Hugging Face Kernels: Swap Faster Kernels Into Models Without Rewriting
RisingSayak · x · 2026-09-18
Hugging Face's Sayak Paul demos the 🤗 Kernels library, which lets you swap optimized kernels into supported model layers without rewriting the model. A decorator (usekernelforwardfromhub) marks layers as swappable, a mapping binds layers to Hub kernels per device, and kernelize(model, mode=Mode.INFERENCE) replaces forward() on eligible layers only. You can specify separate kernels for inference vs training and per device type, registered globally or locally.
Related event: Hugging Face Launches Kernels Library for Drop-in Faster Kernels(2 posts)→
More from coding & agent
- PSAISuite routes PowerShell AI workflows through Vercel AI Gateway; switching models takes one string — dfinke · 2026-09-18
- Jev, a decisions-only model, matches gpt-4o-mini 50/50 while being 2.4x faster and 3.4x cheaper — Al_Grigor · 2026-09-18
- Laya: Open-Weights 421M System 1 Decision Model Outputs Calibrated Probabilities, No Text — victormustar · 2026-09-18
- AI agents can take your calls, but what if they called you to discuss decisions? — RichardsonDx · 2026-09-18
- OpenConfer: open-source voice infra lets AI agents call you to resolve decisions — RichardsonDx · 2026-09-18
- Hugging Face dev runs 5 Claude threads on top and Codex threads below, in parallel — NielsRogge · 2026-09-18