Cascade Token Selection for Transformer Attention Acceleration
Abstract
A method is presented for reducing the cost of representative token selection in transformer attention layers by exploiting the coherence of the representative set across depth. Activation Decorrelation Attention (ADA) selects representative tokens at each layer via a Gram threshold and computes attention on the compressed problem, but the selection requires a Gram matrix at every layer. The cascade mechanism introduced here inherits the representative set from layer to layer , validates it via a cross-Gram computation, and updates it with a small number of additions and removals. The cost of the selection step drops from to per layer. Validation on three model families (GPT-2 124M, GPT-J 6B, OPT 6.7B) on AMD MI300X demonstrates Gram operation savings of to with mean Jaccard overlap of to between consecutive layers. The cascade reveals that the set of informative tokens is a structural property of the input that propagates coherently through the depth of the network: the same tokens carry the non-redundant information at layer and at layer .
Cite
@article{arxiv.2605.03110,
title = {Cascade Token Selection for Transformer Attention Acceleration},
author = {Stephen J. Thomas},
journal= {arXiv preprint arXiv:2605.03110},
year = {2026}
}