PyTorch Universal Docker Template Versions Save

Removed brew because of installation failures. Changed the color of the pure current directory to cyan as the default dark blue was difficult to read.

Added Brew to images. Remove hub service due to its role being similar to simple. Renamed CCA to CCC.

Update the ruff version. This is just to create a new release with Zenodo enabled.

Release for Zenodo publication. Update PyTorch compilation to use modern CMake and the nvcc bugfix. Update documentation to fix typos and be more descriptive.

The non-interactive images now have basic support for shared .zshrc configurations. However, they no longer have zsh as a requirement, allowing users to change shells, etc. This will likely be useful if the non-interactive images are used interactively post-download, such as in interactive Kubernetes or as a reproducibility image uploaded to Docker Hub, etc.

Clean up docker-compose.yaml for better readability, remove HOST_NAME variable. Lots of bugfixes and usability enhancements as well.

Aside from many updates and bug fixes, the biggest change is that the Makefile now reads the .env file and no longer takes any variables from the host shell unless given during the make command invocation. Also, lots of new rules for ruff that I found reasonable.

The simple service can now install an environment from conda-lock files.

This is a very large release that adds many new services. Until now, Cresset has mainly focused on customizability at the expense of very complicated configurations. Also, it has had no support for Kubernetes. Now, the focus has changed to providing a simplified user interface as opposed to making the template customizable. Also, the old deploy stage was removed to reduce clutter in the main Dockerfile. While the Dockerfile is still recommended for maximum customizability, users who do not need any custom compilation and just wish for an easy-to-set-up development and training environment can use the other services. Also, the new build-only Makefile command and INTERACTIVE_MODE=exclude option allow users to create images for distribution more easily. Finally, the simple service has been set up to allow the use of conda-lock which is the strictest form of reproducibility in conda. This will help those with the strictest requirements needs. Moreover, because it consists only of official or verified images, the simple service will also help allay security concerns.

Backwards compatibility breaking changes to support PyTorch 2.0. A known issue is that the pip installation does not work on the deploy stage for Ubuntu 22.04. This will be fixed ASAP.