Kernels designed with KDA took a podium place on every track of the MLSys 2026 FlashInfer kernel contest, in the Full-Agent division, where no human writes kernel code: first on MoE, second on DSA, third on GDN. The team entered as HAN Lab Kernel Mafia.
Swipe the diagram sideways to see all of it →
1/0 Three tracks, each scored on the organisers’ machines.
A kernel contest is the closest thing to an ideal test for an agent loop. The score is a measurement, the machine belongs to the organisers, the reference implementation is theirs, and correctness is a gate rather than a matter of interpretation. A kernel that is faster because it has quietly become incorrect does not place. It fails.
The loop that wrote them
The agent is the Humanize coder–verifier loop: one agent plans and writes a kernel change, and an independent verifier checks it for correctness and speed before anything is kept. Two skills sit beside it. KernelWiki is a searchable store of production kernels, past submissions and documentation, and ncu-report-skill turns an Nsight Compute report into evidence an agent can act on.
Swipe the diagram sideways to see all of it →
1/0 A coder agent plans and writes the next kernel change.
The speedups, as submitted
A placement ranks teams against each other; it does not say how fast the kernel is. The team's technical report gives the speed of each submitted kernel against the official FlashInfer baseline, as the arithmetic mean over the official workloads.
- KDA, as submitted
Two of the five kernels were slower than the baseline. One of them, MoE FP8 at 0.65×, still won its track: the baseline there calls a closed-source TensorRT-LLM GEMM, and no Full-Agent entry placed ahead of it.An earlier version of this post gave 1.4× on MoE, 33.3× on DSA and 17.6× on GDN. Those figures do not match the technical report, so they are gone.
After the contest: what each piece is worth
After the deadline, the team reran KDA and K-Search, another agentic kernel system, on the same workloads under the same strict 48-hour budget. KDA came out ahead on all three tracks. Switch the baseline to K-Search to read the margin directly.
On the DSA top-k indexer, the report adds the pieces one at a time. The loop matters most, and each skill adds to it.
- Speedup
DSA top-k indexer, full loop
8.58×Humanize, KernelWiki and the profiler skill together, from 1.37× without them.
Best submitted kernel
19.08×The DSA indexer as entered in the contest, against the FlashInfer baseline.
The kernels, the prompts and the report are published.
The contest kernels · prompts and report · the results · KDA