On Friday, September 8, 2017 at 3:16:12 AM UTC+2, Fetchinson . wrote: > On 9/7/17, Thomas Jollans <tj...@tjol.eu> wrote: > > On 2017-09-06 16:14, Fetchinson . via Python-list wrote: > >> Hi folks, > >> > >> I'm trying to install a binary package (tensorflow) which contains > >> some binary C extensions. Now my system glibc is 2.15 but the binaries > >> in the C extensions were created (apparently) with glibc 2.17. So I > >> thought no problemo I installed glibc 2.17 to a custom location, built > >> python2.7 from source (hopefully using my custom glibc) and installed > >> pip and everything else using this custom built python. But still when > >> I try to import tensorflow I get: > >> > >> ImportError: /lib64/libc.so.6: version `GLIBC_2.17' not found > >> (required by > >> /home/nogradi/fetch/custom/lib/python2.7/site-packages/tensorflow/python/_pywrap_tensorflow_internal.so) > >> > >> > >> So apparently it's trying to use my system glibc, not the custom one. > >> > >> How do I tell this extension to use the custom glibc? Is it even > >> possible? > > > > It's going to use the same libc as python, so first of all check which > > libc your python interpreter is actually linked to. Maybe your custom > > build didn't do quite what you wanted. > > > > ldd `which python` # or something like that > > > > Once you've convinced yourself that python has the correct libc, you > > could try building tensorflow from source rather than installing the > > binaries. > > > > Maybe something in here helps: > > https://github.com/tensorflow/tensorflow/issues/53 > Thanks a lot for all the comments, my problem was indeed that the > compiled python was still using the system glibc. The solution was to > set the environment variables > > p=/path/to/custom/glibc > export CFLAGS=-I${p}/include > export LDFLAGS="-Wl,--rpath=${p}/lib > -Wl,--dynamic-linker=${p}/lib/ld-linux-x86-64.so.2" > > And then the compiled python was using the new glibc. > > Side note 1 on tensorflow: the compiled tensorflow binary uses unicode > ucs4, so for python I had to ./configure --enable-unicode=ucs4 because > the default is ucs2 > > Side note 2 on tensorflow: it also depends on libstdc++ and my version > was too old for that as well. Instead of compiling gcc from source > (which includes libstdc++) I copied a binary libstdc++ from a newer > linux distro and it was working fine. > > And the reason, if anyone cares, I had to go through the above is that > I couldn't compile tensorflow from source. > > Thanks again, > Daniel > > > >> > >> But maybe I have an even more basic issue: how do I link python not > >> with the system glibc but with my custom glibc? > >> > >> Cheers, > >> Daniel > >> > >> > >> > > > > > > -- > > Thomas Jollans > > -- > > https://mail.python.org/mailman/listinfo/python-list > > > > > -- > Psss, psss, put it down! - http://www.cafepress.com/putitdown
First of all, thank you a lot Fetchinson, it helped on my system (Centos 6) to today recompile the newest Python (3.10.7)! Of course I had foss2018 available on the system, so I "just" had to compile glibc, zlib, and openssl 1.1.1q ... but hey it was worth it! I have my environment up and running! FYI: 1. to me it helped that since I had nonstandard openssl location to edit Modules/Setup.dist as suggested here https://stackoverflow.com/questions/5937337/building-python-with-ssl-support-in-non-standard-location 2. to me it helped to configure with ./configure --with-libc=/tmp/glibc/lib/libc.so --prefix=/tmp/python --with-openssl=/tmp/openssl Hope this helps someone else who will try to do the same. Cheers -- https://mail.python.org/mailman/listinfo/python-list
