PLDI 2026

Writing Performance-Portable Kernels Simplified with Helion

A 4-hour hands-on tutorial at the PLDI 2026 AI Summit

Helion is PyTorch’s Python-based kernel DSL that compiles to optimized Triton code, delivering portable performance across NVIDIA, AMD, Google TPU, and Intel – with autotuning that explores thousands of configs per kernel. Attendees will write, autotune, and benchmark real Helion kernels live alongside the team. Open-source at github.com/pytorch/helion.

Agenda:

• Introduction to Helion (15 min)
• Autotuning in Helion (1 hr, hands-on)
• Compiler Architecture & TorchInductor integration (45 min)
• Distributed support – compute/comm fused kernels (45 min)
• Kernel benchmarking – incl. attention kernels (1 hr 15, hands-on)