Conifer’s compute kernels are hand-tuned per architecture. If you can write a faster one, prove it: upload yours in the Kernel Lab, clear a bit-exact parity gate, and win the slot head-to-head on the bench. No cheating — real numbers, on your own machine.
Beat one of Conifer’s built-in compute kernels on its own ground. Pick a slot — attention, matmul, MoE routing — drop your kernel into the Kernel Lab, and it runs against the built-in under a bit-exact parity gate, then benchmarks both. You win when you’re correct and faster.
Point the Kernel Lab at your kernel source. It's referenced by path and compiled locally — the bytes never leave your machine.
Your kernel runs beside the built-in on real tensors. Its output must match within tolerance. Fail the gate and there's no score — correctness first.
Same model, same context, head-to-head tokens per second. Beat the built-in and the slot — and the bounty — is yours.