Helion: A High-Level DSL for Performant and Portable ML Kernels (pytorch.org)

Helion is a new high-level, Python-embedded DSL that compiles into automatically tuned Triton kernels, offering “PyTorch with tiles”: familiar PyTorch host code plus a small device DSL (hl.tile, standard torch ops) that Helion turns into high-performance GPU kernels. The system separates host setup from device code, lets developers write kernels using PyTorch operators (mapped via TorchInductor), supports lambda templating for reusable epilogues, and dramatically reduces boilerplate (e.g., Attention in ~30 lines vs hundreds/thousands in Triton/CUDA). Crucially, Helion generates a rich, implicit autotuning search space from simple constructs (one hl.tile can spawn thousands of Triton configurations) and runs an ahead-of-time autotuner (Differential Evolution/Pattern Search) to find an optimal config for the target hardware and shapes. For the AI/ML community this means you no longer trade performance for productivity or accept heavy porting debt: a single Helion kernel can be automatically adapted across GPU architectures, memory-access methods (pointer/block/TMA), tile sizes, loop orders, PID mappings, register/thread staging, and cache-eviction hints. Autotuning typically takes minutes initially, after which the discovered config can be baked into the decorator for deterministic, production-ready performance equivalent to hand-tuned kernels. The net effect is faster kernel development, better performance portability, and lower maintenance cost—at the expense of an upfront autotuning pass that can be skipped once tuned.