aten::conv3d onnx export bug when bias is missing

[bas-aarts]

This issue has not repro case, it just points out the issue.

aten::conv3d has the following code when bias is missing:

onnxscript/onnxscript/function_libs/torch_lib/ops/core.py

Lines 2128 to 2132 in c72e810

	if bias is None:
	weight_dim_0 = op.Shape(weight, start=0, end=1)
	bias_shape = op.Concat(weight_dim_0, op.Constant(value_ints=[2]), axis=0)
	zero = op.CastLike(0.0, input)
	bias = op.Expand(zero, bias_shape)

This code is copy-pasted for the aten::conv3d_complex code:

onnxscript/onnxscript/function_libs/torch_lib/ops/core.py

Lines 2173 to 2177 in c72e810

	if bias is None:
	weight_dim_0 = op.Shape(weight, start=0, end=1)
	bias_shape = op.Concat(weight_dim_0, op.Constant(value_ints=[2]), axis=0)
	zero = op.CastLike(0.0, input)
	bias = op.Expand(zero, bias_shape)

and it therefore generates a 2D bias, which is incorrect for non-complex conv3d. This will cause downstream issues.

the aten::conv1d and aten:::conv2d onnx export is correct, and that code should be used instead:

onnxscript/onnxscript/function_libs/torch_lib/ops/core.py

Lines 2038 to 2042 in c72e810

	if bias is None:
	weight_dim_0 = op.Shape(weight, start=0, end=1)
	bias_shape = op.Expand(weight_dim_0, op.Constant(value_ints=[1]))
	zero = op.CastLike(0.0, input)
	bias = op.Expand(zero, bias_shape)

[justinchuby]

Thanks for discovering the issue. Would you be interested in sending a PR?

[bas-aarts]

just checking why this simple fix is taking it's time to be merged into main.