Supported Model Signatures

The following pages explores many kind of signatures for a forward method and how they translate into ONNX when they can. The results are summarized by the following pages. It tries model taking tensors, list of tensors, integers or floats. It also tries tests and loops.

• torch.export.export and static shapes

• torch.export.export and dynamic shapes

• conversion to onnx with static shapes

• conversion to onnx with dynamic shapes shapes

• Exported Programs with Static Shapes
• Exported Programs with Dynamic Shapes
• Exported ONNX with Static Shapes
• Exported ONNX with Dynamic Shapes

Tested Scenarios

• jit: tries to capture the graph by using torch.jit

• jit-decall: tries to capture the graph by using torch.jit then applies decompositions

• nostrict: call torch.export.export(..., strict=False)()

• nostrict-decall: call torch.export.export(..., strict=False)(), then applies decompositions

• strict: call torch.export.export(..., strict=True)()

• strict-decall: call torch.export.export(..., strict=True)(), then applies decompositions

• tracing: trace the execution of the model, it does flatten list, dictionaries or custom classes but the graph does not always produce aten function.

And for ONNX:

• custom-fallback: tries to export with different sceanrios then to convert into ONNX with the custom exporter

• custom-tracing: traces the model and then converts into ONNX with the custom exporter

• dynamo-ir: calls torch.onnx.export(..., dynamo=True)() and then optimizes the model

• script: calls torch.onnx.export(..., dynamo=False)()