Accelerating Autotuning in Helion with Bayesian Optimization

Helion, a PyTorch-like domain-specific language (DSL) for writing high-performance machine learning kernels, has addressed a major user pain point by deploying a new, machine learning-enhanced autotuning algorithm. The new method, called Likelihood-Free Bayesian Optimization (LFBO) Pattern Search, is now the default engine and significantly accelerates the previously slow process of finding optimal kernel configurations. This tackles the core issue where autotuning complex kernels could take hours, forcing developers to choose between long wait times or suboptimal performance. The algorithm intelligently navigates a vast, high-dimensional search space of implementation parameters like block sizes and loop orders.

The technical breakthrough lies in applying Bayesian Optimization principles to filter candidate configurations without requiring extra user data. On benchmark NVIDIA B200 kernels, LFBO reduces autotuning time by 36.5% while actually improving kernel latency by 2.6% on average. For specific kernels like LayerNorm, it achieves up to a 50% reduction in wall-clock time. This efficiency gain is critical because the search space for even a simple kernel can exceed 8 quadrillion configurations, most of which perform poorly. By using an ML model to predict and skip unpromising configs, Helion eliminates the inefficiency of exhaustively evaluating neighboring options, enabling faster development cycles for AI engineers building custom, hardware-optimized operations.