> 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.

# When Should I Use Relational Deep Learning (RDL) or Kumo Relational?

> Choose between training a dedicated AutoML model and using Kumo Relational for immediate predictions

Use **Kumo Relational** when you want to make predictions from relational data without training a task-specific model first. Use **AutoML** when you want to train, evaluate, and operate a dedicated model with control over its model plan.

Both products work with relational data and use Predictive Query Language (PQL) to define prediction tasks. The main difference is whether the workflow uses a pre-trained foundation model at inference time or trains a model for your task.

## Quick decision guide

| If you need to...                                                               | Choose              | Why                                                                                                    |
| ------------------------------------------------------------------------------- | ------------------- | ------------------------------------------------------------------------------------------------------ |
| Test a prediction idea without waiting for model training                       | **Kumo Relational** | It uses a pre-trained relational foundation model and adapts through in-context examples.              |
| Make predictions for several exploratory or changing tasks                      | **Kumo Relational** | You can change the PQL task without training a separate model for each question.                       |
| Integrate immediate predictions into a Python application                       | **Kumo Relational** | The SDK sends the graph, task, and sampled relational context to a Kumo Relational NIM.                |
| Train and evaluate a dedicated model for a defined task                         | **RDL**             | It creates training, validation, and holdout splits and searches for an effective model configuration. |
| Control encoders, neighborhood sampling, architecture, or optimization settings | **RDL**             | The model plan exposes fine-grained training and architecture controls.                                |
| Inspect training experiments and compare holdout metrics                        | **RDL**             | The training workflow records experiment results and evaluates the selected model on holdout data.     |

## Choose Kumo Relational for immediate, flexible predictions

Kumo Relational is a good starting point when:

* You want a prediction before investing in a training workflow.
* You are still refining the target, prediction horizon, filters, or entity population.
* You have several useful prediction tasks that do not each need a separately trained model.
* Your application needs to submit a relational graph and PQL query through the NVIDIA SDFM SDK.

Kumo Relational keeps its pre-trained model weights fixed. It adapts to your task by using labeled historical examples and relevant relational context in the prediction request. See [How Kumo Relational Works](/rfm/introduction).

## Choose AutoML for a dedicated trained model

AutoML is a better fit when:

* You have a stable, well-defined prediction task.
* You want to train a model on your data and measure performance on a holdout set.
* You need to customize the model plan or optimize for a specific metric.
* You want a repeatable training and batch-prediction workflow managed through the platform or Python SDK.

AutoML generates candidate model configurations, runs training experiments, selects a model using validation results, and evaluates it on holdout data. Advanced users can override parts of the generated plan. See [Model Plan](/model-planner) and [Training](/training).

## Use both during the product lifecycle

You can start with Kumo Relational to validate that the data and PQL task produce useful predictions. If the task later requires a dedicated trained model, define the corresponding predictive query in AutoML and run its training and evaluation workflow.

This is a workflow choice rather than an automatic model conversion: Kumo Relational and AutoML use separate execution paths.

#### [Try Kumo Relational](/quick-start/rfm)

Build a graph and make a prediction without task-specific training.

#### [Try AutoML](/quick-start/fine-tune)

Connect data, define a predictive query, and train a model.