<@tflink:fedora.im>
16:30:16
!startmeeting fedora-ai-ml-sig
<@meetbot:fedora.im>
16:30:17
Meeting started at 2026-04-23 16:30:16 UTC
<@meetbot:fedora.im>
16:30:17
The Meeting name is 'fedora-ai-ml-sig'
<@tflink:fedora.im>
16:30:32
!hello
<@zodbot:fedora.im>
16:30:36
tflink: Tim Flink (tflink)
<@jflory7:fedora.im>
16:30:38
!hi
<@zodbot:fedora.im>
16:30:38
Justin Wheeler: Justin Wheeler (jflory7) - he / him / his
<@gordonmessmer:fedora.im>
16:30:46
!hi
<@zodbot:fedora.im>
16:30:48
Gordon Messmer: Gordon Messmer (gordonmessmer)
<@cybette:matrix.org>
16:31:13
!hi
<@zodbot:fedora.im>
16:31:15
Carol Chen: Carol Chen (cybette) - she / her / hers
<@tflink:fedora.im>
16:31:59
I know some folks ended up with conflicts today so we'll probably get started without them in a few
<@tflink:fedora.im>
16:33:41
ok, lets get this party started :)
<@tflink:fedora.im>
16:33:55
!topic Discussion about a framework for AI/ML app development, and onboarding path for Fedora Infra?
<@tflink:fedora.im>
16:34:29
Justin Wheeler: this is yours
<@jflory7:fedora.im>
16:34:45
Oh my 😄
<@jflory7:fedora.im>
16:34:58
OK, I promised I would get this into a ticket, but I can try and riff a bit on what's in my head
<@jflory7:fedora.im>
16:35:05
Maybe this meeting will help out with refining
<@jflory7:fedora.im>
16:35:08
OK, so—
<@tflink:fedora.im>
16:35:36
tom is in another meeting and will probably be late, xanderlent said that he had a last minute $DAYJOB conflict which leaves us with this topic to start
<@tflink:fedora.im>
16:36:26
wasn't that the deal for approving your PR for default ticket types? :-D
<@jflory7:fedora.im>
16:36:35
Ugh, I know 😄
<@jflory7:fedora.im>
16:36:42
Darn Release Party!!
<@jflory7:fedora.im>
16:37:00
!info 1. Related to the framework, I had this novel use case for building a small prompt project / reusable AI definition for summarizing Fedora meetings. I didn't know where to put this or host it, so I asked to build in our SIG's Forgejo organization. Now, the project is mostly "done" and is being refined. It provides a use case / example of how to build things with AI for reuse in the community.
<@jflory7:fedora.im>
16:37:03
!link https://forge.fedoraproject.org/ai-ml/meeting-secretary
<@jflory7:fedora.im>
16:38:14
!info 2. However, while this works for _me_ (@jflory7), what about others in the community building apps and tools with AI? We do not have a good way to connect these people together, help them collaborate, and share learnings. I wondered if there might be way that the AI/ML SIG might be able to provide some support or guidance to developers who want to integrate their `$THING` into Fedora Infrastructure officially, but the AI use is questionable or possibly costly.
<@jflory7:fedora.im>
16:38:21
I guess this is the net-sum of the idea.
<@jflory7:fedora.im>
16:38:26
!link https://abompard.fedorapeople.org/twif/
<@jflory7:fedora.im>
16:38:49
Another example, but I want to have a place to talk to Aurelien about his TWIF project or the Meetbot meeting summarizer bot.
<@jflory7:fedora.im>
16:39:00
How could those become "official-official" instead of a solo project by a single person?
<@jflory7:fedora.im>
16:39:14
I think our SIG could be a useful place to have some conversations about this?
