To overcome the quadratic cost of self-attention, recent works have proposed various sparse attention modules, most of which fall under one of two groups: 1) sparse attention under a hand-crafted patterns and 2) full attention followed by a sparse variant of softmax such as $\alpha$-entmax. Unfortunately, the first group lacks adaptability to data while the second still requires quadratic cost in training. In this work, we propose SBM-Transformer, a model that resolves both problems by endowing each attention head with a mixed-membership Stochastic Block Model (SBM). Then, each attention head data-adaptively samples a bipartite graph, the adjacency of which is used as an attention mask for each input. During backpropagation, a straight-through estimator is used to flow gradients beyond the discrete sampling step and adjust the probabilities of sampled edges based on the predictive loss. The forward and backward cost are thus linear to the number of edges, which each attention head can also choose flexibly based on the input. By assessing the distribution of graphs, we theoretically show that SBM-Transformer is a universal approximator for arbitrary sequence-to-sequence functions in expectation. Empirical evaluations under the LRA and GLUE benchmarks demonstrate that our model outperforms previous efficient variants as well as the original Transformer with full attention.

Large Language Models (LLMs) are increasingly prevalent in the field of long-context modeling, however, their inference computational costs have become a critical bottleneck hindering the advancement of tasks such as agents and multimodal applications. This report conducts a preliminary investigation into the effectiveness and theoretical mechanisms of the Top-$k$ Attention mechanism during both the decoding and training phases. First, we validate the effectiveness of exact Top-$k$ Decoding through extensive experimentation. Experiments demonstrate that retaining only the pivotal Keys with the highest similarity to the Query as the context window during the decoding stage achieves performance comparable to, or even surpassing, full attention on downstream tasks such as HELMET and LongBench v2. Second, we further explore the native Top-$k$ Attention training strategy. Experiments confirm that ensuring the consistency between training and inference regarding Top-$k$ Attention operations facilitates the further unlocking of Top-$k$ Decoding's potential, thereby significantly enhancing model performance. Furthermore, considering the high computational complexity of exact Top-$k$ Attention, we investigate the impact of approximate Top-$k$ algorithm precision on downstream tasks. Our research confirms a positive correlation between downstream task performance and approximation fidelity, and we provide statistical evaluations of the Lightning Indexer's precision within the DeepSeek-V3.2-Exp model. Finally, this report provides a theoretical interpretation from the perspective of Entropy. Experimental observations indicate that models subjected to Top-$k$ Attention SFT exhibit a distinct phenomenon of entropy reduction in downstream tasks, which validates the hypothesis that low-entropy states are better adapted to Top-$k$ Decoding.

Reasoning language models have demonstrated remarkable capabilities on challenging tasks by generating elaborate chain-of-thought (CoT) solutions. However, such lengthy generation shifts the inference bottleneck from compute-bound to memory-bound. To generate each token, the model applies full attention to all previously generated tokens, requiring memory access to an increasingly large KV-Cache. Consequently, longer generations demand more memory access for every step, leading to substantial pressure on memory bandwidth. To address this, we introduce SparseSpec, a speculative decoding framework that reuses the same model as the draft and target models (i.e., self-speculation). SparseSpec features a novel sparse attention mechanism, PillarAttn, as the draft model, which accurately selects critical tokens via elegantly reusing information from the verification stage. Furthermore, SparseSpec co-designs self-speculation with three system innovations: (1) a unified scheduler to batch token drafting and verification, (2) delayed verification for CPU/GPU overlap, and (3) dynamic KV-Cache management to maximize memory utilization. Across various models and datasets, SparseSpec outperforms state-of-the-art solutions, with an up to 2.13x throughput speedup.

Diffusion Transformers dominate video generation, but the quadratic complexity of attention computation introduces substantial latency. Attention sparsity reduces computational costs by focusing on critical tokens while ignoring non-critical tokens. However, existing methods suffer from severe performance degradation. In this paper, we revisit attention sparsity and reveal that existing methods induce systematic biases in attention allocation: (1) excessive focus on critical tokens amplifies their attention weights; (2) complete neglect of non-critical tokens causes the loss of relevant attention weights. To address these issues, we propose Rectified SpaAttn, which rectifies attention allocation with implicit full attention reference, thereby enhancing the alignment between sparse and full attention maps. Specifically: (1) for critical tokens, we show that their bias is proportional to the sparse attention weights, with the ratio governed by the amplified weights. Accordingly, we propose Isolated-Pooling Attention Reallocation, which calculates accurate rectification factors by reallocating multimodal pooled weights. (2) for non-critical tokens, recovering attention weights from the pooled query-key yields attention gains but also introduces pooling errors. Therefore, we propose Gain-Aware Pooling Rectification, which ensures that the rectified gain consistently surpasses the induced error. Moreover, we customize and integrate the Rectified SpaAttn kernel using Triton, achieving up to 3.33 and 2.08 times speedups on HunyuanVideo and Wan 2.1, respectively, while maintaining high generation quality. We release Rectified SpaAttn as open-source at https://github.com/BienLuky/Rectified-SpaAttn .

