[NLPL Task Force (A)] [uninett.no #196965] Tensorflow issues, pt. 2

[oe@ifi.uio.no via RT]

hi thomas,

> I can load tensorflow without any issues.

[...]

> python3.7 is a personal conda env with cudatoolkit 10.0.130 and
> tensorflow 1.13.1.
> 
> So, to me it seems the installation requiring GLIBC_2.23 may be wrong.
> Comparing GLIBC strings in the shlibs in Henrik's and my installations
> shows some differences

could it be that the difference just lies in the origin of the 
pre-compiled TensorFlow binaries: PyPi (via pip) for henrik vs. 
(ana)conda in your case?

i believe we have seen in the past that the conda repository can be more 
accomodating when it comes to older OS environments.  but seeing as 
things like NCCL currently are not available through (ana)conda, i 
expect we (Sigma2, in collaboration with a user community like NLPL) 
need to continue to curate our own ecosystem of environment modules, and 
preferably seek to compile locally with the best applicable 
optimizations.

for example, the current NLPL modules for TensorFlow 1.15 (and 2.0) 
merely provide the standard PyPi binary wheel, lacking support for AVX2, 
AVX512F, and FMA.  in a similar spirit, i suspect it would (often) pay 
off to locally compile NumPy with the MKL back-end, which currently 
neither our NLPL modules nor the system-wide Python3 installations seem 
to provide?

oh well, modularization and version deluge ...

cheers, oe