GPU/TensorFlow install instructions incomplete - Githubissues

jeflem / ananke

Jupyter distribution with LMS integration via LTI

GNU Affero General Public License v3.0

5 stars 1 forks source link

GPU/TensorFlow install instructions incomplete #13

Closed jeflem closed 1 year ago

jeflem commented 1 year ago

Install instructions for GPU support and TensorFlow in Ananke doc yield GPU access from inside the Podman container, but TensorFlow fails as soon as computations shall be started on a GPU. The problem originates from missing CUDA installation inside the container (CUDA toolkit and cudnn do not suffice). Approximate steps for repair:

install libcuda1 via apt
append path to libcuda1.so to LD_LIBRARY_PATH env variable
run conda install -c nvidia cuda-nvcc in python3 env to get libdevice.so
set env variable XLA_FLAGS=--xla_gpu_cuda_data_dir=/opt/conda/pkgs/cuda-nvcc-12.2.140-0

Procedure should be tested and optimized, the added in detail to the docs.

jeflem commented 1 year ago

Solved on current dev branch and in upcoming release 0.3