Code block from tensors.md gets ValueError with numpy>=2.0

[leshabirukov]

docs/intermediate_representation/tensors.md contains this code block:

at line 157:

```{eval-rst}
.. exec_code::

    from onnxscript import ir
    import numpy as np

    array = np.array([0b1, 0b11], dtype=np.uint8)
    # The array is reinterpreted using the ml_dtypes package
    tensor = ir.Tensor(array, dtype=ir.DataType.FLOAT8E4M3FN)
    print(tensor)  # Tensor<FLOAT8E4M3FN,[2]>(array([0.00195312, 0.00585938], dtype='float8_e4m3fn'), name=None)
    print("tensor.numpy():", tensor.numpy())  # [0.00195312 0.00585938]

    # Compute
    times_100 = tensor.numpy() * 100
    print("times_100:", times_100)

    # Create a new tensor out of the new value; dtype must be specified
    new_tensor = ir.Tensor(times_100.view(np.uint8), dtype=ir.DataType.FLOAT8E4M3FN)
    # You can also directly create the tensor from the float8 array without specifying dtype
    # new_tensor = ir.Tensor(times_100)
    print("new_tensor:", new_tensor)  # Tensor<FLOAT8E4M3FN,[2]>(array([0.1875, 0.5625], dtype='float8_e4m3fn'), name=None)
    print("new_tensor == times_100", new_tensor.numpy() == times_100)  # array([ True,  True])
```

Itried to run it with numpy==2.0.2

and get

Traceback (most recent call last):
  File "D:\quick\2024\November\onnxscript-main\.my-stuff\tryit.py", line 24, in <module>
    print("new_tensor == times_100", new_tensor.numpy() == times_100)  # array([ True,  True])
ValueError: operands could not be broadcast together with shapes (8,) (2,) 
Tensor<FLOAT8E4M3FN,[2]>(array([0.00195312, 0.00585938], dtype='float8_e4m3fn'), name=None)
tensor.numpy(): [0.00195312 0.00585938]
times_100: [0.1953125 0.5859375]
new_tensor: Tensor<FLOAT8E4M3FN,[8]>(array([0, 0, 4, 1.75, 0, 0, 0.0546875, 1.875], dtype='float8_e4m3fn'), name=None)

With numpy-1.26.4 it works well.
Also, it works well if uncomment line 176

new_tensor = ir.Tensor(times_100)

Actually, I hit this while tried to build docs, with 'wrong' for requirements-dev.txt numpy version, but I suppose, if onnxscript itself not restricts numpy version, so must its docs, so error.pipfreeze.txt

[justinchuby]

Thanks! It appears the line

ir.Tensor(times_100.view(np.uint8), dtype=ir.DataType.FLOAT8E4M3FN)

Has an encoding problem and does not create a correct tensor. We need to fix that.

[justinchuby]

The line in documentation needs to be changed:

- times_100 = tensor.numpy() * 100
+ tensor.numpy() * np.array(100, dtype=ml_dtypes.float8_e4m3fn)

This is because in numpy 2 the result of tensor.numpy() * 100 is promoted to float32. The following line times_100.view(np.uint8) will then incorrectly view the 32-bit dtype as 8-bit.

[leshabirukov]

Great! Thanks for fast responce.
But for me it needs also

+ import ml_dtypes

- print("new_tensor == times_100", new_tensor.numpy() == times_100)  # array([ True,  True])
+ print("new_tensor == times_100", new_tensor == times_100)  # array([ True,  True])