KernelEvolve: How Meta’s Ranking Engineer Agent Optimizes AI Infrastructure

Meta's KernelEvolve is an agentic system that autonomously optimizes AI kernel code for heterogeneous hardware including NVIDIA GPUs, AMD GPUs, MTIA silicon, and CPUs.

• •Treats kernel optimization as a search problem by evaluating hundreds of candidate kernels and feeding diagnostics back to an LLM, achieving 60% inference throughput improvement for Andromeda Ads model on NVIDIA GPUs
• •Uses a retrieval-augmented knowledge base that injects platform-specific documentation and optimization patterns into the generation context, enabling support for new hardware without prior training
• •Leverages tree search to explore implementation alternatives for any operator including novel ones, with successful optimizations distilled into reusable patterns across model families
• •Automates evaluation of hundreds of kernel candidates in parallel, replacing weeks of manual expert hand-tuning with hours of automated search