Agentic Space Factory

Turn a Hugging Face model card into a private, testable Gradio Space using an agentic HF Job.

Recommended workflow

1. Build from model card
   → creates a private Space
   → attempts ZeroGPU first
   → falls back to a fixed GPU if automatic hardware assignment is available
   → otherwise marks the run as manual_hardware_required

2. If hardware had to be changed manually
   → set the GPU in the generated Space Settings
   → run Validate existing Space
   → smoke-test generation
   → measure latency
   → store the output artifact in the Bucket

Each launch returns quick links to open the HF Job, generated Space, Space settings, and run artifacts in new tabs.

Honest guarantees

• Spaces are private by default.
• Nothing is published automatically.
• Runs, reports, generated files, traces, validation results, and artifacts are written to your private Bucket.
• Success is based on the deployed Space, not only generated code.
• ZeroGPU and fixed-GPU upgrades are best-effort through OAuth; manual hardware selection is an expected fallback.

Limits

This app attempts model-card builds; it does not guarantee that every model will run. Multi-GPU models, Docker-only apps, custom CUDA/FlashAttention stacks, gated models, very large models, or models with unclear documentation may produce technical_blocker, health_only, or manual_hardware_required instead of a full inference success.

Run Bucket: by default each signed-in user writes to their own private bucket: <username>/space-factory-runs. Use Check run bucket or Create private run bucket before launching Jobs.

Run storage

Runs are stored in a private Storage Bucket under the signed-in user's namespace. Create it once here, then use the same bucket name for Build and Validate.

Sign in, then check or create your private run bucket before launching Jobs.