Autoregressive decoding becomes bandwidth-limited at long contexts, as generating each token requires reading all nk key and value vectors from KV cache. We present Stochastic Additive No-mulT Attention (SANTA), a method that sparsifies value-cache access by sampling S≪nk indices from the post-softmax distribution and aggregates only those value rows. This yields an unbiased estimator of the post-softmax value aggregation while replacing value-stage multiply-accumulates with gather-and-add. We introduce stratified sampling to design variance-reduced, GPU-friendly variants, demonstrating 1.5× decode-step attention kernel speedup over FlashInfer and FlashDecoding on an NVIDIA RTX 6000 Ada while matching baseline accuracy at 32k-token contexts. Finally, we propose Bernoulli qKT sampling as a complementary technique to sparsify the score stage, reducing key-feature access through stochastic ternary queries. Both methods are orthogonal to upstream techniques such as ternary quantization, low-rank projections, and KV-cache compression. Together, they point toward sparse, multiplier-free, and energy-efficient inference. We open-source our kernels at: https://github.com/OPUSLab/SANTA.git
Cite
@article{arxiv.2605.01910,
title = {Stochastic Sparse Attention for Memory-Bound Inference},
author = {Kyle Lee and Corentin Delacour and Kevin Callahan-Coray and Kyle Jiang and Can Yaras and Samet Oymak and Tathagata Srimani and Kerem Y. Camsari},
journal= {arXiv preprint arXiv:2605.01910},
year = {2026}
}
Comments
Accepted to ICML 2026. Code available at https://github.com/OPUSLab/SANTA