> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/sdgm/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/sdgm/_mcp/server.

# User Guide

> Install SDM, select pretrained weights, prepare model inputs, and run local inference.

Use SDM to load pretrained structured data models and run inference in Python.
The steps below cover environment setup, model weights, and input preparation.

## 1. Set Up Your Environment

SDM requires Python 3.11 or later and PyTorch 2.7 or later. For GPU inference,
install a CUDA-enabled PyTorch build that matches your environment. Refer to the
[SDM installation guide](https://nvidia.github.io/structured-data-models/install.html)
and the [PyTorch installation instructions](https://pytorch.org/get-started/locally/).

Check your selected model's device requirements and any additional dependencies
needed by its examples. Installation and the first model download require
internet access.

## 2. Install SDM

```bash
pip install structured-data-models
```

## 3. Select and Load Pretrained Weights

Browse the [NVIDIA Kumo model collection](https://huggingface.co/collections/nvidia/nvidia-kumo-structured).
Use the model card to check supported tasks, input requirements, and weight
license. Follow the model's SDM reference for its constructor and checkpoint
loading options.

If the model requires authentication or acceptance of terms, complete those
steps on its Hugging Face page first. SDM uses the Hugging Face login configured
in your environment or the `HF_TOKEN` environment variable when accessing
private repositories. Refer to
[Hugging Face authentication](https://huggingface.co/docs/huggingface_hub/quick-start#authentication)
and [cache configuration](https://huggingface.co/docs/huggingface_hub/package_reference/environment_variables).

## 4. Prepare Inputs and Run Inference

Input structures, preprocessing, and output formats depend on the model and
prediction task. Use the [model reference](https://nvidia.github.io/structured-data-models/api/models.html)
and [repository examples](https://github.com/NVIDIA/structured-data-models/tree/main/examples)
for the model you selected.

The upstream guides cover
[table semantics](https://nvidia.github.io/structured-data-models/tensor.html),
[data processing](https://nvidia.github.io/structured-data-models/processing.html),
and [in-context learning](https://nvidia.github.io/structured-data-models/icl.html).
Match your column types, context data, and prediction inputs to the model's
supported schema before running inference.

## Kumo Tabular Example Notebook

Classify vehicle silhouettes from 18 numerical features using the OpenML
Vehicle dataset. The notebook runs the small model on CPU, fits labeled context,
and scores 85 held-out rows. It compares accuracy and macro one-vs-one ROC AUC
with a majority-class baseline.

The notebook installs its dependencies, downloads the dataset and pretrained
weights, and runs without a GPU.

Download the Kumo Tabular notebook

For import, download, device, and input errors, refer to
[Troubleshooting](/structured-data-models/troubleshooting).