix. The gated residual is the preparation angle of a single ancilla, with the additive identity at a mixing angle of π/2. Every learnable parameter is a rotation-gate angle. The composed layer is exact in the infinite-shot limit with one measure-and-reload step per attention score; a fully-coherent variant is ε-approximate via quantum singular value transformation in the infinite depth limit. The algebraic core is machine-checked in Lean 4.
paperNeurIPS 20252025269 citations
Zihan Qiu, Zekun Wang, Bo Zheng et al.
Gating mechanisms have been widely utilized, from early models like LSTMs and Highway Networks to recent state space models, linear attention, and also softmax attention. Yet, existing literature rarely examines the specific effects of gating. In this work, we conduct comprehensive experiments to systematically investigate gating-augmented softmax attention variants. Specifically, we perform a comprehensive comparison over 30 variants of 15B Mixture-of-Experts (MoE) models and 1.7B dense models trained on a 3.5 trillion token dataset. Our central finding is that a simple modification-applying a head-specific sigmoid gate after the Scaled Dot-Product Attention (SDPA)-consistently improves performance. This modification also enhances training stability, tolerates larger learning rates, and improves scaling properties. By comparing various gating positions and computational variants, we attribute this effectiveness to two key factors: (1) introducing non-linearity upon the low-rank mapping in the softmax attention, and (2) applying query-dependent sparse gating scores to modulate the SDPA output. Notably, we find this sparse gating mechanism mitigates 'attention sink' and enhances long-context extrapolation performance, and we also release related $\href{https://github.com/qiuzh20/gated_attention}{codes}$ and $\href{https://huggingface.co/QwQZh/gated_attention}{models}$ to facilitate future research.
cs.LGAug 19, 2026new
Sotirios P. Chatzis, Loukas Papadoulas
Deep models for irregularly-sampled time series answer queries at arbitrary continuous timestamps, yet report nothing about how far each answer should be trusted. We show the attention layer itself can close that gap: with the right stochastic formulation, the pass that makes each prediction also reports, in closed form and at no extra cost, how far it should be trusted. We introduce Lévy Attention, a cross-attention operator whose output is a stochastic integral against an inhomogeneous Poisson random measure: query-key compatibilities assemble an intensity over a continuous (time x channel) index space, the measure scatters atoms under it, and the output averages an interpolated value field at those atoms. In expectation it reduces to a mollified cosine-kernel attention, so it replaces a softmax layer and trains with exact gradients.
What softmax discards, the Poisson construction preserves in closed form: the evidence $Λ_q$ (total compatibility mass) and the disagreement $\mathrm{tr}\,Σ_V(q)$ (value spread). An exact variance identity makes their combination $\hatσ(q)=\sqrt{\mathrm{tr}\,Σ_V(q)\,\varphi(Λ_q)}$ the root-mean-square deviation of the sampled operator, emitted by the deterministic pass with no trained head.
Empirically, disagreement carries the signal, while the evidence factor swings from uninformative on dense data to strongly informative on sparse. On t-PatchGNN the operator swap costs at most 5.6% accuracy against a matched control and nothing on the sparsest dataset. The free disagreement signal improves on 20-pass MC dropout across matched five-seed suites, and $\hatσ$ scales a calibrated Gaussian whose zero-sample CRPS beats a fifty-draw sampler; a split-conformal wrapper reaches nominal coverage at every level, and one pass ranks 3,383 unseen patients by trust in 1.4 seconds.
paperAug 9, 2026▲ 7
Burc Gokden
The Large Language Model from Power Law Decoder Representations (PLDR-LLM) and its attention, Power Law Graph Attention (PLGA), replace the fixed bilinear form of scaled dot-product attention (SDPA) with a learned, input-generated bilinear operator G_{LM}, built from a positive tensor A_{LM} by elementwise power laws. The architecture is fully specified, verified against pinned reference releases; claims are labeled theorem, conditional theorem, measurement, or conjecture. Unconditionally: PLGA contains SDPA exactly at G_{LM}=I; A_{LM} and A_P are strictly entrywise positive, with Perron-Frobenius structure on A_{LM}; the DAG regularizer has the NOTEARS walk-counting form and positivity obstructs exact acyclicity; and, under nonresonance (satisfied by standard rotary frequencies), a commutant criterion identifies which operators preserve relative-position dependence. An inference-collapse theorem: exact input invariance of deductive outputs collapses inference to generalized SDPA with a constant operator. Measured invariance: relative fluctuations of 10^{-6} and below; perturbation bounds quantify but do not certify cached inference; the assembled proxy misses the decoding margin. A conditional three-stage mechanism (rotary twirl, concentration, row-map contraction) is measured on a released checkpoint. Blockwise training and scoring under the global Gram are stated with explicit target exposure; on tested samples, block and sequential scoring select identical answers and agree on the published TruthfulQA probability-mass metric within 5times 10^{-5} per item. Self-organized criticality enters as a phenomenological framework with an intrinsic order parameter; open claims become falsifiable conjectures. Selected proof cores are machine-checked in Lean 4.
paperJul 2, 2026▲ 44
Xiang Hu, Xinyu Wei, Hao Gu et al.
