In May, KDA took top-three placements on every track of the MLSys 2026 FlashInfer kernel contest. That was a good result against the field. It was not a good result against the field's best entries, which were written by people.
KDA 1.5 changes three things and reruns the contest:
- CuteDSL support, so the agent can express the layouts it wants rather than the ones the previous toolchain made convenient;
- the IKET profiling skill, which turns a profiler report into something an agent can reason from instead of a wall of counters; and
- a better agent loop flow — the arrangement of who writes, who checks and what carries between attempts.
The rerun
All numbers are speedups over the FlashInfer baseline, on B300.
| Track | KDA 1.0 · 2026-05 | Human SOTA · 2026-05 | KDA 1.5 · 2026-07 | KDA 1.5 over human |
|---|---|---|---|---|
| MoE FP8 block scale | 0.67× | 1.80× | 2.25× | 1.25× |
| DSA sparse attention | 11.91× | 22.99× | 29.95× | 1.30× |
| GDN prefill | 1.16× | 4.42× | 6.10× | 1.39× |
The MoE row is the one worth sitting with. KDA 1.0 was slower than the baseline there — 0.67× — and the same application, with a profiling skill and a different loop, is now past the best human entry. Nothing about the model changed.
A workload nobody had tuned
Contest tracks have been optimized by many people, which makes them a fair test and a stale one. So we also pointed KDA 1.5 at Flash-KMeans, which nobody had worked over: balanced k-means for the Wan 2.2 workload on B200 came back 6.1× faster.