1Helion x 🤗 HF Kernels: Building and Shipping Out-of-the-box Performant Kernelspytorch.orgHigh09·11 17:45Z
Meta's Helion, a high-level DSL for writing portable ML kernels, is now supported within Hugging Face's Kernels project, enabling developers to package, autotune, and distribute kernels through the Hugging Face Hub1. A shipped attention kernel outperforms PyTorch SDPA on 19 of 19 pre-tuned shapes with a 1.20x geomean speedup on Nvidia H100s. Seven linear-attention kernels, pre-tuned for Nvidia B200s, beat flash-linear-attention across all variants with a 1.41x geomean device-time speedup on pre-tuned shapes and 1.35x on held-out shapes.