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 64times 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.
paperICLR 20242024202 citations
Xiangming Gu, Tianyu Pang, Chao Du et al.
Language Models (LMs) assign significant attention to the first token, even if it is not semantically important, which is known as attention sink. This phenomenon has been widely adopted in applications such as streaming/long context generation, KV cache optimization, inference acceleration, model quantization, and others. Despite its widespread use, a deep understanding of attention sink in LMs is still lacking. In this work, we first demonstrate that attention sinks exist universally in LMs with various inputs, even in small models. Furthermore, attention sink is observed to emerge during the LM pre-training, motivating us to investigate how optimization, data distribution, loss function, and model architecture in LM pre-training influence its emergence. We highlight that attention sink emerges after effective optimization on sufficient training data. The sink position is highly correlated with the loss function and data distribution. Most importantly, we find that attention sink acts more like key biases, storing extra attention scores, which could be non-informative and not contribute to the value computation. We also observe that this phenomenon (at least partially) stems from tokens' inner dependence on attention scores as a result of softmax normalization. After relaxing such dependence by replacing softmax attention with other attention operations, such as sigmoid attention without normalization, attention sinks do not emerge in LMs up to 1B parameters. The code is available at https://github.com/sail-sg/Attention-Sink.
cs.LG · cs.AIJul 7, 2026new
Shervin Khalafi, Igor Krawczuk, Sergio Rozada et al.
Denoising graphs is a fundamental problem in graph learning and the core operation of graph diffusion models. Attention-based architectures like graph transformers have recently shown promise in denoising graphs. However, our principled understanding of attention-based graph denoising remains limited, making it unclear whether standard attention is the right mechanism for this task. Here we show that, under a denoising objective, linear attention is suboptimal and can only learn an average spectral denoising filter over the training distribution. This creates a fundamental limitation as graphs often vary spectrally across the distribution. To overcome this limitation, we introduce Spectral Attention, which directly utilizes the input graph spectrum and provably outperforms linear attention by a margin governed by the spectral diversity of the distribution. We then derive Graph Convolutional Attention (GCA), a practical and permutation-equivariant realization of this idea that implements spectral denoising through graph-filtered queries and keys. For stochastic block models, GCA provably matches the idealized Spectral Attention mechanism. We further show that the softmax operation, that follows the attention, provides additional denoising by approximately projecting noisy eigenvectors onto the clean eigenspace. Empirically, replacing linear attention with GCA consistently improves graph denoising and diffusion on synthetic and real datasets, with gains strongly correlated with spectral diversity. In DiGress, GCA matches standard graph-transformer performance without computing expensive structural features, and when combined with the recently proposed PEARL positional encodings, avoids explicit eigendecomposition computations resulting in faster inference without degrading quality. The code can be found here: github.com/shervinkhalafi/graph_conv_att
paperJun 25, 2026▲ 1
Yuzhen Mao, Michael Y. Li, Emily B. Fox
Sparse attention can reduce the cost of long-context inference, but most variants introduce new architectural components. We introduce Simplified Sparse Attention (SSA), a simpler approach to sparse attention that requires no architectural changes. Concretely, we first perform continued pretraining on sequences interleaved with gist tokens. We optimize the standard next-token loss as usual, but the gist tokens use an attention mask to restrict what parts of the context the language model can attend to; this teaches the model to pack each chunk's important information into the gist tokens. At inference time, SSA scores chunks via attention between the current query and the small set of gist tokens, selectively unfolding the top-k chunks by reintroducing their corresponding raw tokens. Since the query is scored only against the gist tokens, we avoid the memory-bandwidth cost associated with naive scoring against the full KV cache, without requiring the auxiliary KV cache approach used by sparse attention methods. On LongBench, SSA consistently outperforms compression and inference-time sparse-attention baselines under the same compression ratio. More strikingly, in retrieval-augmented generation, SSA can even outperform full attention after continued pretraining by over 5.7 points. We attribute this to the ability of SSA's selective unfolding, which concentrates attention on the query-relevant chunks and effectively filters out noise. SSA further extends to a hierarchical gist-of-gist variant (H-SSA) that achieves log-linear decoding complexity while maintaining or improving accuracy at high compression ratios up to 32x. The code is available at https://github.com/yuzhenmao/simplified-sparse-attention/.
