Skip to content

Model Export

Exporting Ludwig Models

Ludwig supports exporting trained models in several formats for deployment.

Export Formats

SafeTensors (Default)

SafeTensors is the default format for Ludwig model weights. It provides secure, zero-copy serialization.

model = LudwigModel.load("results/experiment_run/model")
model.export_model("exported/", format="safetensors")

CLI: 

ludwig export_model -m results/experiment_run/model -o exported/ -f safetensors

torch.export

torch.export is the official replacement for TorchScript (deprecated in PyTorch 2.9). It captures the full computation graph as an ExportedProgram at the ATen operator level.

model = LudwigModel.load("results/experiment_run/model")
model.export_model("exported/", format="torch_export")

The exported .pt2 file can be:

  • Loaded back via torch.export.load()
  • Compiled with torch.compile() for runtime optimization
  • Used as input to the dynamo-based ONNX exporter
  • Deployed via ExecuTorch for on-device inference

CLI: 

ludwig export_model -m results/experiment_run/model -o exported/ -f torch_export

Note

export_model accepts an optional sample_input dict mapping input-feature names to tensors. When omitted (the default), Ludwig builds a dummy batch from the model's training-set metadata — so the snippet above works as-is. Pass a real example only if you need to trace against specific shapes or dtypes, for example when exporting models that rely on dynamic axes.

ONNX

ONNX export uses the dynamo-based exporter (torch.onnx.export(dynamo=True)) for cross-platform deployment.

model = LudwigModel.load("results/experiment_run/model")
model.export_model("exported/", format="onnx")

As with torch_export, sample_input is optional and is auto-generated from the training-set metadata when omitted.

The exported ONNX model can be loaded with ONNX Runtime:

import onnxruntime as ort
session = ort.InferenceSession("exported/model.onnx")

CLI: 

ludwig export_model -m results/experiment_run/model -o exported/ -f onnx

MLflow

Export to MLflow format for model registry and deployment:

ludwig export_mlflow -m results/experiment_run/model -o mlflow_model/

See Integrations for details on the MLflow integration.

Loading Exported Models

The load_exported_model utility auto-detects the format:

from ludwig.utils.model_export import load_exported_model

# torch.export format
program = load_exported_model("exported/model.pt2")

# ONNX format (requires onnxruntime)
session = load_exported_model("exported/model.onnx")