Exporter Recipes Gallery

A model can be converted to ONNX if torch.export.export() is able to convert that model into a graph. This is not always possible but usually possible with some code changes. These changes may not be desired as they may hurt the performance and make the code more complex than it should be. The conversion is a necessary step to be able to use ONNX. Next examples shows some recurrent code patterns and ways to rewrite them so that the exporter works.

See torch.onnx.export: export to ONNX for an organized version of this gallery.

A couple of examples to illustrate different implementation of dot product (see also sphinx-gallery).

Getting started

pytorch nightly build should be installed, see Start Locally.

git clone https://github.com/sdpython/experimental-experiment.git
pip install onnxruntime-gpu nvidia-ml-py
pip install -r requirements-dev.txt
export PYTHONPATH=$PYTHONPATH:<this folder>

Common Errors

Some of them are exposed in the examples. Others may be found at Frequent Exceptions or Errors with the Exporter.

Check the exporter on a dummy from HuggingFace

Check the exporter on a dummy from HuggingFace

Export Phi-3.5-mini-instruct piece by piece

Export Phi-3.5-mini-instruct piece by piece

Export Phi-3.5-mini-instruct with draft_export

Export Phi-3.5-mini-instruct with draft_export

Export Phi-3.5-mini-instruct with report_exportability

Export Phi-3.5-mini-instruct with report_exportability

Export a model with a loop (scan)

Export a model with a loop (scan)

to_onnx and a custom operator inplace

to_onnx and a custom operator inplace

to_onnx and a custom operator registered with a function

to_onnx and a custom operator registered with a function

to_onnx and a model with a test

to_onnx and a model with a test

to_onnx and padding one dimension to a mulitple of a constant

to_onnx and padding one dimension to a mulitple of a constant

to_onnx and submodules from LLMs

to_onnx and submodules from LLMs