paperJun 5, 2026▲ 1
Prateek Kumar Sikdar
Standard transformers apply self-attention uniformly at every layer and token, regardless of whether the input requires dynamic cross-token interaction. We propose CHIAR-Former (Chiaroscuro Attention), a 4-layer hybrid transformer that routes each token to one of three operators - DCT spectral mixing, RBF kernel mixing, or full self-attention - based on per-token spectral entropy, a theoretically justified complexity signal. Through systematic ablation on WikiText-103, we discover routing collapse: the router consistently rejects RBF in favour of DCT and attention, revealing that spectral mixing and dynamic attention are complementary and sufficient. A purpose-designed DCT+Attention-only variant achieves Val PPL 36.54 on WikiText-103 - a 45% improvement over a full-attention baseline (PPL 66.62) at 62.5% fewer attention FLOPs. We extend evaluation to WikiText-2, IMDB sentiment classification, and synthetic ListOps operations, establishing a clear operating regime: CHIAR-Former excels on large-scale naturalistic text where token diversity supports spectral specialisation, while full attention retains an edge on small datasets and synthetic pattern-matching tasks. These findings - both the wins and the losses - together define when and why spectral routing earns its keep.
paperJun 28, 2026▲ 41
Disen Lan, Jianbin Zheng, Yuxi Ren et al.
Hybrid attention models improve long-context efficiency by retaining only a subset of full-attention layers and replacing the remaining layers with linear attention. However, the effectiveness of Transformer-to-hybrid conversion critically depends on which layers preserve full attention. Existing hybrid layer selection methods typically rely on heuristic strategies such as fixed placement patterns or layerwise scoring, implicitly treating layer importance as isolated and overlooking the interdependent layer effect under a global hybrid configuration. In this work, we formulate hybrid layer selection as a budget-constrained subset optimization problem. We further propose FlashMorph (Fast LAyer Selection for Hybrid MORPHing), an effective, efficient and scalable layer selection method for Transformer-to-hybrid conversion. FlashMorph first constructs a morphable model by equipping each full-attention layer with a converted linear-attention branch. It then freezes all model weights and jointly optimizes layerwise gates on synthetic long-context retrieval data, with a linearization regularization that encourages the model to rely on linear attention for efficiency. The learned gates are discretized under a preset full-attention budget to instantiate the hybrid architecture, followed by standard logits distillation and long-context finetuning. Extensive experiments show that FlashMorph discovers more effective hybrid configurations, preserves strong long-context recall and general benchmark performance while substantially reducing layer selection cost compared with existing layer selection methods, demonstrating its effectiveness, efficiency, and scalability.
paperJun 12, 2026▲ 3
Ziqing Qiao, Yinuo Xu, Chaojun Xiao et al.
Modern language models increasingly adopt hybrid architectures that combine full attention with efficient attention modules, such as sliding-window attention (SWA) and recurrent sequence mixers. However, how these efficient modules shape model capabilities remains poorly understood. To address this gap, we conduct a systematic analysis across hybrid architectures from three perspectives: scaling behavior, mechanism analysis, and architecture design. First, from a scaling perspective, we find that efficient-attention design primarily affects how fast long-context capability emerges, while different hybrids eventually converge to comparable long-context performance under sufficient training. Second, mechanistically, we show that long-range retrieval is mainly carried by full attention, whereas efficient attention shapes its optimization trajectory. This explains a counter-intuitive phenomenon we call Large-Window Laziness: larger SWA windows can delay the formation of retrieval heads in full-attention layers. Third, guided by this mechanism, we show that applying NoPE to only the full-attention layers of a small-window SWA hybrid substantially improves long-context performance with negligible impact on short-context performance.
cs.AIJun 4, 2026new
Zhuoming Chen, Xinrui Zhong, Qilong Feng et al.
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.
paperJun 10, 2026▲ 126
Xunhao Lai, Weiqi Xu, Yufeng Yang et al.
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.
paperNeurIPS 20242024272 citations
Dongchen Han, Ziyi Wang, Zhuofan Xia et al.
Mamba is an effective state space model with linear computation complexity. It has recently shown impressive efficiency in dealing with high-resolution inputs across various vision tasks. In this paper, we reveal that the powerful Mamba model shares surprising similarities with linear attention Transformer, which typically underperform conventional Transformer in practice. By exploring the similarities and disparities between the effective Mamba and subpar linear attention Transformer, we provide comprehensive analyses to demystify the key factors behind Mamba's success. Specifically, we reformulate the selective state space model and linear attention within a unified formulation, rephrasing Mamba as a variant of linear attention Transformer with six major distinctions: input gate, forget gate, shortcut, no attention normalization, single-head, and modified block design. For each design, we meticulously analyze its pros and cons, and empirically evaluate its impact on model performance in vision tasks. Interestingly, the results highlight the forget gate and block design as the core contributors to Mamba's success, while the other four designs are less crucial. Based on these findings, we propose a Mamba-Inspired Linear Attention (MILA) model by incorporating the merits of these two key designs into linear attention. The resulting model outperforms various vision Mamba models in both image classification and high-resolution dense prediction tasks, while enjoying parallelizable computation and fast inference speed. Code is available at https://github.com/LeapLabTHU/MLLA.
paperJul 26, 2026▲ 28
Haopeng Li, Yitong Li, Junsong Chen et al.
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.
paperJul 16, 2026▲ 8
Hao Liu, Chenghuan Huang, Ye Huang et al.
