PyTorch engineers described an LLM-guided autotuning method for Helion kernels that matches the performance of the existing Likelihood-Free Bayesian Optimization (LFBO) baseline while evaluating roughly 10 times fewer configurations in 6.7 times less wall-clock time. The approach feeds an LLM the kernel source, hardware specs, and configuration space across multiple rounds, using compiled benchmark feedback to iteratively refine candidate configurations. A hybrid strategy combining LLM seeding with LFBO refinement closes remaining performance gaps while still autotuning 3 times faster than a full LFBO search, and the team found the technique was largely model-independent, with Claude Opus 4.8, GPT-5.5, and Claude Sonnet 4.6 performing within a few percentage points of each other.
