[NLPL Task Force (A)] Fwd: installation of CUDA 9.0 on Abel

[Stephan Oepen]

just fyi: i think i mentioned my custom glibc trick a few weeks ago.  i
have now re-built things a little more systematically and documented the
process; see below.  that gives us working versions of PyTorch (cpu and
gpu) and TensorFlow (cpu only, so far) on Abel.

martin, could you put me in contact with the CSC maintainer for TensorFlow
and PyTorch?  i would like to compare notes with them.  my impression is
that they prefer building from sources, and i would be curious to discuss
the benefits and pitfalls of either approach.

best wishes, oe


---------- Forwarded message ---------
From: Stephan Oepen <oe at ifi.uio.no>
Date: Wed, 19 Sep 2018 at 17:14
Subject: installation of CUDA 9.0 on Abel
To: <hpc-drift at usit.uio.no>

colleagues,

we have finally worked out a way of getting PyTorch and TensorFlow to
run natively (sans containers) on Abel; please see

http://wiki.nlpl.eu/index.php/Infrastructure/software/glibc
http://wiki.nlpl.eu/index.php/Infrastructure/software/tensorflow

for GPU support in TensorFlow, however, we would need CUDA 9.0 (not
8.0 nor 9.1, i am afraid).  the pre-compiled TensorFlow wheels
deliberately try to load specific minor version of shared libraries.

could we get CUDA 9.0 installed system-wide?

with thanks in advance, oe
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.nlpl.eu/archives/infrastructure/attachments/20180919/ada97a3c/attachment.htm>