Custom Kernels for All from Codex and Claude

2026-02-16 · Source: Hugging Face - Blog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, long

Summary

Hugging Face has developed an agent skill that enables coding agents like Claude and Codex to autonomously write production-grade CUDA kernels. Published on February 13, 2026, this skill addresses the complexity of integrating custom kernels with `transformers` and `diffusers` libraries, which typically involves intricate hardware-specific optimizations and PyTorch bindings. The agents successfully generated working RMSNorm kernels for the LTX-Video `diffusers` pipeline and the Qwen3-8B `transformers` model, complete with PyTorch bindings and benchmark scripts. Benchmarking on an H100 80GB GPU showed isolated RMSNorm kernel speedups averaging 1.88x for LTX-Video and 1.94x for Qwen3-8B, with an end-to-end video generation speedup of 1.06x for LTX-Video. The skill also facilitates publishing these kernels to the Hugging Face Kernel Hub for easy, compilation-free distribution.

Key takeaway

For NLP Engineers and Computer Vision Engineers seeking to optimize model performance, this agent skill offers a streamlined path to custom CUDA kernel development. You can leverage coding agents to generate, benchmark, and integrate highly optimized kernels for `transformers` and `diffusers` models, potentially achieving significant speedups. This approach reduces the manual effort and specialized knowledge traditionally required, allowing you to focus on higher-level model development while offloading low-level optimization to AI.

Key insights

Coding agents can autonomously generate optimized CUDA kernels and integrate them into complex ML frameworks.

Principles

• Agent skills package domain expertise for complex tasks.
• Hardware-specific optimizations are critical for kernel performance.

Method

Install the `cuda-kernels` skill into a coding agent, then prompt it to generate and benchmark kernels for specific models or pipelines, leveraging its structured guidance and reference scripts.

In practice

• Use `kernels skills add cuda-kernels` to install the skill.
• Prompt agents for H100-optimized RMSNorm kernels.
• Publish generated kernels to the Hugging Face Kernel Hub.

Topics

• CUDA Kernels
• AI Agents
• Hugging Face Diffusers
• Hugging Face Transformers
• Performance Optimization

Code references

• huggingface/blog
• huggingface/kernels
• huggingface/kernels
• burtenshaw/kernel-skill

Best for: NLP Engineer, Computer Vision Engineer, AI Engineer, Machine Learning Engineer, Deep Learning Engineer

Related on AIssential

• Your Coding Agent Should Do AI System Engineering — Ben Burtenshaw, Hugging Face — Coding agents can perform complex AI systems engineering tasks, from kernel optimization to automated research.
• CuBridge: An LLM-Based Framework for Understanding and Reconstructing High-Performance Attention Kernels — CuBridge adapts expert CUDA attention kernels using LLMs and a structured lift-transfer-lower workflow for high performance.
• CUDA Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generation — Agentic reinforcement learning can significantly enhance LLMs' intrinsic CUDA kernel optimization capabilities.
• Boosting MoE Training Throughput with Advanced Fusion Kernels — Fusing MoE block operations with custom kernels significantly boosts training throughput by eliminating bottlenecks.
• Stanford CS 25 Transformers Course (OPEN TO ALL | Starts Tomorrow) — Transformers are highly expressive, optimizable, and efficient general-purpose computing models for diverse AI tasks.
• Automating GPU Kernel Translation with AI Agents: cuTile Python to cuTile.jl — AI-assisted workflows can translate GPU kernels between DSLs by encoding domain-specific rules and pitfalls.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Hugging Face - Blog.