We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Hugging Face Blog·Jul 6·5 min read
🤗 Kernels: Major Updates
TL;DR
Users faced challenges in identifying suitable compute environments for their projects. Hugging Face introduced a new repository type called 'kernels' to address these needs.
✦ Why It Matters
Engineers can now easily find and evaluate compute environments tailored to their specific project needs.
Key Takeaways
- Review the new kernel repository for compute-specific needs.
- Utilize trusted kernel publishers for enhanced security.
- Experiment with the Torch Stable ABI for framework compatibility.
- Leverage HF Jobs for efficient kernel performance benchmarking.
Full Summary
The Kernels project at Hugging Face has undergone a major redesign, introducing a new repository type called 'kernel' to better support compute-specific needs. This update enhances security by implementing trusted kernel publishers and code signing, ensuring that only verified kernels are loaded by default.
The project now supports multiple frameworks, including the Torch Stable ABI and Apache TVM FFI, allowing for greater interoperability among different AI frameworks. Additionally, the separation of concerns between the kernels and kernel-builder tools has streamlined the development process.
Users can now easily benchmark kernel performance across various hardware configurations, leveraging integration with HF Jobs for efficient testing. These updates not only improve the safety and usability of kernels but also lay the groundwork for future developments in agentic kernel creation.