Tensorflow v2.0rc* is impossible to install with pip 18.1

[Ark-kun]

I'm installing TF on Debian 10 with python 3.6: Debian 4.19.37-5+deb10u1rodete2 (2019-08-06 > 2018) x86_64 GNU/Linux

TF 2.0beta1 had support for the manylinux1 tag which allowed me to successfully install it.
TF 2.0rc1 switched to the manylinux2010 tag which is not supported on my Debian installation with pip 18.1 and rc1 cannot be installed.

Can TF not require upgrading pip to install?

https://pypi.org/simple/tensorflow/ :
tensorflow-2.0.0b1-cp36-cp36m-manylinux1_x86_64.whl
tensorflow-2.0.0rc0-cp36-cp36m-manylinux2010_x86_64.whl

python3 -c "import wheel.pep425tags as w; print(w.get_supported())" | tr '\n' '\0' | sed -E 's/\),/),\n/g'
[('cp36', 'cp36m', 'linux_x86_64'),
 ('cp36', 'abi3', 'linux_x86_64'),
 ('cp36', 'none', 'linux_x86_64'),
 ('cp35', 'abi3', 'linux_x86_64'),
 ('cp34', 'abi3', 'linux_x86_64'),
 ('cp33', 'abi3', 'linux_x86_64'),
 ('cp32', 'abi3', 'linux_x86_64'),
 ('cp36', 'none', 'any'),
 ('cp3', 'none', 'any'),
 ('cp35', 'none', 'any'),
 ('cp34', 'none', 'any'),
 ('cp33', 'none', 'any'),
 ('cp32', 'none', 'any'),
 ('cp31', 'none', 'any'),
 ('cp30', 'none', 'any'),
 ('py3', 'none', 'linux_x86_64'),
 ('py36', 'none', 'any'),
 ('py3', 'none', 'any'),
 ('py35', 'none', 'any'),
 ('py34', 'none', 'any'),
 ('py33', 'none', 'any'),
 ('py32', 'none', 'any'),
 ('py31', 'none', 'any'),
 ('py30', 'none', 'any')]

[mihaimaruseac]

Unfortunately, no, we need to be manylinux2010 compatible if we want our pips to reside on PyPi.

Please update your pip via python -m pip install --upgrade pip or similar.

Closing this issue as there is a fix (upgrade pip) and there is no way to not use the manylinux2010 tag.

[Ark-kun]

we need to be manylinux2010 compatible if we want our pips to reside on PyPi.

Can you please elaborate a bit more on that or point me to the relevant document?

The official python packaging doc says: "Note PyPI currently supports uploads of platform wheels for Windows, macOS, and the multi-distro manylinux1 ABI. Details of the latter are defined in PEP 513." and does not mention "manylinux2010" anywhere.

The official python platform compatibility just mentions that "manylinux2010 is not yet widely recognised by install tools."

TF Model analysis does not publish manylinux2010 wheels and neither does numpy (https://pypi.org/simple/numpy/), so I'm not sure what the issue is.

[steffanjensen]

Read more about it here

#32627

All the Google employees in this repo will not take your issue serious, pls avoid talking to them and only talk to core TensorFlow devs.

Keep posting your issue, google mostly only support 64 bit in the cloud, this is the reason they don't have any interest in getting this fixed, but most of the community and the devs know this is an issue they will look into.

Python is 32 bit as default.

[mihaimaruseac]

Hi there @reliefs

This issue has no relationship with 32 bits support (as you can see from the tags, all are 64 bits). Please don't spam inadvertently.

[mihaimaruseac]

Pinging @gunan and @angersson for more details about manylinux2010 issue.

[mihaimaruseac]

Hi @Ark-kun

Our manylinux1 wheels were incompatible since a long time ago, see for example #8802 and #5033 (comment)

In the end, it all boils down to the standard C/C++ library used to provide symbols in precompiled code that is then used by Python. Maybe the other packages you mentioned above don't run into these issues.

Regarding the need to update tag to stay on PyPi, see pypi/warehouse#5420