[ONNX] Use the dynamo=True option from PyTorch 2.5

[justinchuby]

Feature request

PyTorch 2.5 introduces the torch.onnx.export(..., dynamo=True) option and an improved export logic for converting models to ONNX. We recommend optimum to leverage this feature.

Motivation

The torch.onnx.export(..., dynamo=True) logic leverages new model capturing mechanism offered by torch.export. It is how PyTorch will capture models moving forward. Additionally, we have better optimization and external data support built into this logic path.

This is a continuation of #1712

Your contribution

I can create a PR

[IlyasMoutawwakil]

Hi ! that's super cool, so we won't need to call torch.onnx.dynamo_export(..) ? and how much are these two different ?

[justinchuby]

Hi! The dynamo=True option in PyTorch 2.5 uses a new logic the leverages torch.export ExportedProgram and the ONNX IR for capturing graph and constructing ONNX models. It is more robust and produces more optimized graphs than the current dynamo_export. dynamo_export will be updated to use the same logic starting from PyTorch 2.6 (tentative).