<@tflink:fedora.im>
16:39:49
I think that a lot of that will require infra's input at a minimum
<@jflory7:fedora.im>
16:39:57
I guess this is where I'm at on this. Because of the Release Party tomorrow and being on PTO next week, I haven't had much time to dig into this. I am not even sure I will have time for it before RH Summit 😄 But it would be good to have an open conversation
<@jflory7:fedora.im>
16:40:08
If we do it, of course, they will have to have input
<@jflory7:fedora.im>
16:40:33
But I also feel like Infra would love to have someone else helping vet AI usage before someone tries to launch something expensive in CommuniShift and use the new GPU hardware we have coming in
<@jflory7:fedora.im>
16:40:57
I wanted to start the conversation here first, because it would be a change in scope for this SIG a bit
<@jflory7:fedora.im>
16:41:08
If we feel this is not the best place, then I don't need to talk to Infra 🙂
<@tflink:fedora.im>
16:41:10
in general, I'm fine with having these conversations as part of the SIG. If we get to a point where it should be split off, we can cross that bridge when we get there
<@jflory7:fedora.im>
16:42:39
So, maybe the action might be an actual proposal and written framework/process for how this might work
<@tflink:fedora.im>
16:42:50
but I also think there are some underlying questions in this that would need to be addressed. I assume that you're talking about apps which could utilize a service that's hosting an openai compatible api or something similar?
<@jflory7:fedora.im>
16:42:53
Then we could meaningfully discuss details, and then engage with Infra on what we come up
<@jflory7:fedora.im>
16:43:43
Potentially! Once we get some sort of MCP functionality in the community (I don't know this will happen but I suspect it is inevitable), it would be nice to provide guidance and tips for how to use these things in Fedora apps, but that is getting a bit ahead of myself, I think?
<@tflink:fedora.im>
16:45:12
I don't even mean providing API keys to anyone, even something like the meetbot summerizer would use some endpoint, no?
<@tflink:fedora.im>
16:45:40
and after that there are questions about how much hosting we'd do, how we'd do it and which models would be available
<@trix:fedora.im>
16:47:11
!hi
<@zodbot:fedora.im>
16:47:12
Tom Rix: Tom Rix (trix)
<@trix:fedora.im>
16:47:25
sorry had a conflict...
<@jflory7:fedora.im>
16:48:22
I think _we_ would not host, this is what CommuniShift and proper infra is for. But, our SIG could be a place to get feedback and ideas. This could shape a CommuniShift request to get infra access if needed
<@jflory7:fedora.im>
16:48:45
I think Infra does not want to spin up a CommuniShift environment and then have it get hammered to death from AI queries that are badly designed or work poorly
<@jflory7:fedora.im>
16:49:10
Yeah, I think for now, the app developer can provide that, and part of "officialiazing" an app would be figuring out things like this endpoint
<@jflory7:fedora.im>
16:49:26
Definitely some coordination required with Infra to do something like this
<@tflink:fedora.im>
16:49:27
in my mind there are those components: HW needed for hosting, interface to said HW (cloud, on prem etc.), what is hosting the model(s), which models are hosted and then how apps utilizing those models would be hosted. there's also the potential need to rate limit or prioritize things since capacity would be limited
<@jflory7:fedora.im>
16:49:36
I am also kinda wondering if I am over-engineering this
<@jflory7:fedora.im>
16:49:49
But an AI/ML experiment home and place to collaborate with others is really want I want
<@tflink:fedora.im>
16:50:26
but maybe I'm trying to answer a different question than the one you're asking
<@jflory7:fedora.im>
16:50:53
This makes sense. On providing hardware, we immediately touch on the Fedora Infra touchpoint. This is where we need to have a chat. For the parts like rate-limiting or prioritizing what is most important, I think our SIG could be good advisors/stewards about this
<@tflink:fedora.im>
16:51:20
is the idea to emphasize what could be done with a 100% Fedora and open source stack or would the idea be more to emphasize a community building things on top of stuff that might not be Fedora?
<@jflory7:fedora.im>
16:51:32
Ideally, our SIG does a bit less on the actual hardware access – this is where we need infra. But sometimes people bring their own stuff or their own keys. If they have their own local dev environment, that works, and we could help them get to CommuniShift, if they wanted that help
<@cybette:matrix.org>
16:52:36
my 2 cents: I don't think either the ai-ml group nor CommuniShift / infra should host. It's one thing to create a repo for apps / gems / agents / whatever is the correct term nowadays, in this case gems for Google Gemini. Others can take the gems and reuse them with their own access to Gemini (personal account, their own Google enterprise account etc.) but I don't think the expectation is for Fedora to host the models
<@tflink:fedora.im>
16:52:39
does infra have the capacity or interest in dealing with LLMs and/or accelerator HW?
