Helion Kernels

A collection of Helion kernels and their equivalent PyTorch models with example inputs to measure their performance.

Dependencies

• Helion
• [optional] Intel XPU Backend for Triton
  • XPU requires special nightly wheels of PyTorch and Triton

Kernels

The available kernels are based on and follow KernelBench categories:

• Level 1: Single-kernel operators - The foundational building blocks of neural nets
• Level 2: Simple fusion patterns - A fused kernel would be faster than separated kernels
• Level 3: Full model architectures - Optimize entire model architectures end-to-end