Helion: A High-level DSL for Kernel Authoring

Helion: A High-level DSL for Kernel Authoring - Jason Ansel, Meta The Meta PyTorch Team plans to announce a public beta for Helion at PyTorch conference, with the goal of attracting users. Helion is a new kernel authoring DSL tightly integrated with, and using many parts of, PyTorch 2. Helion is a Python-embedded domain-specific language (DSL) designed for authoring machine learning kernels. It compiles down to Triton, a high-performance backend for programming GPUs and other accelerators. By raising the abstraction level compared to Triton, Helion simplifies kernel development, minimizes boilerplate, and significantly enhances maintainability and portability. Key features of Helion include extensive automation in autotuning processes, implicit search space definitions, kernel templating via Python closures, and advanced integration with PyTorch 2, including support for tensor subclasses. This talk will introduce Helion, detail its design principles, demonstrate current capabilities, and provide insights into its implementation.