Spend context with intent.
See what a sparse connectivity pattern changes in the pair budget before you profile a real kernel. The arithmetic is explicit; the benchmark boundary is too.
Pattern inputs
tokens, not weightstokens
keys per query, nominal
broadcast keys
Ready to compare an attention pattern.
Boundary: this models pair connectivity. It does not run custom kernels, use a GPU, measure throughput, or evaluate quality.
Dense pairs--
Sparse pairs--
Theoretical reduction--
Fewer-pairs ratio--
Connection budget
Awaiting comparisonPattern
Choose inputs to show the local-plus-global pattern.
Benchmark boundary
The result will explain why pair reduction is not measured speed.
Calculation trace
No result yetLoad the sample and compare to show the equations.
Computed theory appears before export.