Arch Linux

As far as I can tell, pytorch actually uses all of those dependencies. Arch is not a minimal distro, despite what people say. We're a practical distro and enabling some common deps is quite reasonable to do. I'm not even sure what technical or practical reason we'd have to build it against fewer deps. I'm not sure how this makes running our pytorch in docker less worthwhile but would like you to explain that part.

Let's discuss.

I have been using pytorch for a couple of years and never encountered a need for QT or FFMPEG. When running with CUDA you typically don't need any of the competing Intel libs. OpenCV is debatable, the mainline PIP packages manage to do without, and you can install OpenCV for python separately if you need it. The problem with these packages is that they always end up pulling in a full desktop setup, which you don't want or need for a headless machine learning server or cloud image. Using Ubuntu and PIP for the pytorch images allows me to achieve this (but Ubuntu is painful to use in so many other ways, and the PIP images do not always work on Arch due to them lagging a bit with CUDA versions.)

Wrt using in Docker, e.g., for pushing images to the cloud, you will want as few dependencies as possible, to shorten your docker image build times and the amount of data to down- and upload.

Obviously I want as few hard deps as possible but I'll have to investigate whether we can do dynamic loading of deps and then have some opt deps. However, I don't see which of our deps pulls on a desktop environment.

Turns out pytorch doesn't support dynamic loading of libraries and I'd rather offer more features in the package than fewer.