# Set up model inputs

TensorViz needs a way to create your model and an example input for one forward call. It reuses a matching saved setup when one exists.

![Set up this project sidebar with Python and Model inputs steps beside a source-only preview](/images/project-setup-card.png)

## Complete the setup

The setup sidebar shows **Python → Model inputs**. Once Python is ready, the
next step uses the same agent/manual choices. **Learn more** explains the saved
configuration. If you close setup, choose **Complete setup** to return.

1. Open the model with **Open as Graph**.
2. Choose **Set up manually**, or **Set up project with agent** in the editor. The plugin browser offers **Copy setup request**.
3. Fill in missing constructor values and input shapes. Choose a Python factory when your model needs custom initialization.
4. Review the setup and save it.

A complete initial setup saves `tensorviz.toml` and starts the capture. Suggested values are examples; check that they make sense for your model.

If automatic input inference cannot resolve the forward inputs, fill them in the setup form. Agents can call `suggest_graph_inputs` or `configure_graph_setup`. The failure includes known unresolved input names and available worker details; manual inputs can still capture the model.

## Change an existing input

Open **Inputs** in the Canvas. Review the saved profile, change the values, and use the run controls to capture again. When editing `tensorviz.toml` directly, save it before choosing **Run again**.

## Use your own model builder

Choose **Model creation → Python function** when your project already has a builder or the model is part of a larger pipeline. The factory must return the selected model class.

## Declare an input in Python

Add a real Python comment such as `# tensorviz: input=2,8` next to the model class to declare one tensor. Separate shapes with semicolons for multiple arguments. Declarations belong to their model; conflicting declarations need manual setup. See the [declaration rules](/reference/graph-setup.html#class-scoped-input-declarations).

## Keep a source-only preview

Choose **Inspect source only** to explore without executing Python. Unmeasured shapes stay unknown.

[Python environments](/environment-recovery.html) · [Saved profile format](/model-validation-profiles.html)

<details>
<summary>Setup fields, factories, and supported cases</summary>

[Full setup reference](/reference/graph-setup.html) · [Supported model libraries](/graph-libraries.html)

</details>