Helion 1.0: A High-Level DSL for Performance Portable Kernels
summary
Helion is a Python-embedded DSL that preserves PyTorch APIs, lowers through TorchInductor to multiple backends (Triton today, QTESL/Palace/Metal planned), and generates tiled device kernels plus host code automatically. The built-in autotuner explores ~1,200 Triton kernels in ≈10 minutes and a likelihood-free Bayesian variant cuts tuning to ~5–10 minutes and often 3–4x faster than prior search. Benchmarks show Helion matches Triton/QDSL performance, approaches CUDA, and outperforms Torch Compile and many bespoke Triton kernels across NVIDIA B200 and AMD MI350X.
AI-written summary · about
chapters (13)
- 0:00Helion 1.0 introduction
- 1:43What Helion is
- 4:29Why Helion exists
- 6:35Matrix multiply in Helion
- 8:39Helion auto-tuning
- 13:10Auto-tuner efficiency
- 15:58Production without auto-tuning
- 17:03Performance case studies
- 19:10Speedup comparisons
- 20:34PyTorch Foundation & governance
- 21:38Helion 2026 roadmap
- 22:41Helion compiler internals
- 24:06Getting started with Helion
related talks
28mHelion: A High-level DSL for Kernel Authoring
26mVLLM-Helion: SOTA LLM Performance by Advanced Autotuning and Fine-grained Dispatching
41mHelion's New CuTe DSL and Pallas Backends and How We Built Them Via Agents
08mA Halide Backend for TorchInductor
16mLightning Talk: Triton Compiler
25mLeveraging AI Ecosystem for Portable and Sustainable GPU Kernels in HPC
10mMaximizing Kernel Development Productivity Under Performance Constraints
25mRace to Superintelligence: Building Low-Friction AI Infrastructure for a Heterogen...