Code Review

This pull request updates make_tensor_proto in tensorflow/python/framework/tensor_util.py to normalize numpy arrays with non-native byte orders to the host's native byte order. It also adds comprehensive unit tests in tensor_util_test.py to verify this behavior for 32-bit integers, 32-bit floats, and 64-bit integers. No review comments were provided, and there is no additional feedback.

Hi @Ijtihed,

Thanks for submitting this fix! The logic to normalize non-native NumPy arrays before serialization looks correct and targets the root cause of the silent data corruption.

Could you add a unit test case to your PR to verify this behavior and prevent future regressions? Specifically, the test should cover various numeric data types (like int32, float32, and int64) initialized with a forced non-native endianness to assert they serialize and deserialize correctly.

For your convenience, here is a test implementation targeting tensorflow/python/framework/tensor_util_test.py that you can directly copy into your PR:

def testNonNativeByteOrder(self):
  for dtype in [np.int32, np.float32, np.int64]:
    arr = np.array([1, 2, 3], dtype=dtype)
    arr_swapped = arr.astype(arr.dtype.newbyteorder("S"))
    self.assertFalse(arr_swapped.dtype.isnative)

    t = tensor_util.make_tensor_proto(arr_swapped)
    a = tensor_util.MakeNdarray(t)
    self.assertTrue(a.dtype.isnative)
    self.assertEqual(arr.dtype.type, a.dtype.type)
    self.assertAllClose(arr, a)

Once you've added the tests, let us know and we'll proceed with reviewing and importing it internally!