numpy array metadata change

If you want to reinterpret the existing data in an array you need numpy.ndarray.view. That's the main difference between .astype and .view (i.e. the former converts to a new type with the values being preserved, while the latter maintains the same memory and changes how it's interpreted):

import numpy as np 

a = np.array([127,127,127,127,127,127,127,127], dtype=np.uint8)
b = a.view(np.uint64) 
print(a) 
print(b) 
print(b.view(np.uint8))                                        

This outputs

[127 127 127 127 127 127 127 127]
[9187201950435737471]
[127 127 127 127 127 127 127 127]

Note that 127 has a leading zero in its binary pattern, so it's not all ones, which is why the value we get in b is different from what you expect:

>>> bin(b[0])
'0b111111101111111011111110111111101111111011111110111111101111111'

>>> bin(72057594037927935)
'0b11111111111111111111111111111111111111111111111111111111'

What you seem to assume is a set of uint7 values of one bits...

Anyway, the best part about .view is that the exact same block of memory will be used unless you explicitly copy:

>>> b.base is a
True

The corollary, of course, is that mutating b will affect a:

>>> b += 3

>>> a
array([130, 127, 127, 127, 127, 127, 127, 127], dtype=uint8)

To control endianness you'd want to use string-valued dtype specifications, i.e. a.view('<u8') (little endian) or a.view('>u8') (big endian). We can use this to reproduce the faulty number in your question:

>>> a2 = np.array([0] + [255] * 7, dtype=np.uint8)
... a2.view('>u8')
array([72057594037927935], dtype=uint64)