AutoKernel: Autoresearch for GPU Kernels
Summary
AutoKernel is an open-source project that automates autoresearch for GPU kernel optimization using Triton. It profiles a PyTorch model to identify bottlenecks, extracts bottleneck kernels into standalone Triton kernels, and autonomously edits code, benchmarks, and either keeps or reverts changes, guided by Amdahl's law. It ships self-contained models like GPT-2, LLaMA, and BERT, along with nine kernel types, a fixed evaluation pipeline, and TSV-based results logging, making it a practical example of AI-driven optimization for real-world workloads.