Fedora AI Developer Desktop Objective

[ngompa (Neal Gompa)]

@dmellado

And the idea that we bring people in through the proprietary path and then nudge them toward open alternatives… in my experience that just doesn’t play out. The proprietary path becomes the default, the open one stays “nice to have,” and nobody gets around to pushing it. I do totally support what @ngompa said.. if we don’t do the hard thing here, nobody will. ROCm has in-tree drivers. Vulkan Compute and OpenCL work across vendors. Tim’s question about whether this is essentially “make CUDA easier on Fedora” deserves a straight answer.

To reinforce this point, one reason that NVIDIA became an outlier in the first place is because we exerted pressure and reinforced our ethical and ideological stance here that non-free/closed technologies were unwanted. We have compromised in some respects (e.g. firmware), but in most respects we historically haven’t. That made AMD eventually capitulate and retire fglrx for amdgpu. And now AMDVLK has been replaced with the openly developed RADV. These are wins that you don’t get unless you hold firm.

I use Fedora because we do stuff like that to help drive the future of open technologies and promote Free Software. I use and develop Fedora KDE because I believe firmly that we have to be the change we want to see, and that means working through hard things around multimedia, working through hard things around usability, and working through hard things around hardware support.

We have intentionally kept the NVIDIA proprietary stack at arm’s length on purpose. Even the existing AI/ML SIG (which was not engaged or consulted on for any of this proposal) threads this needle at least reasonably well by focusing on stacks that leverage open technologies and actively working in upstream projects to enable them.

I respect that @gordonmessmer has a particular view of problems he wants to solve. We even had a productive in-person conversation about it early last week and there’s even some alignment on things to fix. That being said…

@gordonmessmer

I honestly wish I knew how to spend less time discussing NVIDIA. I think that enabling NVIDIA hardware is important, in as far as the status quo drives a lot of people toward other projects by virtue of the hardware they are using, and the loss of their participation hurts the project. But I also think that this is the smallest problem to solve, and that is why I predict that it will be completed before any other work.

It’s pretty easy, actually. Instead of talking about NVIDIA, talk about what they use NVIDIA for and work with people to close the gaps there. The Linux market is majority AMD/Intel, not NVIDIA for GPUs. The client-side “AI” market is almost 40% non-NVIDIA between Apple Silicon and Intel/AMD laptops. The cost of NVIDIA GPUs has skyrocketed too, making non-NVIDIA hardware more attractive (even in DCs), provided the software enablement is present.

Most AI/ML projects do not interface with CUDA directly, they use abstractions like libtorch, litert, blas, and so many other things that allow mathematics to be represented more cleanly. Enable those to use open technologies and promote the fact you’re making things more accessible by using more accessible hardware.

@gordonmessmer

But that’s just the first problem. The interesting and difficult and very important work will take longer. As I said above, developers are also telling us that it’s hard to use Fedora or target Fedora, independent of NVIDIA. There’s an audience of application developers that we’re not serving well, and I see evidence of that everywhere. In our meeting this morning, we discussed the need to contact maintainers and coordinate updates of dependencies, and I think that’s backward. Fedora’s structure works well for building a collection of mature components that move in unison.

Honestly, this is more mixed in my experience. It entirely depends on the audience, the technologies, and their awareness. At least for classical numerical work, Fedora serves that audience pretty well and has for a very long time. Scientific research is an area we’ve done reasonably okay in. Commercial ML has been weaker because for a long time it was centered on TensorFlow, but these days it’s on libtorch through PyTorch and other more accessible technologies. Furthermore, Fedora’s AI/ML SIG (which absorbed the existing high-performance computing folks too), has not actively marketed their work. That results in a perception that those particular technologies aren’t easy to use in Fedora. On a lark, I actually tried out ollama from the repositories with the Gemma4 model on my Framework 16, and it just worked. Other open source models worked as well. So we have the basic pieces already, we just need to promote them in the way we want to push the world to look at things.

The audiences I tend to talk to actually say Fedora is the easiest platform for them to get started on. Game development, multimedia production, software development, and other similarly advanced tasks are extremely well-supported in Fedora. Enterprise software development in Fedora has gotten worse as Red Hat has slowly pulled out of Fedora Java and .NET still lacks real guidance on how to use it, but all the rest of the software development spaces are reasonably healthy.

@jflory7

The Council reviewed the Fedora AI Developer Desktop Initiative proposal with @gordonmessmer. Discussion covered naming concerns and hardware enablement, specifically the need to address the LTS kernel feedback in the Fedora Wiki proposal. Recognizing that the technical work is already underway and aiming to provide official guidance and legitimacy, the Council voted to approve the initiative. The approval is pending a short lazy consensus period ending on Friday, May 8th to accommodate absent Council members during the meeting.

But, I am extremely unhappy with the Fedora Council essentially ignoring the community discussion that indicated that at the minimum this was probably not okay to greenlight as-is. I’m especially disappointed at how we’re being told that as our opinions and interests in the project as highly engaged contributors do not matter.

Fedora as a community project does not function if we all decide to pack up and leave. There are even more disastrous outcomes that can occur if people continue to feel steamrolled.