Bridging the Hardware Gap With Code Harnesses on the Hugging Face Kernels Hub
summary
Benchmarks compare six harnesses across ten models on CUDA and Metal, measuring agent cost, kernel latency, VRAM use and end-to-end inference performance; one agent-generated kernel boosts Qwen3-8B throughput ~94% on H100 and a kernel for H200 is published. The platform provides a reproducible Nix build with build.toml, downloadable wheels and PyTorch integration via get_kernel and kernelize to swap ops, plus a distribution UI. Demonstrated tooling includes Kernels Hub for packaging and HF Skills for agent-driven kernel generation and benchmarking.
AI-written summary · about
chapters (12)
- 0:00Custom kernels with code agents
- 1:03Need for kernel standardization
- 3:10PyTorch kernel optimization
- 5:59Challenges in kernel building
- 7:24Kernels library features
- 8:27Kernel structure on Hugging Face Hub
- 9:10Using kernels in PyTorch
- 10:56Bridging hardware gaps
- 12:22Kernel generation agents
- 14:51Kernel optimization results
- 17:47Getting started with kernels
- 18:50Q&A
related talks
19mAI Kernel Generation: What's working, what's not, what's next
25mLeveraging AI Ecosystem for Portable and Sustainable GPU Kernels in HPC
18mKernelAgent: Hardware-Guided GPU Kernel Optimization Via Multi-Agen...
13mKernelBot: The World's First Competitive GPU Programming Platform
08mLightning Talk: Standardizing CPU Benchmarking with TorchBench for PyTorch...
21mUnlocking How To Efficiently, Flexibly, Manage and Schedule Seven AI Chips in Kubernetes
20mUnlocking Performance: Harnessing LLMs To Streamline GPU Kernel Development in...
01h 53mMastering Agentic Coding and GPUs