-2

I am following This tutorial to build tensorflow from source because my GPU only has 3.0 CUDA compatibility (If you know where to find a build that would work for me let me know as well!). I am using Ubuntu 17.10, and Bazel 0.13.0. CUDA 9.0, CuDnn 1.7.3, nvidia-driver version 396. I am running bazel build --config=opt --config=cuda //tensorflow/tools --cxxopt="-D_GLIBCXX_USE_CXX11_ABI=0"

It builds most of the way then gives me a huge error saying it didn't find various CUDA libraries.

Connor O'Hara
  • 139
  • 2
  • 10

1 Answers1

1

I recently experienced similar problems. It seems that recent Bazel version 0.13 is unable to find CUDA libraries set in LD_LIBRARY_PATH. I found two workarounds :

  1. Add cuda library paths to /etc/ld.conf.so (see comment at the end of https://github.com/tensorflow/tensorflow/issues/15889), or
  2. Use Bazel version 0.12 instead of 0.13

First one works, if you have only one cuda installation in your system. However I'm using different cuda versions for different projects, so I had to find an another solution. Latter solution seems to find libraries from LD_LIBRARY_PATH.

Arto Vuori
  • 11
  • 1