Hugging Face Kernels primer: load optimized CUDA kernels from the Hub in one line
ariG23498 · x · 2026-11-03
Hugging Face's ariG23498 shares a primer on the Kernels library, with docs, a blog post, and an interactive notebook.
- Download community-optimized kernels straight from the Hub: getkernel("kernels-community/activation", version=1) then call fast implementations like gelufast
- Kernels are versioned by major version; version=1 pulls the latest v1 build
- Ships with cache management: kernels cache ls lists cached kernels, sizes, and build variants
- Handy for devs who want optimized kernels without writing CUDA themselves
More from Infra
- Your $99 Apple developer account hides 1PB CloudKit, 25hrs CI/mo and an on-device LLM — vincent_koc · 2026-11-03
- Marvell inks record-breaking $115B AI deal — LinkedInNews · 2026-11-03
- Hash partitioning cuts Neo4j parallel import from 250.6s to 186.5s on 3.2M rows — JeremyCMorgan · 2026-11-03
- ai& x AMD host one-day Tokyo hackathon with ¥10.2M in credits, 750K JPY tokens per winner — DavidBennett__ · 2026-11-03
- Musk mocks compute-satellite heat dissipation skeptics: 'They'll never get a job at SpaceX' — elonmusk · 2026-11-03
- tinygrad's exabox drops to $5M: 256x MI300X, 669 PFLOPS, delivery Q3 2027 — andrew_n_carr · 2026-11-03