BUG: AttributeError: module 'numpy.linalg._umath_linalg' has no attribute '_ilp64'

[gunaxprofessional]

Describe the issue:

Whenever I try to run this code, from peft import PeftModel , it throws an error.

Reproduce the code example:

from peft import PeftModel

Error message:

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-39-c22a8a6046a0> in <cell line: 1>()
----> 1 from peft import PeftModel

18 frames
/usr/local/lib/python3.10/dist-packages/numpy/testing/_private/utils.py in <module>
     55 IS_PYSTON = hasattr(sys, "pyston_version_info")
     56 HAS_REFCOUNT = getattr(sys, 'getrefcount', None) is not None and not IS_PYSTON
---> 57 HAS_LAPACK64 = numpy.linalg._umath_linalg._ilp64
     58 
     59 _OLD_PROMOTION = lambda: np._get_promotion_state() == 'legacy'

AttributeError: module 'numpy.linalg._umath_linalg' has no attribute '_ilp64'

Runtime information:

numpy version = 1.23.5
system version = 3.10.12 (main, Jun 11 2023, 05:26:28) [GCC 11.4.0]

Context for the issue:

Well, I'm currently working on building a chatbot using the LLM. When I try to execute the from peft import PeftModel command, I receive the above error. You can access my notebook at the following link

[mattip]

Looking at your notebook and the stack trace, the interesting part is where the code does

from scipy.stats import norm

which then calls

from numpy.testing import suppress_warnings

This was removed in response to scipy/scipy#19344 in the latest SciPy HEAD.

As to why it fails: that seems very strange and points to a problem with the numpy module. What kernel/runtime are you using? What happens if you try to only import that attribute, which of these lines fail?

import numpy
import numpy.linalg
import numpy.linalg._umath_linalg
import numpy.linalg._umath_linalg._ilp64

[alexanderkulesza]

Can reproduce with numpy 1.23.5
import tellurium as te on a google collab

Trying the sequence you suggested gave

---------------------------------------------------------------------------
ModuleNotFoundError                       Traceback (most recent call last)
[<ipython-input-17-d2a9a824a7cd>](https://localhost:8080/#) in <cell line: 4>()
      2 import numpy.linalg
      3 import numpy.linalg._umath_linalg
----> 4 import numpy.linalg._umath_linalg._ilp64

ModuleNotFoundError: No module named 'numpy.linalg._umath_linalg._ilp64'; 'numpy.linalg._umath_linalg' is not a package

[mattip]

on a google collab

What kernel? How did you install numpy 1.23.5?

[alexanderkulesza]

To be honst, no idea. Collab comes with python3.10 and had that numpy version pre-installed. It also seems that I can not simply uninstall numpy1.23.5 and install a later, as then google-collab 1.0.0 does not run anymore

[mattip]

Maybe worth re-installing.

[contravarianceisuseful]

Also having this issue.

pip show says 1.26.2 but in the python kernel numpy.__version__ says 1.23.3.

Error happened when calling from ray.rllib.agents.ppo import PPOTrainer.

Forcing it to pip install version 1.24.1 fixed the issue.

[Mrw33554432]

Same, newest numpy version 1.26.2 and 1.24.3, 1.24.1 and more

[Mr-Twave]

In Spyder (anaconda3), I'm using print(np.__version__) -> 1.24.3 and it's throwing this same error.

Edit: Do I have two numpy versions installed? I'm very confused.
pip show numpy

Name: numpy
Version: 1.26.2

conda list numpy
# packages in environment at C:\Users\[user]\anaconda3:
#
# Name                    Version             
numpy                     1.26.2         
numpy-base                1.26.2         
numpydoc                  1.5.0           

!pip install --upgrade numpy
Requirement already satisfied: numpy in c:\users\[user]\anaconda3\lib\site-packages (1.26.2)

import numpy as np

print(np.__version__)
1.24.3

I am using the default spyder python installation path directory.

[mattip]

I think these all are errors due to two numpy versions installed on top of eachother. It just so happens ilp64 is the first thing to trip up the imports when numpy.linalg._umath_linalg is from a later version, and numpy/testing/_private/utils.py is from an earlier version.

[seberg]

Agreed with Matti, everyone here reporting seems to have a messed up NumPy installs that somehow have two versions mingled.

I might suggest manually uninstalling NumPy (possibly more than once with pip), but it's hard to guess how to fix broken environments remotely. In either case these don't seem like NumPy issues.

[Mrw33554432]

Agreed with Matti, everyone here reporting seems to have a messed up NumPy installs that somehow have two versions mingled.

I might suggest manually uninstalling NumPy (possibly more than once with pip), but it's hard to guess how to fix broken environments remotely. In either case these don't seem like NumPy issues.

Not really. During problem solving I actually tired uninstall - reinstall, and it doesn’t work. And inside the site package there is only 1 numpy, as well as in the package list. So it is an unsolved problem in my case, and the bug only appeared in that piece of code - if there are two numpy then bug will be everywhere, as it used to be.

Additionally, having 2 versions of numpy is a very common issue, perhaps it worth to monitored during installation (actually I have encountered this type of issue a couple of times).

[Mr-Twave]

Agreed with Matti, everyone here reporting seems to have a messed up NumPy installs that somehow have two versions mingled.

I might suggest manually uninstalling NumPy (possibly more than once with pip), but it's hard to guess how to fix broken environments remotely. In either case these don't seem like NumPy issues.

I attempted this and I appeared to achieve the same result with only numpy 1.24.3 installed. I would reconsider closing this issue.

[charris]

@Mr-Twave Have you done a file search to see if there is another version lurking about?

• What OS are you running on?
• How did you install 1.24.3?

[KlevertonSilva]

It worked here using pip install numpy==1.24.4

[finch-wren]

The issue still exists in colab, while pip show shows the latest version of numpy - the actual actual import's numpy.version shows a different version causing errors in codes as mentioned above. Reopen the issue please.

[mattip]

@finch-wren please open a new issue and describe exactly how to reproduce it so we can help. So far we have not seen anyone explain how they get to this situation in a way that is reproducible.

[ShivamGupta92]

Encountered same issue when working with Scikit-learn (version== 1.4.0) and NumPy (version==1.25.2). Had a version conflict with NumPy tried upgrading to NumPy version==1.26.3

[mattip]

I locked this issue. Please do open an new issue with a description of your environment and complete reproducer. Much of the discussion here is not anything NumPy can control, but maybe we can help you fix the problem with a good description of what you are doing.