Helion, PyTorch's DSL for ML kernels, traditionally relies on Likelihood-Free Bayesian Optimization (LFBO) for autotuning, which requires hundreds of compile-and-benchmark cycles. A new LLM-guided autotuner uses models like Claude Opus-4.8, GPT-5.5, and Sonnet-4.6 to propose kernel configurations via structured prompts, achieving LFBO-level kernel performance (geomean 1.009X) while exploring ~10X fewer configurations and reducing wall-clock tuning time by ~6.7X (39s vs 261s). A hybrid strategy — LLM seeding followed by LFBO refinement — closes remaining performance gaps in 6 of 8 edge cases while remaining ~3X faster than full LFBO. Notably, results are largely model-independent, with all three tested LLMs performing within a few percent of each other.