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 , a training-free sparse-attention system that improves the distributed execution efficiency of adaptive sparse attention under multi-GPU sequence parallelism. 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, reduces average load imbalance from 1.34 to 1.08 and delivers a 4.41times attention speedup over FlashAttention, while achieving a 2.02--2.11times DiT inference speedup with competitive video quality.
cs.CVJun 29, 2026new
Seongro Yoon, Donghyeon Cho, Jinsun Park et al.
Understanding facial expressions in videos requires modeling subtle and localized facial dynamics under unconstrained conditions. Although recent Vision Transformer~(ViT)-based video models have shown strong performance through large-scale self-supervised pretraining, their attention mechanisms often emphasize dominant global motions and coarse temporal dynamics, limiting sensitivity to fine-grained facial variations. To address this limitation, we propose MiRA (Marginal-induced Attention Redistribution), a plug-in frame-marginal attention redistribution framework for ViT backbones that enhances spatio-temporal selectivity toward subtle facial dynamics without introducing additional trainable parameters. MiRA derives frame-level confidence and intra-frame concentration statistics from self-attention maps to estimate frame-wise marginal importance and redistribute attention toward spatiotemporally localized facial cues. We first introduce a principled \textit{exact mode} based on post-softmax attention redistribution. To further improve efficiency, we propose \textit{flashLite mode}, a lightweight pre-softmax approximation that integrates frame-marginal redistribution into FlashAttention kernels while preserving the effectiveness of the exact formulation. Experimental results on challenging Facial Expression Recognition~(FER) benchmarks demonstrate consistent improvements over strong ViT baselines.
paperJun 17, 2026▲ 13
Zhentao Tan, Wei Chen, Jingyi Shen et al.
The quadratic complexity of attention poses a critical bottleneck for long-context processing, spurring interest in hybrid attention designs. Most open-source hybrid models adopt a layer-wise strategy. Yet, prior work has noted the inherent difficulty of integrating Linear Attention (LA) with Full Attention (FA), suggesting that the design space of attention hybridization remains underexplored. To probe this space, we conduct interpretability analysis and observe that layers exhibit block-wise functional similarity, while individual heads within the same layer display distinct functional specialization despite sharing input features. This head-level heterogeneity suggests that the head dimension provides a natural and principled granularity for fusing heterogeneous attention signals. Building on this insight, we introduce HydraHead, a novel architecture that hybridizes FA and LA along the head axis. HydraHead features two key innovations: (1) an interpretability-driven selection strategy that identifies retrieval-critical heads and preserves FA only for them, and (2) a scale-normalized fusion module that reconciles the distributional gap between FA and LA head outputs. By leveraging a three-stage transfer pipeline with parameter reuse and distillation, we achieve high-performance hybrid models with minimal training overhead. Under a unified training setup, HydraHead outperforms other hybrid designs in long-context tasks while maintaining strong general reasoning. With interpretability-driven head selection, it matches a 3:1 layer-wise hybrid's long-context performance at a 7:1 LA-to-FA ratio. Crucially, trained on only 15B tokens, HydraHead achieves over 69% improvement over the baseline at 512K context length, approaching Qwen3.5, a leading model of comparable size with a native context length of 256K. This highlights the significant scaling potential of head-level hybridization.
paperJun 2, 2026▲ 2
Yaosheng Fu, Guangxuan Xiao, Xin Dong et al.
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 1.25times prefill speedup and 1.7times decode speedup over the sparse-attention offload baseline. By enabling larger feasible batch sizes on a single GPU, SparDA further reaches up to 5.3times higher decode throughput than the non-offload sparse baseline. Our source code is available at https://github.com/NVlabs/SparDA.
cs.CVAug 14, 2026new
Jihun Park, Kyoungmin Lee, Jongmin Gim et al.
Subject-driven image personalization---generating new images that preserve the identity of one or several reference subjects in novel scenes---is a foundational capability for modern visual content creation. It is currently dominated by generalized methods that fine-tune a pretrained multimodal diffusion transformer (MMDiT) on hundreds of thousands to millions of paired \emph{(reference, composed-target)} examples, where each composed target is a synthesized image of the subject in a novel scene. Producing such targets demands a costly multi-stage curation pipeline---LLM-based prompt generation, T2I-based composed-target synthesis, reference-subject extraction, VLM-based quality filtering, and correspondence labeling---and tightly couples each method to a particular target synthesizer and curation choice. We introduce \emph{CRAFT} (Constrained Reward via Attention Fine-Tuning), a single-step ReFL framework that fine-tunes a pre-trained \emph{reference-aware} MMDiT via LoRA adapters using a compact reference-only data construction---$10$K reference images and subject masks, with no composed-target supervision. CRAFT realizes a \emph{Where to look} principle: attention-level rewards align noise- and phrase-token attention with the correct reference subject, and the resulting per-subject attention masks gate a pixel-level identity reward to keep image-space supervision consistent with the learned attention routing. Applied to FLUX.2-klein-9B, CRAFT achieves state-of-the-art performance on XVerseBench \rev{while using no composed-target supervision---only $10$K reference-only samples, whereas prior generalized methods require $150$K to over $2$M composed-target pairs}. The same recipe transfers to other reference-aware backbones, consistently improving performance. Project page: https://jihun999.github.io/projects/CRAFT/.
cs.CVAug 3, 2026new
Jiayu Chen, Zhikun Jiang, Maoliang Li et al.
