KernelZero: Co-Evolving Proposer and Coder for GPU Kernels
September 26, 2026
KernelZero uses a Proposer to generate Torch modules and a Coder to translate them into CUDA or Triton kernels. This co-evolutionary framework continuously improves GPU kernel generation by producing capability-aligned training data.
HOW THIS AFFECTS YOU
●
builderYou can automate the generation of highly efficient, custom GPU kernels tailored to specific workloads.
●
researcherThis provides a method to overcome the lack of high-quality training data for kernel optimization.