From 070a3161e7042d92e2aa726fda4ffdcff1db18c1 Mon Sep 17 00:00:00 2001 From: Farhad Ramezanghorbani Date: Thu, 22 Oct 2026 18:14:42 -0600 Subject: [PATCH] docs: link HyenaND paper from README Signed-off-by: Farhad Ramezanghorbani --- README.md | 2 ++ 1 file changed, 2 insertions(+) diff --git a/README.md b/README.md index f564c825..25ca9148 100644 --- a/README.md +++ b/README.md @@ -10,6 +10,8 @@ keeps its last result while runs wait for a down runner. A unified PyTorch-native library for subquadratic alternatives to quadratic attention methods. +**Paper:** [Native Multi-Dimensional Subquadratic Operators via Input Dependent Long Convolutions](https://arxiv.org/abs/2607.19378) (HyenaND). + ## Overview nvSubquadratic consolidates efforts from across NVIDIA Research teams (nvResearch, NeMo, BioNeMo) into a single, consistent API. Currently supporting multi-dimensional (1D, 2D, 3D) Hyena operators with optimized CUDA kernels. Hyena operators provide subquadratic alternatives to attention mechanisms, achieving O(N) or O(N log N) complexity compared to the O(N²) scaling of traditional attention.