>>>np.allclose([1,2,float('nan')], [1,2,float('nan')], equal_nan=True)
True
>>>np.allclose([1,2,float('nan')], [1,2,float('nan')], atol=[1,1], equal_nan=True)
True
>>>np.allclose([1,2,3], [1,2,3)], atol=[1,1,1], equal_nan=True)
True
>>>np.allclose([1,2,float('nan')], [1,2,float('nan')], atol=[1,1,1], equal_nan=True)
ValueError: operands could not be broadcast together with shapes (3,) (2,)

It appears that allclose is basically removing entries where a or b is NaN before passing them on to isclose with the same atol provided to allclose. This of course doesn't work if I expect there to be no NaNs some of the time and need to provide a longer atol because my a and b are different scales in different positions.

What should happen is allclose removes the same positions from atol before passing it on, so the dimensions always agree.