Where the browser forward pass actually spends its time, measured in this tab right now. This page exists because a performance assumption in this project turned out to be wrong and was only caught by measuring. Two sequence lengths: T=77 is what the explainer runs on every interaction, T=154 is two full periods. ← back to the explainer · parity check
running…
How to read it. xs and ys are the two dense
encoder matmuls, scores is the causally-masked Gram matrix, rope is
the rotary phase application, nzscan indexes the non-zero entries so the score
loop can skip zeros, and count is the sparsity measurement itself.
What it is used for. The two 4-way unrolled loops in bdh.js were
added on a hypothesis and first measured at no effect, on a machine that was running two
training jobs at the time. Re-measured on a quiet machine they are worth roughly 1.9× on
xs and 2.0× on ys. A benchmark taken on a busy machine is not a
weak result, it is not a result. Interaction latency end to end is recorded separately in
experiments/results/latency.md.