Blame NVIDIA instead. If they had open source their blobs, it would have been everywhere out of the box. Most evil company for open source ever.

>Most evil company for open source ever.

I won't ever defend nvidia, but... I really think Broadcom is worse. They've ruined tons of routers and wireless cards. Also, the Raspberry Pi is full of blobs because of them.

Raspberry Pi seems to be doing just fine with them, whereas the FOOS competition is where exactly?

Allwinner, Rockchip and Mediatek seem do be doing just fine without all the blobs.

So where are their Raspberry Pi like boards being adopted like crazy by the Maker community?

I’m not looking to blame anybody. System76’s approach just conflicts with my priorities when choosing an Ubuntu-based distro. I prefer to not have my OS surprise me with requirements that I replace a major portable dev tool standard with one of their bespoke unmaintained side projects. I doubt I am alone and maybe others will find the info useful. I just can’t trust their work any longer after realizing they were serious about that tool of theirs being a replacement.

Hollywood and HPC are don't seem to care about it.

`sudo apt install nvidia-docker2` works on 21.10. I have been packaging NVIDIA's container toolkit since tensorman was created, because tensorman doesn't work without nvidia's container toolkit. It's a dependency.

Nvidia-docker is not the same as nvidia-container-toolkit, and neither of them have official support for 21.10 the last time I checked (a few months ago).

It is a component of nvidia-container-toolkit. Take one minute to browse the Pop repository and you'll find three packaging repositories forked from NVIDIA. libnvidia-container, nvidia-container-toolkit, and nvidia-docker. Installing nvidia-docker installs support for the whole system. Now you have functioning NVIDIA support in containers.

You mean the one that you forked 12 days ago? What are you trying to accomplish by acting like it has been packaged all along?

https://github.com/pop-os/nvidia-docker/commit/aa9fbcff6e6e5...

This changes nothing about my story of the Pop OS developers previously thinking it could be replaced with their bespoke custom tooling.

This is a new repository that replaces a pre-existing nvidia-container-runtime repository that we've been using since 20.04 that does the same thing. It still remains that tensorman would not function without NVIDIA docker container support, because it's literally just a convenient frontend for interacting with Docker's CLI.

I never said otherwise. As you know, nvidia-containter-toolkit only very recently merged with nvidia-docker, and you know perfectly well that you have not always packaged or supported nvidia-docker.

They occasionally have driver mismaatches between the upstream containers and the distro-provided driver. This happens because nVIDIA only supports Ubuntu LTS releases and Pop moves faster. I solved it forever by pinning nvidia's repos to a higher priority. Otherwise I can't figure out what "unmaintained project" you're talking about, or what "tooling standards" are.

They are most likely referring to `tensorman` https://support.system76.com/articles/tensorman/ which is regrettably premised on the notion that you have any intention of using Tensorflow in the first place.

Unfortunately, the `tensorman` article shows up first (after several complaint/support threads) in a google search result for the subject https://www.google.com/search?hl=en&q=nvidia%20docker%20pop%...

In all actuality, the correct article to reference is https://support.system76.com/articles/cuda/ but alas, this is presently broken on PopOS 21.04 unless you pin nvidia's repo's as you've suggested.

At the very least, an unpleasant experience. Probably some blame to be shared by system76 and nvidia.

It wouldn’t be a significant if the Pop OS developers weren’t telling users that they don’t need nvidia-docker. Of course, they can tell users whatever they want, but I’m not going to trust their OS if this is how they plan to handle things like this.

I am actually participating in a thread right now where contributors are happily trying to solve this issue. Their devs are very helpful if you reach out.

https://github.com/pop-os/pop/issues/1708

Since when have we told anyone that they don't need nvidia docker packages installed? You can validate that the Debian packaging has always shown nvidia-container-runtime/nvidia-docker as a recommends for CUDA support.

Can you expand on what in your view System76 did to impede it?

I’m sorry, I don’t quite recall the details but it was a few different things. The main one is forcing certain Nvidia libraries in the package manager so you need to take a few steps to undo some of the Pop OS Nvidia installation, which breaks other Pop OS stuff, of course.

There is no need to undo anything. We regularly test CUDA functionality with our packaging, and it's the whole point of having the packaging in the first place. So you don't have to manually install things from a third party repository or deb.

“CUDA functionality” and nvidia-docker are not the same thing and you know it. This is the attitude that has me turned off off. Just acknowledge and we could move on! Unless you deleted them, anyone can search your github to find your comments telling users to use your bespoke tensorman tool instead of installing nvidia-docker, acting like one is a replacement for the other. And with the same presumptive attitude.

I have no idea why you're trying to argue with me over absolutely nothing. You could have saved so much time by asking questions instead of accusing me of things I haven't done, and spreading misinformation about it.

CUDA support in Docker requires nvidia-docker+nvidia-container-toolkit+libnvidia-container, and that is the whole point of having nvidia-docker installed. We validate that CUDA functions inside of Docker when pushing updates to nvidia-container-toolkit and friends.

Tensorman is not a replacement for nvidia-docker, and you're not going to get anywhere by trying to convince me about the functionality of something that I personally wrote. It is quite literally just a simple command-runner that runs docker commands, specifically for the purpose of managing the official Tensorflow Docker images, and getting a more streamlined setup for managing your local Docker images based on them. The sole purpose is to replace our previous tensorflow packaging in Pop!_OS that became impossible to package because newer versions of Tensorflow did not build beyond 18.04. Tensorflow themselves recommend using their docker images instead of trying to package and install their libraries on your host OS.

OK here is a question: Why do you tell people to use tensorman instead of nvidia-docker when they have problems with nvidia-docker?