Current status

Apple is aware of this issue, see this comment below. Hopefully an OS update will have a fix soon. If you are experiencing this issue on a Mac with an M4 CPU it is not necessary to comment on this thread about it.

Describe the issue:

I guess its self explanatory. Following code:

np.identity(n=784) @ np.identity(n=784)

produces many wrong runtime warnings.

For small matrices is works as expected, for eg. np.identity(14) @ np.identity(14) doesn't produce warning but np.identity(15) @ np.identity(15) and bigger does.

Reproduce the code example:

import numpy as np
np.identity(n=15) @ np.identity(n=15) # produces warning
np.identity(n=14) @ np.identity(n=14) # no warning

# same result for
import numpy as np
np.ones(shape=(1, 2)) @ np.ones(shape=(2, 784))

Error message:

../test.py:3: RuntimeWarning: divide by zero encountered in matmul
  np.identity(n=784) @ np.identity(n=784)
../test.py:3: RuntimeWarning: overflow encountered in matmul
  np.identity(n=784) @ np.identity(n=784)
../test.py:3: RuntimeWarning: invalid value encountered in matmul
  np.identity(n=784) @ np.identity(n=784)

Python and NumPy Versions:

2.2.4
3.13.2 | packaged by Anaconda, Inc. | (main, Feb 6 2025, 12:55:35) [Clang 14.0.6 ]

I've checked also 3.11 & 3.9 but same result.

Runtime Environment:

No response

Context for the issue:

No response