Custom Kernels for All From Codex and Claude

Why It Matters

The skill dramatically reduces the time and expertise needed to create high‑performance GPU kernels, unlocking faster AI inference and easier distribution through the Kernel Hub.

Key Takeaways

• Skill packages GPU optimization knowledge for agents
• Claude and Codex generated working kernels for diffusers and transformers
• RMSNorm kernels achieved ~1.9× isolated speedup on H100
• End‑to‑end video pipeline improved 6% with custom kernels
• Published kernels load instantly via HuggingFace Kernel Hub

Pulse Analysis

Writing custom CUDA kernels has long been a bottleneck for AI engineers, requiring deep knowledge of GPU micro‑architectures, memory hierarchies, and library integration quirks. By encapsulating this expertise into a concise agent skill, Hugging Face enables large language models to act as autonomous kernel developers. The skill’s structured guidance—covering H100, A100, T4 specifics, PyTorch bindings, and template code—lets agents synthesize complete, buildable projects without manual low‑level debugging, effectively turning high‑performance GPU programming into a repeatable, automated workflow.

The performance data underscores the practical value of this approach. Isolated RMSNorm kernels achieved an average 1.9× speedup on H100, with larger sequence lengths seeing up to 2.5× gains. When integrated into the LTX‑Video diffusers pipeline, the custom kernels delivered a modest 6% end‑to‑end acceleration, which compounded to a 1.43× boost when paired with torch.compile. For the Qwen3‑8B transformer, the same kernel reduced RMSNorm latency by nearly half at long contexts, directly translating to lower inference costs for large‑scale language models. These results demonstrate that even single‑operator optimizations can have measurable impact on real‑world AI workloads.

Beyond raw speed, the skill dovetails with Hugging Face’s Kernel Hub, creating a seamless path from development to distribution. After an agent produces a kernel project, developers can publish pre‑compiled binaries that the Hub resolves automatically based on the user’s PyTorch, CUDA, and hardware versions. This eliminates the traditional compile‑time friction and democratizes access to optimized kernels across the community. As more models adopt custom operators, the combination of agent‑driven development and centralized hosting promises to accelerate innovation, lower engineering overhead, and broaden the reach of high‑performance AI applications.