ClusterAttention is a training‑free technique that speeds up bidirectional attention by recursively clustering keys and queries into fixed‑size, power‑of‑two blocks, enabling block‑sparse attention to match dense attention latency on GPUs. The method derives error bounds for sparse attention, showing tighter clusters can reduce error when compensated via centroids, and demonstrates significant speedups—up to six‑fold on large tabular data and 1.8× on video generation—while preserving over 99% of dense accuracy.
arXiv:2607. 03012v1 Announce Type: cross Abstract: Video Diffusion Transformers (VDiTs) have demonstrated significant capabilities in high-fidelity video generation.
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.
arXiv:2508. 18224v3 Announce Type: replace-cross Abstract: Recent advances in sparse attention mechanisms have demonstrated strong potential for reducing the computational cost of long-context training and inference in large language models (LLMs).
The paper introduces SparsePR, a training‑free block‑sparse attention method for video transformers that partitions query‑key responses and reconstructs the residual via probe‑fitted affine corrections. By pairing sampled‑query key responses into K/V groups and using centroids to guide shared routing, SparsePR reduces attention‑reconstruction error across diverse video generation and world‑model tasks. Experiments show consistent error reductions, with probe fitting contributing most of the improvement, while maintaining generation quality at 22.0–26.0% executed‑pair density and delivering 1.48×–2.61× speedups.
arXiv:2607. 13241v1 Announce Type: cross Abstract: Traffic forecasting is highly challenging due to complex and nonlinear spatial and temporal dependencies.
arXiv:2608. 12032v1 Announce Type: cross Abstract: 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.
The paper introduces a reproducible benchmark for evaluating attention mechanisms in tabular foundation models, focusing on the distinct row and column attention patterns that differ from language model attention. It compares several backends—Torch SDPA, FlashAttention variants, vLLM, and SageAttention—across realistic tabular shapes on A100, H100, and B200 GPUs, revealing that optimal backend choice varies by attention type, hardware, and model specifics. The study finds FlashAttention generally performs best, but CuDNN can outperform it for column attention on longer sequences, while SageAttention excels for large row sequences beyond 16k rows.
arXiv:2609.15810v1 Announce Type: new Abstract: Diffusion Transformers deliver state-of-the-art video generation, but their long spatiotemporal sequences make attention the dominant deployment cost,...
arXiv:2605. 18848v3 Announce Type: replace Abstract: This paper introduces Exact Linear Attention (ELA), a mechanism that achieves linear computational complexity for Transformer attention by exploiting the exact decomposition property of kernel functions, thereby eliminating approximation error.
RBS-Attention introduces a training‑free, radius‑bounded sparse prefill strategy for long‑context large language models, addressing the mean dilution problem where a block centroid can miss highly relevant tokens. The method employs two complementary selection branches: a centroid base branch that captures average relevance and a rescue branch that uses the maximum key‑block radius to flag under‑estimated blocks. Experiments on Qwen3 models demonstrate significant speedups—over 20× in standalone prefill‑attention and nearly 6× in end‑to‑first‑token time—while maintaining competitive accuracy compared to dense attention.
arXiv:2606. 13392v1 Announce Type: new Abstract: 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.
arXiv:2609.37001v1 Announce Type: cross Abstract: Diffusion Transformers (DiTs) enable high-quality video generation but suffer from substantial inference latency, primarily attributable to the compu...