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.
paperJul 21, 2026▲ 8
Cheng Luo, Zefan Cai, Junjie Hu
Transformers propagate information across depth through a single additive residual stream: every sublayer reads only the most recent state. Attention residuals relax this by letting each sublayer attend, through a learned softmax. However, that read uses a single query shared across the entire width, so every feature subspace must read the depth history through one distribution. The cost of this forced compromise grows with how much the subspaces disagree about which layers to read, and disagreement grows with model width. We introduce Multi-Head Attention Residuals (MHAR): the routing query is reshaped into H per-subspace heads, each with its own softmax over the depth history. The read becomes block-diagonal, the reshape adds zero parameters and negligible compute, and H = 1 recovers attention residuals exactly. Trained from scratch on a deduplicated Nemotron-based anneal corpus that is quality-filtered and STEM- and code-heavy, MHAR improves validation loss over a standard Transformer at 100M, 350M, and 1B (-0.061, -0.149, and -0.140). It achieves the best result among four methods in every setting, with the gain increasing from 100M to the larger scales. The head count is a real design axis rather than a free knob: validation loss is U-shaped with respect to H, with a flat optimum at H = 4 or H = 8 across scales. We adopt H = 8 for large-scale models; over-splitting beyond this point (H = 16) consistently gives back part of the gain. A direct probe of the trained queries confirms that learned subspace disagreement is the underlying driver. Fused Triton routing kernels increase attention-residual training throughput from 0.2-0.5x to 0.55-0.88x of the baseline while maintaining near-baseline peak memory. An identity-preserving conversion using delta attention residuals supports 8B mid-training, yielding improvements of +3.2 on GSM8K and +3.1 on GPQA.
paperJul 7, 2026▲ 13
Tommaso Cerruti, Tim Rieder, George Rowlands et al.
Self-attention lets each token retrieve information from the full context, but its quadratic cost in sequence length limits training and inference at long context. This paper presents a comparative study of softmax attention and four recent recurrent linear-attention architectures: DeltaNet, Gated DeltaNet, Kimi Delta Attention, and Gated DeltaNet-2. We express these mechanisms in a common recurrent-memory notation, making explicit how they differ in expressivity, memory decay, erase and write control, training throughput, and implementation complexity. Our experiments center on 350M-parameter models trained for 15B tokens, and include optimizer and learning-rate comparisons, hybrid-versus-pure stack comparisons, sequence-length runtime measurements, larger DeltaNet runs at 1.3B and 3B parameters, and a small set of downstream evaluations. The reported speed results measure training throughput and iteration time; we do not provide an empirical inference-speed benchmark. Within the reported 350M-parameter, 15B-token sweep, Kimi Delta Attention with Muon reaches the lowest final validation loss, a pure Gated DeltaNet stack trained with AdamW has the highest normalized training throughput, hybrid stacks generally improve loss at a throughput cost, and Muon consistently lowers final validation loss relative to AdamW in the matched architecture settings we evaluate. We introduce and evaluate lightweight cross-layer routing mechanisms for DeltaNet-style memories. The most natural DeltaNet-inspired formulation, forwarding a lower layer's delta-rule write error into the next layer's value target, does not improve over matched baselines. Routing into the aligned hidden stream and forwarding the write value instead yields a modest improvement in the matched runs we report: Cross-Layer Value Routing (CLVR) lowers final validation loss for both DeltaNet and Gated DeltaNet.
cs.HC · cs.AIJun 29, 2026new
Jessica Hutchison, Ian Tyler Applebaum, Kenneth Angelikas et al.
AI code completion tools, such as Github Copilot, provide students with code suggestions to help them write programs. However, recent qualitative studies suggest that students fail to critically evaluate these suggestions. We present Clover, a code completion tool that logs students' interactions with code suggestions and additionally offers attention checks to probe reflective engagement during programming tasks. We also develop a taxonomy of behavioral interaction metrics for AI-assisted programming, informed by literature. We analyzed relationships between interaction patterns, engagement with attention checks, and task performance. We observed that higher rates of tab accept were associated with lower attention check performance, while increased dwell time was associated with higher attention check performance. We conclude by discussing how programming process data and attention checks might support reflective engagement in AI-assisted programming.
