To build a conda tensorflow package with GPU support

• Build the required Docker containers:

  Dockerfiles can be found at: https://github.com/jjhelmus/docker-images

  CUDA 9.0: pkg_build_cos6_cuda90 CUDA 9.2: pkg_build_cos6_cuda92 CUDA 10.0: pkg_build_cos6_cuda100

• Start the docker container using:

  sudo nvidia-docker run -v `pwd`:/io -it pkg_build_cos6_cuda100
  

• Create a symlink for libcuda.so.1

  ln -s /usr/local/cuda/lib64/stubs/libcuda.so /usr/local/cuda/lib64/stubs/libcuda.so.1
  

• Update conda and conda-build, and navigate to the recipe root folder.

  Modify conda_build_config.yaml in this directory to specifiy the CUDA, CuDNN, compiler and python versions.

  To start a build use:

  conda build .
  

  To time the build and log the build output use:

  time conda build . 2>&1 | tee ../tf_build_gpu.log