<@jflory7:fedora.im>
16:52:47
<@jflory7:fedora.im>
16:52:47
I am not sure the goal is to provide a 100% FOSS stack for AI development, given the closed nature of many LLMs. I feel optimistic this will eventually begin to change and more options will become available in the future. For now though, 100% FOSS in the sense of LLM dependency might not be possible. But the source code implementing LLM use should be FOSS of course.
<@jflory7:fedora.im>
16:52:47
<@jflory7:fedora.im>
16:52:47
The apps and tools can be community-focused, but I _do_ think it needs to be Fedora-focused. Just like my Meeting Secretary repo is
<@jflory7:fedora.im>
16:52:47
Hmmm. I think it is in the middle.
<@cybette:matrix.org>
16:52:48
Btw, Gems can also be used with the Gemma (open source) models
<@jflory7:fedora.im>
16:53:52
There is some backchannel, I think Adam Williamson is the person I am thinking of here, about future GPUs being installed in our Fedora infrastructure for more AI enablement sort of work. I think Fedora providing an appointment or hosting a model for community-specific use might be possible, but likely not any time soon?
<@jflory7:fedora.im>
16:54:01
Oh, this is neat, did not know that 👀
<@jflory7:fedora.im>
16:54:22
They are actively building this capacity. It might be slow-going until Pagure.io is gone and Flock happens.
<@cybette:matrix.org>
16:56:49
I'm not saying that Fedora shouldn't be hosting for AI related needs, but for the framework Justin is proposing, a place for some form of AI development, maybe it doesn't need to be tightly coupled with infra.
<@trix:fedora.im>
16:56:50
how are we doing wrt GPU's installed into fedora infra ? Specifically AMD ones ?
<@jflory7:fedora.im>
16:57:36
This is the hardware I specifically know about that is incoming, but I think you'd have to ask in #admin:fedoraproject.org. Adam Williamson or Greg Sutcliffe might have more context.
<@gordonmessmer:fedora.im>
16:57:38
I'm talking to infra people about where 1 host lives, and how it's configured, and what workloads it can support
<@gordonmessmer:fedora.im>
16:58:08
we had planned to get more hosts up this quarter, but right now it looks like a delay due to RAM shortages
<@jflory7:fedora.im>
16:58:11
I guess it does not have to be coupled with infra. What do you see as the alternate? I admit that of the people in the room, I feel like I have the least in-depth knowledge in this space, so I am happy for y'all to suggest other things to me too 😄
<@jflory7:fedora.im>
16:58:19
At least, to help me guide a "first draft" proposal
<@trix:fedora.im>
16:58:21
is it up and running ?
<@tflink:fedora.im>
16:58:24
either way, I wonder if the conversation might be better advanced with a more specific proposal
<@tflink:fedora.im>
16:58:34
kinda
<@jflory7:fedora.im>
16:58:49
Oh yeah, Gordon Messmer would probably know this too 😄
<@gordonmessmer:fedora.im>
16:59:10
it is, but I've been asked to reconfigure it. Trying to get some questions answered right now
<@jflory7:fedora.im>
16:59:21
I heard through the grapevine that a lot of Dell stuff on backorder got cancelled by Dell because of supply chain issues?
<@trix:fedora.im>
16:59:30
i have a lot of interest in not having my own lab of stuff nor having to do all the ci
<@jflory7:fedora.im>
16:59:53
For sure. It is super abstract right now, there is no ticket, and that's on me. But if y'all aren't totally opposed to this, let me work on something when I'm back from PTO next week.
<@jflory7:fedora.im>
17:00:06
Oh right. RH Summit too 😅
<@jflory7:fedora.im>
17:00:20
Let's schedule me for a follow-up in the week of May 4th, at least
<@jflory7:fedora.im>
17:00:32
I might not have a lot of progress but there should absolutely be a tracking ticket by then 🙂
<@trix:fedora.im>
17:00:36
what will be usable by F45 branch .. August ?
<@jflory7:fedora.im>
17:01:11
!info No formal ticket or proposal yet for the AI/ML app development framework concept. It was pitched and tested out as a concept in this meeting. Overall, feedback is warm but there are some major questions to consider for this to be viable.
