perf(cuda): measure SDPA plan-cache bucketing, recommend against it - #1906

Merged

inureyes merged 13 commits into

mainfrom

update/issue-1820-sdpa-plan-cache-bucket

Sep 19, 2026

Commits on Sep 17, 2026

perf(cuda): bucket the cuDNN SDPA plan-cache key
inureyes
committed
5c3e7a8
docs(benchmark): record the #1820 plan-cache bucketing measurements
inureyes
committed
16c6d82
docs(benchmark): preserve the partial #1820 context-ladder sweep
inureyes
committed
04fe01f
fix(cuda): separate the causal and array-mask axes in the SDPA gate
inureyes
committed
0a9f984
docs(benchmark): re-baseline the #1820 short-context sweep after the rebase
inureyes
committed
192784b
docs(benchmark): record the #1820 ladder 2634-token rung
inureyes
committed
af25d03
docs(benchmark): preserve the halted #1820 8k rung as raw evidence
inureyes
committed
e7bc8c2
test(bench): gate the #1820 harness on the driver, memory and CI
inureyes
committed
e0a6fe2
test(bench): count the node toolchain as host-busy in the #1820 gate
inureyes
committed
fa76f7d
docs(benchmark): quarantine the unusable #1820 ladder rows
inureyes
committed
773e4d2
docs(benchmark): record the #1820 decision against shipping bucketing
inureyes
committed
0e6e06a
perf(cuda): default the SDPA plan-cache bucketing off
inureyes
committed
7a43b85

Commits on Sep 19, 2026

docs(benchmark): extract the #1820 traces and cite the quarantine file
inureyes
committed
995c152