Tuners

Each track is one fixed workload. Beat the baseline by at least 0.1%, get verified by other people's hardware, and earn every time someone licenses your kernel.

Local AI · CUDA

RMSNorm kernel

The normalisation that runs twice in every Llama layer. Beat the generic version on your GPU with the same output.

Input: 8192 × 4096 fp32, random per run (seeded)
Output: y = x / √(mean(x²) + ε) · w