MKEvolve Framework Generates Efficient Kernel Code for Accelerators

Despite advancements in large language model (LLM) based code generation, creating accurate and high-performing kernels for hardware accelerators remains a significant challenge in scaling modern machine learning workloads. This bottleneck often limits the efficiency of AI systems. Researchers have introduced MKEvolve, a novel modular multi-agent framework designed to address this issue. MKEvolve operates by iteratively co-evolving two key components: a modular decomposition of complex PyTorch modules and the LLM-generated kernel code for each resulting submodule. The framework continuously refines these decompositions through splitting and fusing operations across iterations. Concurrently, MKEvolve independently enhances each subkernel using an LLM-driven beam search. The outcome is a set of kernels that are programmatic compositions of independently verified subkernels. This modularity makes the kernels highly configurable, allowing for swappable implementations, and interpretable, as errors and speedups can be traced to specific subkernels. Furthermore, they are readily adaptable to related model architectures. Experiments using Triton on KernelBench L2 and L3 demonstrate that MKEvolve improves both correctness and speedup over direct end-to-end synthesis baselines, while also reducing LLM token usage by up to 35%.