<@jflory7:fedora.im>
17:01:46
!action @jflory7 due:2026-05-07 Create AI/ML SIG ticket and/or draft a proposal for the app development framework for further discussion and review.
<@tflink:fedora.im>
17:01:52
not currently known, we're waiting on some input before we can answer that question
<@jflory7:fedora.im>
17:02:03
!agreed Hard to discuss details _without_ a proposal right now, but we will revisit and discuss this more in May.
<@jflory7:fedora.im>
17:02:10
EOM
<@cybette:matrix.org>
17:02:14
well, maybe not the best analogy, but let's say I write an Android app. I don't have to provide you with the phone, you bring your own device. So in this case, you write a gem and share it, and people bring their own models. those without models (devices) may get some access some other way, but it's not tied to the app/gem.
<@jflory7:fedora.im>
17:02:15
I think this is all I need on this topic right now
<@jflory7:fedora.im>
17:02:28
Yeah, that makes sense!
<@jflory7:fedora.im>
17:02:38
Good point
<@jflory7:fedora.im>
17:03:19
!info Open question: Does this need to be tightly coupled to infrastructure? Or could it be more general or basic guidelines for AI/ML-driven development?
<@tflink:fedora.im>
17:03:24
yeah, I think that more detail is needed before we have much more discussion but getting past the "is this a hosting or community building conversation" has at least helped narrow things down
<@jflory7:fedora.im>
17:03:43
Yep! This inches me toward the direction to take the ticket
<@tflink:fedora.im>
17:03:51
anything else on this or should we move on?
<@jflory7:fedora.im>
17:03:58
I will do my best to get it opened at least before I go on leave, but… ahhhh tomorrow is stressing me out 😄
<@jflory7:fedora.im>
17:04:08
I'm fine to move on personally
<@jflory7:fedora.im>
17:04:26
(BTW… don't forget to register for tomorrow's Release Party!!) https://fedoramagazine.org/join-us-for-the-fedora-linux-43-release-party/
<@jflory7:fedora.im>
17:04:32
Wait
<@jflory7:fedora.im>
17:04:35
Not that link 🫣
<@jflory7:fedora.im>
17:04:50
https://fedoramagazine.org/fedora-linux-44-release-party/
<@jflory7:fedora.im>
17:04:54
That's the one 😛
<@tflink:fedora.im>
17:04:59
I was wondering if you had a time machine somehow :)
<@jflory7:fedora.im>
17:05:03
Heheh
<@tflink:fedora.im>
17:05:18
anyhow, moving on to the next topic
<@tflink:fedora.im>
17:05:22
!topic ROCm 7.2.2 update in fedora rawhide, likely EPEL 10.3 target
<@tflink:fedora.im>
17:05:27
Tom Rix: take it away
<@trix:fedora.im>
17:05:54
7.2.2 was a very minor update and maybe the last in 7.2.x series
<@tflink:fedora.im>
17:06:20
!info ROCm 7.2.2 was a minor update
<@trix:fedora.im>
17:06:28
maybe 5 packages ?
<@tflink:fedora.im>
17:07:04
!info ROCm 7.2.2 may end up in EPEL 10.3 but a final decision has not yet been made
<@tflink:fedora.im>
17:07:32
wasn't 7.2.2 mostly a bugfix update?
<@trix:fedora.im>
17:07:46
yes, minor stuff.
<@tflink:fedora.im>
17:08:27
anything else on this?
<@trix:fedora.im>
17:08:54
So it will be 'calm' on the fedora release for a while. development is happening on 8.x front.
<@trix:fedora.im>
17:09:20
the big feature we are chasing is generic gpu support.
<@trix:fedora.im>
17:09:39
and of course the usual not breaking anyone.
<@tflink:fedora.im>
17:09:43
which translates into smaller builds and wider HW support
<@tflink:fedora.im>
17:10:02
generics translates into smaller builds and wider HW support, to be clear
<@trix:fedora.im>
17:10:11
yes. the gpu list is the big reason for slower builds.
<@trix:fedora.im>
17:10:24
we will go from 20+ to maybe 6.
