Scout Before You Attend: Sketch-and-Walk Sparse Attention for Efficient LLM Inference

Hoang Anh Duy Le, Sahil Joshi, Zeyu Yang, Zhaozhuo Xu, Anshumali Shrivastava
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:63033-63059, 2026.

Abstract

Self-attention dominates the computational and memory cost of long-context LLM inference across both prefill and decode phases. To address this challenge, we introduce Sketch&Walk Attention, a training-free sparse attention method that determines sparsity with lightweight sketches and deterministic walk. Sketch&Walk applies Hadamard sketching to get inexpensive approximations of attention scores, then aggregates these estimates across layers via a walk mechanism that captures attention influence beyond direct interactions between tokens. The accumulated walk scores are used to select top-$k$ attention blocks, enabling dynamic sparsity with a single training-free algorithm that applies uniformly to both the prefill and decode phases, together with custom sparse attention kernels. Across a wide range of models and tasks, Sketch&Walk maintains near-lossless accuracy at 20% attention density and can slightly outperform dense attention in some settings, while achieving up to $4.7\times$ end-to-end attention speedup over FlashAttention-2.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-le26c, title = {Scout Before You Attend: Sketch-and-Walk Sparse Attention for Efficient {LLM} Inference}, author = {Le, Hoang Anh Duy and Joshi, Sahil and Yang, Zeyu and Xu, Zhaozhuo and Shrivastava, Anshumali}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {63033--63059}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/le26c/le26c.pdf}, url = {https://proceedings.mlr.press/v306/le26c.html}, abstract = {Self-attention dominates the computational and memory cost of long-context LLM inference across both prefill and decode phases. To address this challenge, we introduce Sketch&Walk Attention, a training-free sparse attention method that determines sparsity with lightweight sketches and deterministic walk. Sketch&Walk applies Hadamard sketching to get inexpensive approximations of attention scores, then aggregates these estimates across layers via a walk mechanism that captures attention influence beyond direct interactions between tokens. The accumulated walk scores are used to select top-$k$ attention blocks, enabling dynamic sparsity with a single training-free algorithm that applies uniformly to both the prefill and decode phases, together with custom sparse attention kernels. Across a wide range of models and tasks, Sketch&Walk maintains near-lossless accuracy at 20% attention density and can slightly outperform dense attention in some settings, while achieving up to $4.7\times$ end-to-end attention speedup over FlashAttention-2.} }
Endnote
%0 Conference Paper %T Scout Before You Attend: Sketch-and-Walk Sparse Attention for Efficient LLM Inference %A Hoang Anh Duy Le %A Sahil Joshi %A Zeyu Yang %A Zhaozhuo Xu %A Anshumali Shrivastava %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-le26c %I PMLR %P 63033--63059 %U https://proceedings.mlr.press/v306/le26c.html %V 306 %X Self-attention dominates the computational and memory cost of long-context LLM inference across both prefill and decode phases. To address this challenge, we introduce Sketch&Walk Attention, a training-free sparse attention method that determines sparsity with lightweight sketches and deterministic walk. Sketch&Walk applies Hadamard sketching to get inexpensive approximations of attention scores, then aggregates these estimates across layers via a walk mechanism that captures attention influence beyond direct interactions between tokens. The accumulated walk scores are used to select top-$k$ attention blocks, enabling dynamic sparsity with a single training-free algorithm that applies uniformly to both the prefill and decode phases, together with custom sparse attention kernels. Across a wide range of models and tasks, Sketch&Walk maintains near-lossless accuracy at 20% attention density and can slightly outperform dense attention in some settings, while achieving up to $4.7\times$ end-to-end attention speedup over FlashAttention-2.
APA
Le, H.A.D., Joshi, S., Yang, Z., Xu, Z. & Shrivastava, A.. (2026). Scout Before You Attend: Sketch-and-Walk Sparse Attention for Efficient LLM Inference. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:63033-63059 Available from https://proceedings.mlr.press/v306/le26c.html.

Related Material