Treatment of optional inputs to nodes when empty

[shubhambhokare1]

Describe the issue

In the pytorch-onnx exporter, when an optional input is not provided, it is defaulted to None, which gets translates to "" in the onnx graph. Semantically, "" and nothing should be the same in the ONNX spec, however this runs into issues for ops such as reduction ops and squeeze.

To reproduce

onnxruntime/onnxruntime/core/providers/cuda/reduction/reduction_ops.cc

Lines 648 to 652 in 493159b

	if (num_inputs == 2) {
	// override the attribute value with the input value for reduction_axes
	const Tensor* axes_tensor = ctx->Input<Tensor>(1);
	ORT_ENFORCE(axes_tensor != nullptr, "Axes input is null");
	ORT_ENFORCE(axes_tensor->Shape().NumDimensions() == 1, "An axes tensor must be a vector tensor.");

In this example for Reduction ops, let's say we have axes=None, this is now exported as axes="". So ORT still counts this as two inputs and the example fails as axes is Null. Ideally, this empty input, should not count as one of the inputs to the node.

Urgency

No response

Platform

Linux

OS Version

Mariner 2.0

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.18.0

ONNX Runtime API

Python

Architecture

X64

Execution Provider

Default CPU, CUDA

Execution Provider Library Version

CUDA 12.4

[skottmckay]

Yeah - that's a bug. It should be checking to handle the 'input with empty name' case.