On Theta GPU, currently we support the major deep learning frameworks through two paths: singularity containers, based off of Nvidia's docker containers, and through bare-metal source builds. The bare-metal builds are so far only for tensorflow 2.X, with plans to support pytorch soon. Tensorflow 1.X is supported only via Nvidia's containers at this time.
As of now, the nvidia containers with tensorflow 1, 2 and pytorch built against cuda11, cudnn8 are available in singularity format here:
$ ls /lus/theta-fs0/projects/datascience/thetaGPU/containers/ pytorch_20.08-py3.sif tf1_20.08-py3.sif tf2_20.08-py3.sif
Execute a container interactively like this:
$ singularity exec --nv -B /lus:/lus /lus/theta-fs0/projects/datascience/thetaGPU/containers/tf1_20.08-py3.sif bash