conda install of 1.5.1 not detected by pip

[pluflou]

Describe the bug
There is a bug in the packaging of v1.5.1. Pip fails to detect it when installed with conda; it seems to be missing the dist-info directory.

To Reproduce
In a fresh conda environment, install the latest cxvpy:
conda install cvxpy -c conda-forge

Expected behavior
I expect to see cvxpy listed as installed when calling pip list or pip freeze.

Output
cvxpy shows up as UNKNOWN when calling pip list or pip freeze:

(test-pip) smiskov@PC99612 cvxpy % pip list
Package         Version
--------------- -----------
build           1.2.1
clarabel        0.7.1
ecos            2.0.13
numpy           1.26.4
osqp            0.6.3
packaging       24.0
pip             24.0
pyproject_hooks 1.1.0
qdldl           0.1.7.post2
scipy           1.13.0
scs             3.2.4.post1
setuptools      69.5.1
UNKNOWN         1.5.1
wheel           0.43.0

Version

• OS: macOS 14.4.1 and Ubuntu 22.04.4
• CVXPY Version: 1.5.1

Additional context
This was working as expected in v1.4.3. We saw this bug when building our conda package that depends on cvxpy. The build fails to pass the pip check test in our conda recipe (see workflow failure here and PR here).

[pluflou]

@phschiele this is in relation to the bug we saw when building spcqe a couple days ago

[phschiele]

@pluflou I am able to reproduce your output on a fresh code space:

Package    Version                                                                                                                                                                                       
---------- -----------                                                                                                                                                                                   
clarabel   0.7.1                                                                                                                                                                                         
ecos       2.0.13                                                                                                                                                                                        
numpy      1.26.4                                                                                                                                                                                        
osqp       0.6.3                                                                                                                                                                                         
pip        24.0                                                                                                                                                                                          
qdldl      0.1.7.post2                                                                                                                                                                                   
scipy      1.13.0                                                                                                                                                                                        
scs        3.2.4.post1                                                                                                                                                                                   
setuptools 69.2.0                                                                                                                                                                                        
UNKNOWN    1.5.1                                                                                                                                                                                         
wheel      0.43.0   

Installing from PyPI works, though.
@h-vetinari have you seen this behaviour before?

[h-vetinari]

have you seen this behaviour before?

No I haven't...

We didn't change anything about the way cvxpy gets installed between 1.4.3 and 1.5.1 (the stdlib things are unrelated), so at a guess this would be something to do with some potential behaviour changes in pip/setuptools?

I added a test in conda-forge/cvxpy-feedstock#92 and could reproduce the failure. On a hunch, I restricted to pip <24, but that didn't change anything. Note that we're deleting pyproject.toml before installation, because the way it is set up forces installation of

cvxpy/pyproject.toml

Lines 26 to 34 in f098629

	[build-system]
	requires = [
	"oldest-supported-numpy",
	"scipy >= 1.1.0",
	# 68.1.0 Promoted pyproject.toml's [tool.setuptools] out of beta.
	"setuptools>=68.1.0",
	"wheel",
	"pybind11"
	]

which cannot work in conda-forge for various reasons (these packages get installed through conda before the build starts).

However, we've been doing that since a long time, so I don't think this should play a role here.

[h-vetinari]

@timkpaine found the issue - it's that name="cvxpy" is missing from setup.py (since d8f74ec). In principle, this information is now present in pyproject.toml, but as I had mentioned, we removed that (mostly because it became painful to constantly patch out the build system).

[timkpaine]

It should be safe to include name in both since its obviously not going to change