cs.LG · cs.CVJun 18, 2026new
Przemyslaw Musialski
We place the attention token on the group: a token is an element $g_i$ of a matrix Lie group $G$ -- a bare transformation, with no feature payload and no external action $ρ(g)$ carrying it. To our knowledge this is the first attention construction whose tokens are bare matrix Lie group elements: their score is the closed-form algebra norm of the relative pose rather than a learned kernel, and it reaches the affine full-frame groups that every irrep- or surjective-exp-based method must exclude. We call it Lie-Algebra Attention. Once tokens are group elements, the rest follows with none of the usual representation-theoretic machinery. The relative geometry of a pair is canonical, $g_i^{-1} g_j$, so the pairwise invariant $w_{ij} = \log(g_i^{-1} g_j)$ is intrinsic rather than designed; equivariance under the diagonal $G$-action is tautological, and the cocycle condition holds automatically. The attention score is the negative squared algebra norm, $s_{ij} = -\|\log(g_i^{-1} g_j)\|_λ^2/τ$: the canonical proximity kernel under a block-weighted Frobenius inner product, with no irreducible representations, spherical harmonics, Clebsch-Gordan products, or learned kernel. The construction applies to any matrix Lie group on a chosen logarithm chart containing the relative poses, including the non-compact non-abelian affine groups with scale and shear that no vector-token attention method reaches: neither the irrep tradition nor surjective-exp methods. Three sequence-completion experiments, on SE(2), SO(3), and Aff(2), bear this out: the closed-form score matches a learned MLP kernel on the same invariant and outperforms it on SE(2), using 50 to 80x fewer score parameters, while a vector-token baseline breaks invariance by five to twelve orders of magnitude.
paperJun 17, 2026▲ 54
Vishesh Tripathi, Abhay Kumar
Self-attention is central to Transformer performance and is often the most expensive part of the Transformer at long context lengths because its pairwise token interactions scale quadratically with sequence length. Standard dense attention also applies the same set of attention heads to every token regardless of token difficulty or information content. This uniform activation can waste compute, especially as sequences grow longer and attention cost increases rapidly. We propose Grouped Query Experts (GQE), a mixture-of-experts layer on top of grouped-query attention (GQA). Within each GQA group, a router selects k query-head experts per token while all key-value (KV) heads remain dense and unchanged. Thus, GQE keeps the KV cache benefits of GQA and reduces only the active query-head computation. On a fixed 30B token budget at the 250M parameter scale, GQE matches the all-active GQA baseline in downstream accuracy while activating half the query heads per token.
paperJun 8, 2026▲ 2
Xin Wang, Hui Shen, Boyuan Zheng et al.
The scalability of Large Language Models (LLMs) to long contexts is fundamentally constrained by the quadratic complexity of standard attention, motivating the adoption of linear attention mechanisms with sub-quadratic cost. To improve representation capacity under long contexts, recent approaches organize memory in a multi-state manner. However, existing multi-state linear attention methods rely on fixed state merging policies that cannot adapt to dynamically varying token importance, irreversibly obscuring critical tokens and causing severe error accumulation over long sequences. To address this limitation, we propose DLA, a dynamic memory modeling framework for multi-state linear attention. DLA introduces (i) Information-Aware Dynamic State Merging, which adaptively determines state boundaries based on token-level information variation, preserving high-resolution representations around semantic transitions while aggressively summarizing stable regions, and (ii) Capacity-Bounded Memory Modeling, which maintains a fixed-size, chronologically ordered state cache by selectively merging adjacent low-information states to control memory growth with minimal information loss. We pre-train DLA on two different linear attention models and evaluate on 16 datasets across three categories. Experimental results demonstrate the superiority of DLA over state-of-the-art.
paperAug 6, 2026▲ 1
Yingtao Ren, Ziyi Zhao, Yiwei Fu et al.
Retrieval-augmented generation (RAG) is indispensable for enhancing large language models. However, RAGs are increasingly susceptible to poisoning attacks, in which adversarial documents are injected to manipulate generator outputs. Previous methods rely on output-side signals such as perplexity and consistency checks to detect such attacks. Nevertheless, our analysis reveals that deliberate attacks often induce false confidence, where poisoned outputs exhibit even lower perplexity than benign ones, rendering uncertainty-based detection ineffective. To address this challenge, we explore the internal dynamics of the generator and identify a distinctive signature termed Attention Collapse. Unlike the dispersed attention in benign generations, attacked generations exhibit a decrease in entropy as attention concentrates on poisoned documents. Building on these findings, we propose D-SCAN (Document-level Signal Collapse Analysis), a lightweight detection framework that monitors attention dynamics to identify attacked generations. Extensive experiments on multiple attack benchmarks demonstrate the effectiveness of our method. Moreover, D-SCAN can detect attacks even when they fail to alter the final answer. Code is available at https://github.com/yingtaoren/D-Scan.git.
cs.CVJun 25, 2026new
Ron Campos, Subhajit Maity, Xin Li et al.
