Built on top of gorodnitskiy/jax-cuda-docker 🫡

my 2 🪙's: jax-cuda-tf compatibility is a nightmare. instead of trying to reconfigure your servers, it is much easier to work with containerized environments like docker.

There are a lot of issues on GitHub about installing JAX with CUDA support, related to JAX and CUDA/cuDNN versions mismatching. This repository contains Dockerfile that can be used to easily run JAX with CUDA support in Docker, though specific modifications may be necessary in places.

For example, for eicl experiments, you need a very specific version of cudnn+jax combination. After many, MANY trials and error, use

pip install "jax[cuda11_cudnn82]==0.4.7" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html

inside the docker container (i.e. docker exec -it into it). While the Dockerfile automatically finds jax-cudnn combo, you may sometimes need to be über specific in case of the pesky lil updates that break everything. Otherwise, the script defaults to the most up-to-date jax/jaxlib available for said cuda/cudnn.

It strictly requires to specify, based on existing nvidia docker images on NVIDIA Docker hub:

• CUDA (eg: 11.4.3)
• OS (eg: ubuntu22.04 or centos7)

In case of JAX and CUDA/CUDNN versions mismatching, you have to change CUDA and JAX_CUDA_CUDNN building variables.

Check JAX versions via Google Storage. Check CUDA/cuDNN versions matching via cuDNN archive.

Each JAX for CUDA compiled with specific cuDNN versions. For example jaxlib==0.4.2 (CUDA=11) compiled for two cuDNN versions: 8.2 or 8.6. So, we might choose:

• CUDA="11.4.3" and JAX_CUDA_CUDNN="cuda11_cudnn82"
• CUDA="11.8.0" and JAX_CUDA_CUDNN="cuda11_cudnn86"

Also, it might be a problem with overall NVIDIA environment, for example incompatible NVIDIA driver version for requested CUDA version. It has to be checked apart.

An example is shown here. You can also specify pip requirements like in the example.

For example docker builds, take a look at this snippet.

See example here.