| name | hf-spaces |
| description | Build and deploy ML demo apps on Hugging Face Spaces: Gradio, Docker, or Static SDKs, ZeroGPU vs paid hardware, and debugging a Space that won't build or run. Use to host a public demo or app for a model.
|
| stack | ml |
Hugging Face Spaces
A Space is a git repo hosting an app. Three SDKs:
| SDK | Use | Hardware |
|---|
| Gradio | default for ML demos, Python, ZeroGPU | compute |
| Docker | non-Python stack (Streamlit, etc.) | compute, no ZeroGPU |
| Static | plain HTML / Svelte / React, in-browser ML (transformers.js) | free, none |
Auth + hardware reality
hf auth whoami
- Static Spaces are free for everyone, no hardware.
- Gradio/Docker need a paid plan, except a free account in good standing can
host up to 2 ZeroGPU Spaces.
- ZeroGPU (
zero-a10g) is dynamic per-request GPU (RTX PRO 6000). Gradio +
PyTorch only. Free to create; visitors spend their own quota.
- Dedicated GPU (T4/L4/A10G/A100/H200) is hourly-billed,
canPay=True only.
Check: hf spaces hardware.
Workflow
- Search prior art first:
hf spaces search "<model or task>" --sdk gradio.
Read an existing app.py + requirements.txt for the working pattern.
- Decide SDK + hardware (Gradio + ZeroGPU for a public demo by default).
- Create the Space, then write
app.py, requirements.txt, and a README with
YAML frontmatter (title, sdk, app_file, pinned).
- Deploy by pushing to the Space repo (or
hf upload ... --type space).
- Debug:
hf spaces logs <ns>/<space> and hf spaces restart <ns>/<space>.
When to use
Hosting a demo, a public app, or an interactive report for a model.
Pitfalls
- ZeroGPU is Gradio + PyTorch only; a non-PyTorch main model needs dedicated
GPU (or a Static Space with in-browser inference).
- Free accounts: 2 ZeroGPU Spaces max. Beyond that needs PRO or a community
grant.
Adapted from huggingface/skills (Apache-2.0).