The softmax activation in multihead attention (MHA) is the de facto standard for attention-based models in visual perception tasks. However, standard softmax can produce noisy attention patterns that dilute relevant features and degrade its performance. In this paper, we propose Denoising Attention or DnA, in which, first, a positive query identifies which image features belong to the correct class, and a negative query identifies closely associated but irrelevant image features. DnA then projects these interactions into two distinct subspaces with larger principal angles, promoting subspace separation and improved discriminability. Using a ViT-B backbone, our proposed DnA achieves an absolute gain of 0.8% on ImageNet-1K compared to the baseline. We further show improvements across multiple visual understanding tasks, including video understanding with video transformers (1.8%) and video LLMs (0.5%). Our extensive empirical analyses justify the design choices involving two interacting subspaces and the denoising effect of DnA.
cs.CR · cs.AIJun 24, 2026new
Tânia Carvalho, Maxime Cordy
Tabular foundation models are commonly assumed to present limited privacy concerns as they are often pre-trained on large collections of synthetic data. However, these models leverage in-context learning, where sensitive records may be provided directly at inference time as labelled context examples. In this paper, we demonstrate that predictions generated via the attention mechanism leak sufficient information to enable effective Membership Inference Attacks (MIAs). To highlight this vulnerability, we propose AMIA (Attention-based Membership Inference Attack), a shadow-model-free attack that exploits the concentration of transformer attention patterns. Our results show that attention mechanisms reveal strong membership signals, which exceed classical confidence-based attacks, achieving an average gain of 7.7\%, specially in low false-positive regimes. To mitigate this risk, we introduce an inference-time defence inspired by $k$-anonymity principles. This approach reduces the uniqueness of context-key representations without introducing random noise or retraining the model. By targeting only high-risk queries identified through AMIA scores, the defence substantially reduces membership leakage of this attack by an average of 50\% and 25\% against confidence-based attacks, while preserving predictive utility with only 3.9\% performance degradation. Beyond showing that context examples are vulnerable, we further demonstrate that fine-tuning introduces an additional source of privacy risk. In particular, samples whose prediction confidence increases after fine-tuning become more susceptible to MIAs, indicating that fine-tuning can amplify memorisation and expose sensitive training information through confidence shifts.
cs.CVJun 23, 2026new
Gargi Panda, Soumitra Kundu, Saumik Bhattacharya et al.
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.
cs.AIJun 18, 2026new
Nityanand Mathur, Hamees Sayed, Wasim Madha et al.
Style-captioned text-to-speech systems use natural language to control voice characteristics, but how individual words influence acoustic output remains unclear. Understanding this is critical for diagnosing failure modes and improving controllability in expressive TTS. We propose cross-attention attribution for speech diffusion models, adapting the DAAM framework to the speech domain for the first time, and apply it to CapSpeech-TTS. Our method extracts per-token heatmaps across 25 layers and 24 ODE steps. We analyze 3,600 (style caption, text transcript) combinations comprising 120 style captions conditioning the generation of 30 text transcripts each, revealing how caption tokens shape waveforms. Results show: (1) style tokens have lower temporal variance than content/function tokens, confirming global conditioning; (2) style attention correlates with F0 and energy; (3) style conditioning peaks in early steps and deep layers; (4) attention entropy reaches its minimum at layer 17, co-occurring with the style importance peak, indicating maximal network selectivity at the most style-critical stage. This is the first study of how natural language influences cross-attention in speech diffusion models
paperJun 14, 2026▲ 5
Zhongzhu Zhou, Qingyang Wu, Junxiong Wang et al.
Hybrid linear attention models offer an appealing path to faster long-context inference: they reduce the quadratic cost and KV-cache burden of full softmax attention while retaining much of the quality of Transformer models. A practical way to obtain such models is to convert a pretrained Transformer instead of pretraining a new architecture from scratch, but this conversion is still brittle. Simply copying the teacher attention projections into a Gated DeltaNet (GDN) student does not specify the new recurrent decay, write, and output-gating dynamics. As a result, the converted model often starts in a poor dynamical regime and must spend many distillation tokens repairing initialization rather than learning the remaining teacher behavior. We propose Taylor-Calibrate, a lightweight initialization method for hybrid GDN students. The method uses Taylor-guided teacher attention statistics to set the value projection, memory timescale, write gates, and output gate, then applies a short per-layer alignment step to match each converted layer to the teacher output. Across four teacher settings and three retained-layer policies, Taylor-Calibrate gives substantially stronger zero-shot students, with up to an 88x improvement in a representative ablation, and reaches matched recovery targets with 4.9x--9.2x fewer training tokens than naive conversion.
cs.SD · cs.AIJun 12, 2026new
Ravi Ranjan, Utkarsh Grover, Xiaomin Lin et al.
