Gradio turns a node graph into an API with gr.Workflow

Gradio added gr.Workflow, a way to describe a process as a graph of typed nodes. Push it to Hugging Face Spaces and it builds both a UI and a REST endpoint for every output node, per Hugging Face's guide. That collapses a step teams usually do by hand. They build an interactive demo, then write a separate API layer so other services can call the same logic.

You can chain nodes in sequence or fan one input out to run several in parallel. Nodes can be Hugging Face Inference Providers, existing Spaces, or plain Python functions, mixed in the same graph.

The minimal setup is gr.Workflow(bind=[your_function]).launch(). GPU nodes use the @spaces.GPU decorator with ZeroGPU, so a node can grab GPU time without managing the allocation yourself.

The guide doesn't say what happens when one node in a parallel fan-out times out, or how an error in one branch affects the others. For a single function or a two-node chain that gap doesn't matter. For a dozen nodes mixing GPU calls and external Providers, it's the first thing to test before trusting it in production.