MLSys 2026 FlashInfer kernel contest

Kernels designed with the [KDA](/applications/kda) workflow, by the MIT HAN Lab team, took top-three placements on tracks of the contest. The contest is the referee, on its own hardware and against its own reference implementation.

#1–#3Track placements at the contestResult

ncuProfiler output read by the agent as evidence, not decorationMethod

PublicThe evaluation and reproduction code, and the workflow itselfSource

Why a kernel contest is a good test

A kernel is either faster or it is not, and the machine that decides is not ours. It is also the failure mode we care about most: a kernel that is faster because it has quietly become incorrect will happily report a speedup, so correctness has to be a gate rather than a hope.

What is published is the placement, not a per-kernel speedup — so that is what we claim.

Check it yourself

The contest repository has the evaluation harness and the reproduction code, and the agent workflow that produced the kernels is public too.

mit-han-lab/mlsys2026-flashinfer-contest · mit-han-lab/kernel-design-agents · KDA