BUG: array_api.vecdot is wrong

[zerothi]

Describe the issue:

Using the numpy.array_api.vecdot I get wrong results.

According to the specification the vecdot collapses the axis dimension to have N-1 dimensions and performs dot-products along the specified axis, thus removing that axis:
https://data-apis.org/array-api/latest/API_specification/linear_algebra_functions.html#vecdot-x1-x2-axis-1

However, by using tensordot it extends the axis for the non-contracted axes.

Reproduce the code example:

import numpy as np

# recreated implementation (see `numpy/array_api/linalg.py`)
def vecdot(x1, x2, axis=-1):
    return np.tensordot(x1, x2, axes=((axis,), (axis,)))

a = np.random.rand(9, 10, 11)
b = np.random.rand(9, 10, 11)
c = vecdot(a, b, axis=1)
d = np.einsum('ijk,ijk->ik', a, b)
assert d.shape == (9, 11)
# This will fail, the output shape will be (9, 11, 9, 11)
assert c.shape == (9, 11), c.shape

Error message:

Traceback (most recent call last):
  File "/home/nicpa/test.py", line 13, in <module>
    assert c.shape == (9, 11), c.shape
AssertionError: (9, 11, 9, 11)

NumPy/Python version information:

1.21.5
3.9.9

[AnirudhDagar]

@charris I think this can be closed. It was fixed in #21928.

[charris]

Thanks for the heads up @AnirudhDagar .