Deploying Models

Models run on the Xisom runtime in ONNX format. You upload a model, then choose it in a flow together with an input datasource and outputs. Enabling the flow loads the model.

Model Manager — upload an ONNX bundle and check its validation status.

Produce the ONNX first

If you are starting from a PyTorch or TensorFlow model — or want to quantize for faster edge inference — prepare it on a workstation with modelctl, then deploy the resulting bundle here.

Supported format

• ONNX — the runtime accepts ONNX models and executes them with the execution provider that matches your hardware. On the Jetson TX2 that is CUDA, with CPU as the fallback. See Hardware Setup for the modes.

Window size and features must match the datasource

A model’s input shape is [window size, feature count]: samples per prediction, and input channels. A flow offers only models whose window size equals the input datasource’s Window size and whose feature count equals its Tag count. The check is per axis, so a transposed model is refused. Note these values before you build the flow.

Upload and run

1. Upload the model

   Go to Models and upload your .onnx file. The platform validates the file (extension, type, and schema) before it is available.

2. Choose it in a flow

   Create a flow and pick this model for an input datasource. If the model is not in the list, its window size or feature count disagrees with the input. Adjust the datasource or regenerate the model to match.

3. Enable the flow and start

   Enable the flow, then press Start. Enabling loads the model into the inference service; Start begins producing predictions. See Your First Inference for the full loop.

Replacing the demo seed model

A fresh install ships a demo model with random weights — predictions are meaningless until you replace it. Upload your trained model as above, or drop it into the model-data directory and restart the inference service. The offline-bundle install page covers the file-drop path: Replace the demo model (step 5).

If something goes wrong

• Upload rejected, the model is missing from a flow’s list, or enabling the flow fails to load it — see the Troubleshooting runbook.

Next steps

Create a flow Bind the model to an input and outputs, then enable it.

Connect an input datasource Feed the model with live sensor data.

Monitor predictions Latency, throughput, and KPIs.

See the pipeline Visualize the Input → Model → Output topology and check live connections.