Lightning Talk: Streamlining Model Export with the New ONNX Exporter
summary
New exporter (beta in PyTorch 2.1) uses TorchDynamo to produce ONNX from dynamic PyTorch graphs via torch.onnx.dynamo_export and runs with ONNX Runtime. It preserves control flow and data types, writes module boundaries as ONNX functions, and supports exporting without loading weights using fake tensors, avoiding OOMs for very large models. ONNX Script lets teams author missing ops or runtime kernels. Most Dynamo-traceable models export successfully today; a few niche ops and graph breaks remain to be implemented.
AI-written summary · about
chapters (9)
- 0:00ONNX and model export
- 1:23Old exporter limitations
- 2:49New ONNX exporter in PyTorch 2.1
- 3:56Key improvements
- 4:37New API and features
- 6:43Debuggability improvements
- 7:26ONNX Script for custom ops
- 8:28Future plans
- 9:15Q&A
related talks
11mWhat's New in Torch.onnx
05mPyTorch ONNX Export Support
10mPyTorch 2.x: What’s coming up next for export-path
30mThe Evolution of ONNX: Driving Interoperability in AI...
33mWhat on earth is ONNX and ONNXRUNTIME
45mLeverage Power of Machine Learning with ONNX
45mMaking Machine Learning Applications Fast and Simple with...
51mMaking neural networks run in browser with ONNX