CodeAuditor

WildCat Attention Implementation Issues

WildCat: Near-Linear Attention in Theory and Practice

Read the original source

What the paper says

WildCat achieves super-polynomial error decay O(n^{-√log(log(n))}) in near-linear O(n^{1+o(1)}) time.

The Critique

Runtime analysis shows only 1.24x speedup vs standard attention at n=8192, while random sampling achieves 1.62x. Missing baselines: uniform random sampling, attention weight thresholding. 'Near-linear' hides large constants. No end-to-end model quality degradation measured.

Why It Matters

If theoretical advantages don't translate to practical speedups, research effort may be misdirected. Understanding when sophisticated approximation is worth complexity is crucial.

What They Missed

The error bound O(n^{-√log(log(n))}) decays extremely slowly—for n=1e6, √log(log(n)) ≈ 1.4, so error decays as n^{-1.4}. This is polynomial, not 'super-polynomial' in any practical sense.

The Big Question

When do WildCat's theoretical advantages actually materialize in practice, and how does approximation error affect downstream task performance?