Potential dependency conflict with Python requirements.txt

[vepadulano]

Check duplicate issues.

• Checked for duplicates

Description

On MacOS we see failures in the CI due to:

>>> import onnx
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages/onnx/__init__.py", line 131, in <module>
    from onnx import (
  File "/Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages/onnx/compose.py", line 8, in <module>
    from onnx import (
  File "/Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages/onnx/helper.py", line 21, in <module>
    from onnx import _mapping, defs, subbyte
  File "/Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages/onnx/_mapping.py", line 104, in <module>
    np.dtype(ml_dtypes.float4_e2m1fn),
             ^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: module 'ml_dtypes' has no attribute 'float4_e2m1fn'. Did you mean: 'float8_e4m3fn'?

e.g. at https://github.com/root-project/root/actions/runs/19784857964/job/56689905732?pr=20570

The error is due to onnx/onnx#7249 . This was fixed in the latest onnx==1.19.1, but on MacOS nodes (macphsft33 specifically), we have

$: pip freeze
[...]
onnx==1.19.0

We can't simply update to onnx=1.19.1, because the transient update of ml_dtypes is conflicting with the current version of the tensorflow package.

The tensorflow package on the node is version tensorflow==2.16.2. I am not sure if we can "blindly" update all package versions, which would include updating tensorflow as well. According to our own requirements.txt, we have no strict dependency on this package version

root/requirements.txt

Line 16 in 1df6d55

tensorflow

, but I am not sure this is true.

Note that this means that currently MacOS will always fail because of the outdated onnx version

@couet @lmoneta please help!

@dpiparo @guitargeek FYI

Reproducer

No need

ROOT version

Any

Installation method

Any

Operating system

MacOS, but probably true for other platforms as well

Additional context

No response

[guitargeek]

TensorFlow tensorflow==2.16.2 is I think the newest version that supports macOS. Afterwards, they stopped caring about Mac support, and we are now at TensorFlow 2.20.

With the support of TensorFlow for macOS being this bad, we will always have problems. What about disabling all TensorFlow features on macOS?

Sorry for the lack of links (for now) to back up the claims. Typing from the phone.

edit: I was not correct with TF 2.16 was the latest supported version on macOS. It's only the the latest supported version on Intel macOS:

Quoting from:
https://www.tensorflow.org/install/pip

Caution: TensorFlow 2.16 was the last TensorFlow release that supported macOS x86

[dpiparo]

@lmoneta ?

[guitargeek]

I see that this is my fault now, because I have recently enabled again the testing of the TMVA SOFIE parsers, because they seemed to work just fine, and we needed test coverage for GSoC developments:

• [tmva][sofie] Disentangle PyMVA and SOFIE #20355

What I get now is that the tests should better be disabled in the case where onnx==1.19.0:

• [tmva][sofie] Require onnx!=1.19.0 and keras<3.5 for tests #20575