🤗 Kernels: Major Updates

Summary

Hugging Face has introduced a dedicated “kernel” repository type on the Hub, allowing each custom kernel to declare supported accelerators, operating systems, and backend versions. This change makes kernels first-class artifacts that can be discovered, trended, and related to the models and applications that depend on them, while giving practitioners immediate visibility into hardware compatibility before installation.

Security has been strengthened through two new mechanisms. Trusted publishers restrict default loading to organizations vetted by the community; kernels from other sources require an explicit trust_remote_code flag. In addition, code signing with Sigstore’s cosign and ephemeral keys, combined with verification against approved GitHub workflows, protects against credential compromise on trusted repositories. Reproducibility remains anchored in Nix-based hermetic builds and embedded Git SHA provenance.

The CLI responsibilities have been split: kernels now focuses strictly on loading and preparation, while kernel-builder handles scaffolding and compilation. The resulting leaner interfaces support non-interactive, programmatically interpretable commands suited to agent-driven workflows. Expanded framework coverage, including Torch Stable ABI and Apache TVM FFI bindings, broadens the range of backends that can be targeted.

Agentic kernel development is supported by repeatable project layouts, backend-specific skills, and direct integration with HF Jobs for cross-hardware benchmarking. Agents can therefore scaffold, compile, measure speedups against baselines on multiple accelerator generations, and feed results back into subsequent optimization cycles. On the packaging side, kernels now link libstdc++ dynamically against the official manylinux_2_28 toolchain to avoid conflicts with PyTorch’s runtime while preserving broad compatibility.

These updates collectively address the full lifecycle of custom kernel packaging, distribution, and consumption, with explicit attention to hardware checks, reproducible builds, and performance validation required in production ML environments.

Why it matters

Provides actionable implementation guidance on kernel tooling, security, compatibility, and benchmarking that ML Engineers can apply to optimize models under latency and hardware constraints. Strong focus on MLOps practices and production deployment aligns with the category.