Gaia

Gaia is yasp's kernel-generation and evaluation product. The KernelGen router proposes optimized GPU kernels from a PyTorch reference model; a Gaia eval worker benchmarks each candidate on real hardware and checks its numbers against the reference, so the kernel you get back is the one that measured fastest on the GPU you're targeting.

A Gaia worker is the evaluation runner that runs on the target GPU and scores every candidate; you drive export, compile, and inspection from your reference machine.

Guides

• Generate your first kernel — from a PyTorch reference model to an optimized, scored GPU kernel in four steps.

Workers

• Gaia workers — deploy the eval worker with Docker (or Helm for a cluster).

Reference

• kernelgen reference — export, compile, and get.
• eval reference — score candidate kernels and inspect evaluators.