Replace attention in a PyTorch model with HyenaND from the nvSubquadratic library. Covers 1D / 2D / 3D hosts (ViT, U-Net, diffusion, causal LM, hierarchical encoders). Trigger when the user wants a subquadratic alternative to attention, ports a model to HyenaND, swaps `nn.MultiheadAttention` or `F.scaled_dot_product_attention` for a Hyena mixer, builds a striped Hyena LM, or asks "how do I use nvSubquadratic with my model." Phrases like "make my ViT subquadratic," "Hyena layer for my U-Net," "swap attention with FFT convolution," "subquadratic alternative for my 3D segmentation network," or "long-context model with O(L log L) scaling" should all activate this skill.
2026-06-03