In Diffusion Transformer (DiT) models, particularly for video generation, attention latency is a major bottleneck due to the long sequence length and the quadratic complexity. We find that attention weights can be separated into two parts: a small fraction of large weights with high rank and the remaining weights with very low rank. This naturally suggests applying sparse acceleration to the first part and low-rank acceleration to the second. Based on this finding, we propose SLA (Sparse-Linear Attention), a trainable attention method that fuses sparse and linear attention to accelerate diffusion models. SLA classifies attention weights into critical, marginal, and negligible categories, applying O(N^2) attention to critical weights, O(N) attention to marginal weights, and skipping negligible ones. SLA combines these computations into a single GPU kernel and supports both forward and backward passes. With only a few fine-tuning steps using SLA, DiT models achieve a 20x reduction in attention computation, resulting in significant acceleration without loss of generation quality. Experiments show that SLA reduces attention computation by 95% without degrading end-to-end generation quality, outperforming baseline methods. In addition, we implement an efficient GPU kernel for SLA, which yields a 13.7x speedup in attention computation and a 2.2x end-to-end speedup in video generation on Wan2.1-1.3B. The code is available at https://github.com/thu-ml/SLA.

Figure 1: Flow-chart demonstrating the compuation for clustered attention. For more details refer to 1.1 or 3.2 in the main paper. Work done at Idiap 34th Conference on Neural Information Processing Systems (NeurIPS 2020), V ancouver, Canada. We then present the flow chart demonstrating the same. This is followed by taking the weighted average of the 3 correponding values.

Large language models (LLMs) now support context windows of hundreds of thousands to millions of tokens, enabling applications such as long-document summarization, large-scale code synthesis, multi-document question answering and persistent multi-turn dialogue. However, such extended contexts exacerbate the quadratic cost of self-attention, leading to severe latency in autoregressive decoding. Existing sparse attention methods alleviate these costs but rely on heuristic patterns that struggle to recall critical key-value (KV) pairs for each query, resulting in accuracy degradation. We introduce Adamas, a lightweight yet highly accurate sparse attention mechanism designed for long-context inference. Adamas applies the Hadamard transform, bucketization and 2-bit compression to produce compact representations, and leverages Manhattan-distance estimation for efficient top-k selections. Experiments show that Adamas matches the accuracy of full attention with only a 64-token budget, achieves near-lossless performance at 128, and supports up to 8x higher sparsity than prior state-of-the-art (SOTA) methods while delivering up to 4.4x self-attention and 1.5x end-to-end speedups on 32K-length sequences. Remarkably, Adamas attains comparable or even lower perplexity than full attention, underscoring its effectiveness in maintaining accuracy under aggressive sparsity.

Recent advance in sparse attention mechanisms has demonstrated strong potential for reducing the computational cost of long-context training and inference in large language models (LLMs). Native Sparse Attention (NSA), one state-of-the-art approach, introduces natively trainable, hardware-aligned sparse attention that delivers substantial system-level performance boost while maintaining accuracy comparable to full attention. However, the kernel implementation of NSA forces a loop order that is only efficient with a relatively large number of query heads in each Grouped Query Attention (GQA) group, whereas existing LLMs widely adopt much smaller number of query heads in each GQA group -- such an inconsistency significantly limits the applicability of this sparse algorithmic advance. In this work, we propose Flash Sparse Attention (FSA), an alternative kernel implementation that enables efficient NSA computation across a wide range of popular LLMs with varied smaller number of heads in each GQA group on modern GPUs. Compared to vanilla NSA kernel implementation, our empirical evaluation demonstrates that FSA achieves (i) up to 3.5x and on average 1.6x kernel-level latency reduction, (ii) up to 1.25x and 1.09x on average end-to-end training speedup on state-of-the-art LLMs, and (iii) up to 1.36x and 1.11x on average for prefill-phase speedup in LLM generative inference. Github Repo at https://github.com/Relaxed-System-Lab/Flash-Sparse-Attention.