Kernels: Major Updates (huggingface.co)

The Kernels project has undergone a significant redesign, introducing a new repository type specifically for kernels that caters to users with unique compute requirements. These kernels can now be explored on the Hub, offering insights into supported accelerators, operating systems, and backend versions. This enhancement increases the discoverability of kernels, models, and their applications, ultimately bolstering the AI ecosystem. Alongside this, critical security measures have been implemented, including the introduction of "trusted publishers" to mitigate risks associated with malicious kernels. Users are now required to opt-in for loading kernels from non-trusted sources, and code signing protocols using ephemeral keys offer an additional layer of protection. The updates further expand support for frameworks, particularly incorporating the Torch Stable ABI and Apache TVM FFI, which allows developers to create kernels interoperable across various frameworks. A structured approach to agentic kernel development has been established, enhancing workflows that facilitate building, benchmarking, and optimizing kernels. Users will benefit from improved installation processes and system cards for each kernel, which document usage and interfaces. This comprehensive update not only streamlines kernel development but also ensures robust performance evaluation across different hardware, reinforcing the project’s commitment to open collaboration and community feedback.