I'm a bit surprised that conda is not mentioned anywhere (I'm not saying as best practice, but as an alternative or why envs are better)
Just right now I'm trying to fire up a new instance on GCP. With a completely clean image, doing a conda install hangs for 30 minutes while it's trying to "solve" something.
I supported a research cluster, and the amount of times conda caused an issue was a lot.
I use this with my m1 mac, and it works great.
[1] -- https://docs.conda.io/en/latest/miniconda.html#linux-install...
You can conda install the CUDA dependencies and then install the required Tensorflow version via conda pip. But that's not much different to installing CUDA manually and then installing tf from system pip.
It's much faster and easier to pull a tf Docker image as it's their "officially supported" way to get up and running.
So... as a tf user and a sysadmin... Nah. No conda for me thanks.
E.g. https://github.com/conda/conda/issues/8087, https://www.anaconda.com/blog/understanding-and-improving-co...
Oh and if there is a conflict somewhere, it goes into some conflict detection routine that will take hours and not produce anything useful.
I could go on, but I have come to really dislike conda.
This is why mamba [0] was created. It is a C++ reimplementation of conda for much better performance. mamba is a drop-in replacement of conda and can operate on the same anaconda, condaforge (and mambaforge) repositories.
I use Gentoo and its package manager is written in python. Even though it is more complex (IMO) it doesn’t have nearly the same slowness when it comes to dependency resolution and conflict detection.
With all scientist packages, conda is pretty good but it is sooooo slow, it is just insane!
It's very disillusioning to see how sheer twitter-followings and "popularity" type metrics drive development these days by forcing alternatives to be de-facto neglected. Everyone does what's "hot", so all the tutorials and bug reports and tests and SO questions and new libraries and and and all go towards that framework or language or tool or method. You can't even argue technical merits towards the neglected options because yes the popular tool is better, but only because we have a metric boat load (millions) of man-hours being pumped into making it better instead of all the alternatives. It's like the tech-equivalent of fashion fads in that it's self-reinforcing. Not to take away from some of the actual and technical achievements that some of these things have made, of course.