Huggingface Zerogpu

This teaches Claude how to write and debug Gradio apps that run on Hugging Face's ZeroGPU infrastructure, where you get fractional GPU slices on demand instead of a dedicated card. It covers the decorator syntax (@spaces.GPU), the quirky process isolation model (models load at module scope but inference runs in forked workers), duration tuning (default 60s blocks users even if your task takes 10s), and all the sharp edges like no torch.compile support and CUDA wheel-only dependencies. The bundled references explain why returning CUDA tensors hangs and how quota pre-checks work. If you're deploying ML demos on Spaces and hitting PicklingError or quota exceeded messages, this is the missing manual.

Install to Claude Code

npx -y skills add huggingface/skills --skill huggingface-zerogpu --agent claude-code