Hugging Face launches Kernel Hub for writing and distributing compute kernels
ariG23498 · x · 2026-10-19
Hugging Face has published documentation for its Kernel Hub, enabling Python libraries and apps to load compute kernels directly from the Hub.
Key points:
- Hub kernels differ from traditional Python kernel packages: they are portable (loadable outside PYTHONPATH), version-coexistent (multiple versions of the same kernel in one process), and broadly compatible across Python versions, CUDA builds, and C++ ABIs.
- The companion kernel-builder toolchain automatically builds kernels for all supported PyTorch configurations (CUDA versions, C++ standards).
- Developers can follow the kernel requirements spec to write kernels and distribute them globally.
More from coding & agent
- Multiplayer agent towns could enable fully autonomous agent teams — omooretweets · 2026-10-20
- Vercel Shop goes GA: Vercel and Shopify make AI agents first-class shoppers — evilrabbit_ · 2026-10-20
- Gemini team builds playable 3D Wizard Chess from scratch with multi-agent Blender workflow — GeminiApp · 2026-10-20
- Engineer: my real AI workflow is 10 rounds of dialogue with Claude, not agent factories — brandon_galang · 2026-10-20
- 84% of devs use or plan to use AI tools, but only 69% of agent users report productivity gains — YvesMulkers · 2026-10-20
- Devin spawned multiple thousand-star open source clones — only OpenHands survived — hargup13 · 2026-10-20