Compute DFT from timeseries data wrong output shape

[cstiewegit]

behavior

I created a model to compute fft-coefficients of an onedimensional timeseries.
The output of the DFT-Operator is 3-dimensional, from these the (complex) coeffitients are gathered.
From them I want to compute the absolute value of these complex numbers.

The output in onnxscript eager mode works fine, but inferencing with onnxruntime throws an error:

Fail: [ONNXRuntimeError] : 1 : FAIL : Node (n11) Op (Gather) [ShapeInferenceError] axis must be in [-r, r-1]

After shape inferencing the model looks like this:

It looks like the output shape of the DFT-Operator is wrong?
Am I missing something?

code to reproduce

from onnxscript import FLOAT, script
from onnxscript import opset20 as op 
from onnxruntime import InferenceSession
import numpy as np
import onnx
    
@script(ir_version=9)
def onnx_dft_abs(X: FLOAT[10]) -> FLOAT[None]:
    X_as_column = op.Reshape(data=X, shape=[-1, 1])
    dft = op.DFT(input=X_as_column, onesided=1, axis=0, dft_length=10)
    # complex coefficients
    dft_complex = dft[:, 0, :]
    # compute absolute value
    squared = op.Mul(dft_complex, dft_complex)
    abs_val = op.Sqrt(op.Sum(squared[:, 0], squared[:, 1]))
    return abs_val

x = np.random.randn(10).astype(np.float32)

model = onnx_dft_abs.to_model_proto()
model = onnx.shape_inference.infer_shapes(model)
sess = InferenceSession(model.SerializeToString())
input_names = sess.get_inputs()[0].name
result = sess.run(None, {sess.get_inputs()[0].name: x})[0]