KernelEvolve: Scaling Agentic Kernel Coding for Heterogeneous AI Accelerators (arxiv.org)

Meta has introduced KernelEvolve, a groundbreaking agentic kernel coding framework designed to optimize deep learning recommendation model (DLRM) training and inference across diverse hardware architectures. Addressing challenges in model architecture, kernel primitives, and hardware heterogeneity, KernelEvolve automates kernel generation and optimization by operating at multiple programming abstractions—from high-level languages like Triton to low-level hardware-agnostic options. This innovative system employs a graph-based search process to dynamically adapt to runtime contexts, significantly reducing development time from weeks to mere hours while maintaining 100% correctness across 250 benchmark problems. The significance of KernelEvolve in the AI/ML community lies in its capacity to enhance performance efficiency and decrease the programmability barrier for developers of new AI hardware. By validating its effectiveness on a variety of NVIDIA and AMD GPUs, as well as Meta’s proprietary AI accelerators, KernelEvolve represents a substantial leap forward in executing AI workloads at scale. The framework not only improves performance over existing PyTorch implementations but also streamlines the deployment process for heterogeneous AI systems, positioning itself as a vital tool for future advancements in AI infrastructure.