> 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.

# Troubleshooting

> Resolve common environment, model download, and input issues when running inference with SDM.

These checks apply to local inference and the example notebook in the
[User Guide](/structured-data-models/user-guide).

## Environment and Dependencies

If an import fails, confirm that SDM and the example dependencies are installed
in the Python environment running your script or notebook. The notebook installs
SDM, `pandas`, `scikit-learn`, and `openml` in its first code cell.

The notebook explicitly selects CPU and does not require a GPU. For a CUDA
configuration, check that `torch.cuda.is_available()` returns `True`. If it
returns `False` on an NVIDIA GPU host, check the driver and CUDA-enabled
PyTorch installation.
See the [SDM installation guide](https://nvidia.github.io/structured-data-models/install.html)
and [PyTorch installation instructions](https://pytorch.org/get-started/locally/).

## Model Download and Access

If weights fail to download, check connectivity to Hugging Face and the error
message. Open the model page to confirm access. If authentication or acceptance
of terms is required, complete those steps and configure your Hugging Face login
in the environment running SDM.

The first model initialization may download pretrained weights. Subsequent runs
can reuse the cache. If storage or cache configuration causes an error, see the
[Hugging Face cache configuration](https://huggingface.co/docs/huggingface_hub/package_reference/environment_variables).

## GPU Memory

If inference raises a CUDA out-of-memory error, check the selected model size
and the available GPU memory. For Kumo Tabular, review the `size` parameter in
the [model reference](https://nvidia.github.io/structured-data-models/api/generated/sdm.models.KumoTabular.html)
to select a smaller model. The notebook sets `size="small"` and runs on CPU,
so it does not use GPU memory.

## Input Columns and Types

Check that the target has the intended semantic type, that query features
exclude the target, and that the selected model supports your input types and
prediction task. See
[table semantics](https://nvidia.github.io/structured-data-models/tensor.html)
and the [model reference](https://nvidia.github.io/structured-data-models/api/models.html).

## Report a Library Issue

[Search or open an issue in the SDM repository](https://github.com/NVIDIA/structured-data-models/issues).
Include your SDM and PyTorch versions, device details, the error message, and a
minimal reproduction. Remove credentials and private data before sharing it.