<@tflink:fedora.im>
17:10:40
!info ROCm 8.x is still under development, plans may change as we get closer to release
<@trix:fedora.im>
17:11:11
and from good HW support to ALL HW support. sorry no old 803 support, we eol that year ago.
<@tflink:fedora.im>
17:11:15
!info hopefully generics will be coming to Fedora's ROCm packages which will translate into wider HW support and smaller/quicker builds
<@tflink:fedora.im>
17:11:32
cool, anything else on this? any questions from other folks?
<@trix:fedora.im>
17:11:43
that's it from me.
<@tflink:fedora.im>
17:12:47
ok, moving on, then
<@tflink:fedora.im>
17:13:04
!topic open floor
<@tflink:fedora.im>
17:13:21
!info xanderlent was not able to make it today, his topics will be bumped to the next meeting
<@jflory7:fedora.im>
17:13:46
Oh? 👀
<@jflory7:fedora.im>
17:13:59
ROCm working with non-AMD hardware? Do I get that right?
<@trix:fedora.im>
17:14:08
no.
<@trix:fedora.im>
17:14:17
it works with all amd hw.
<@tflink:fedora.im>
17:14:24
generics is for AMD hardware but treating it more generically instead of having to specify every single instruction set that it's built for
<@trix:fedora.im>
17:14:25
we have holes in support now.
<@jflory7:fedora.im>
17:14:52
Ahhhh! So, kinda like a more unified API for the hardware, if I translate this in my developer mind?
<@jflory7:fedora.im>
17:14:56
(for AMD hardware)
<@trix:fedora.im>
17:15:02
ex/ gfx1032 on my workstation is not support because i did not build gfx1032.
<@jflory7:fedora.im>
17:15:13
I see. That does sound painful
<@jflory7:fedora.im>
17:15:29
Reminds me of the old days of compiling Realtek WiFi drivers into my Linux kernel 🙂
<@tflink:fedora.im>
17:15:31
for example, instead of having to specify every RDNA3 sub-variant, (there are a lot of them), we can just build for what is effectively RDNA3
<@trix:fedora.im>
17:15:53
in 8.x it will just be there as gfx10-3-generic target will be replace all the individual gfx103x targets we have.
<@tflink:fedora.im>
17:16:08
users won't notice anything other than more supported AMD GPUs for ROCm and hopefully smaller packages to download
<@trix:fedora.im>
17:16:37
less cursing about no support for older or really new hw.
<@trix:fedora.im>
17:17:46
it is a tradeoff atm, some of our libraries would not build in 2 days if we used all the targets.
<@tflink:fedora.im>
17:18:13
There will hopefully be more AMD NPU support in Fedora soon but that's still pretty early
<@jflory7:fedora.im>
17:18:24
Makes sense to me
<@tflink:fedora.im>
17:18:37
I figured that would come up with xandlerlent's topic but that'll have to wait 2 weeks
<@trix:fedora.im>
17:20:04
we are trying to get 'lemonade' in which has the npu backend. i am interested to see how npu does on smaller laptop gpu's. i am sure not everyone has a strixhalo laptop.
<@tflink:fedora.im>
17:20:56
is the lemonade review posted or is that still needing work?
<@trix:fedora.im>
17:21:08
i have the review.
<@tflink:fedora.im>
17:21:22
ok, I couldn't remember if that had been pushed public yet
<@trix:fedora.im>
17:21:38
it needs work. ex/ to see it actually goo, some backend needs to be enabled.
<@trix:fedora.im>
17:22:15
i think it assume ggml is installed, but has not Requires:
<@tflink:fedora.im>
17:22:35
eh, it really likes to download backends if it isn't told not to
<@trix:fedora.im>
17:23:00
yeahhh..
<@tflink:fedora.im>
17:23:05
anyhow, we're kinda getting into the weeds here
<@tflink:fedora.im>
17:23:25
any other topics for open floor?
<@tflink:fedora.im>
17:23:37
or shall we adjourn?
<@trix:fedora.im>
17:23:45
i'm good.
<@tflink:fedora.im>
17:25:35
ok, thanks for coming everyone
<@tflink:fedora.im>
17:25:44
!endmeeting