Transformer-based automatic speech recognition (ASR) models such as Whisper are highly accurate, but their predictions remain difficult to interpret. Existing explainable AI (XAI) methods often lack faithfulness and precise temporal grounding. We propose Listening with Entropy-guided Attention for Faithful explainability (LEAF-X), a model-intrinsic XAI framework for transformer-based ASR. LEAF-X combines entropy-guided attention weighting, multi-layer attention rollout, and optional causal ablations to identify low-entropy, high-impact heads and layers, producing sparse token-to-frame attributions. Unlike perturbation-based explainers or raw attention maps, LEAF-X exploits the internal structure of encoder-decoder and speech-augmented decoder-only models to generate explanations that better reflect model computation. Results show 32% improved faithfulness, 35-39% stronger locality/sparsity, and the most stable attributions, supporting more transparent and auditable ASR.
paperMay 26, 2026new
Xiuying Wei, Caglar Gulcehre
Efficient inference is critical for long-context language models, where attention computation and KV-cache access dominate the cost. Recent work RAT+, introduces a recurrence-augmented attention backbone that enables flexible dilated attention at inference time. In this paper, we investigate whether this exponentially decaying memory can also improve existing query-aware sparse inference methods. Using representative methods including Quest, MoBA, and SnapKV, we show that RAT+ consistently improves accuracy over standard attention across sparse budgets on eight needle-in-a-haystack tasks. We validate these gains both on the released checkpoints from the RAT+ paper and on OLMo2-7B, which we continue pretraining with the added memory module for 10B tokens. Finally, we propose two hypotheses explaining why this memory module benefits query-aware sparse inference and design targeted experiments to support them.
paperNeurIPS 202520252.2k citations
Yunjie Tian, Qixiang Ye, David S. Doermann
Enhancing the network architecture of the YOLO framework has been crucial for a long time, but has focused on CNN-based improvements despite the proven superiority of attention mechanisms in modeling capabilities. This is because attention-based models cannot match the speed of CNN-based models. This paper proposes an attention-centric YOLO framework, namely YOLOv12, that matches the speed of previous CNN-based ones while harnessing the performance benefits of attention mechanisms. YOLOv12 surpasses all popular real-time object detectors in accuracy with competitive speed. For example, YOLOv12-N achieves 40.6% mAP with an inference latency of 1.64 ms on a T4 GPU, outperforming advanced YOLOv10-N / YOLOv11-N by 2.1%/1.2% mAP with a comparable speed. This advantage extends to other model scales. YOLOv12 also surpasses end-to-end real-time detectors that improve DETR, such as RT-DETR / RT-DETRv2: YOLOv12-S beats RT-DETR-R18 / RT-DETRv2-R18 while running 42% faster, using only 36% of the computation and 45% of the parameters. More comparisons are shown in Figure 1.
paperNeurIPS 20242024633 citations
Jay Shah, Ganesh Bikshandi, Ying Zhang et al.
Attention, as a core layer of the ubiquitous Transformer architecture, is the bottleneck for large language models and long-context applications. FlashAttention elaborated an approach to speed up attention on GPUs through minimizing memory reads/writes. However, it has yet to take advantage of new capabilities present in recent hardware, with FlashAttention-2 achieving only 35% utilization on the H100 GPU. We develop three main techniques to speed up attention on Hopper GPUs: exploiting asynchrony of the Tensor Cores and TMA to (1) overlap overall computation and data movement via warp-specialization and (2) interleave block-wise matmul and softmax operations, and (3) block quantization and incoherent processing that leverages hardware support for FP8 low-precision. We demonstrate that our method, FlashAttention-3, achieves speedup on H100 GPUs by 1.5-2.0$\times$ with FP16 reaching up to 740 TFLOPs/s (75% utilization), and with FP8 reaching close to 1.2 PFLOPs/s. We validate that FP8 FlashAttention-3 achieves 2.6$\times$ lower numerical error than a baseline FP8 attention.
paperNeurIPS 20242024425 citations
Huiqiang Jiang, Yucheng Li, Chengruidong Zhang et al.
The computational challenges of Large Language Model (LLM) inference remain a significant barrier to their widespread deployment, especially as prompt lengths continue to increase. Due to the quadratic complexity of the attention computation, it takes 30 minutes for an 8B LLM to process a prompt of 1M tokens (i.e., the pre-filling stage) on a single A100 GPU. Existing methods for speeding up prefilling often fail to maintain acceptable accuracy or efficiency when applied to long-context LLMs. To address this gap, we introduce MInference (Milliontokens Inference), a sparse calculation method designed to accelerate pre-filling of long-sequence processing. Specifically, we identify three unique patterns in long-context attention matrices-the A-shape, Vertical-Slash, and Block-Sparsethat can be leveraged for efficient sparse computation on GPUs. We determine the optimal pattern for each attention head offline and dynamically build sparse indices based on the assigned pattern during inference. With the pattern and sparse indices, we perform efficient sparse attention calculations via our optimized GPU kernels to significantly reduce the latency in the pre-filling stage of long-context LLMs. Our proposed technique can be directly applied to existing LLMs without any modifications to the pre-training setup or additional fine-tuning. By evaluating on a wide range of downstream tasks, including InfiniteBench, RULER, PG-19, and Needle In A Haystack, and models including LLaMA-3-1M, GLM4-1M, Yi-200K, Phi-3-128K, and Qwen2-128K, we demonstrate that MInference effectively reduces inference latency by up to 10x for pre-filling on an A100, while maintaining accuracy. Our code is available at https://aka.ms/MInference.