Hey folks! 
Here is the summary of the July 16, 2026, Fedora AI/ML SIG meeting.
The AI/ML SIG works on AI and machine learning toolchains in Fedora, including packaging, inferencing tools, and hardware enablement. Today’s meeting covered llama.cpp backend recruitment for F45/F46, the newly packaged pi-coding-agent, a vote to create the Skills Reviewers sub-team, and an initial discussion on GPU hardware access governance.
Topics discussed in this meeting:
llama.cppbackend maintainers needed for F45/F46pi-coding-agentin F45/rawhide: testing and documentation- Establishing a Skills Reviewers sub-team (ai-ml/tickets#31)
- GPU hardware access scoping and acceptable use (ai-ml/tickets#34)
Note: Claude Opus 4.6 was used to summarize the Meetbot-generated HTML log of the public meeting. I edited the AI-generated output before making this Fedora Discussion topic. If you notice mistakes, please provide a correction as a reply to this topic.
Links to Meeting Logs
Executive Summary
Tom Rix (@trix) outlined two areas needing community help for F46: llama.cpp backend maintainers (especially for Vulkan, to keep Vulkan support in Ollama) and testing of the newly packaged pi-coding-agent on non-x86 architectures. Kevin Hannon (@kannon92), a newcomer attending his first SIG meeting (
), volunteered to review and expand the pi-coding-agent documentation on the Fedora Wiki. @trix also noted that lemonade-server has landed in Fedora and shared his exploration of skill packages for pi-coding-agent, including pi-graphify for large codebase analysis.
The SIG agreed to create a “skills-reviewers” team on Forgejo (ai-ml/tickets#31). Five people volunteered over the past two weeks, and together with @jflory7, the team will be seeded with six members: @jflory7, @cverna, @davdunc, @lbarczio, @cybette, and @ttomecek. The team gets write access to the ai-ml/skills-library repository with at least one reviewer approval required before merge. Once reviewer guidelines are in place, an open call will go out on Matrix and Discourse for contributions. David Duncan (@davdunc), returning to the SIG after a long absence, and @jflory7 offered to support @lbarczio on drafting reviewer guidelines.
On GPU infrastructure (ai-ml/tickets#34), Tim Flink (@tflink) and @trix both indicated it is too early to formalize access policies given the number of moving parts around hardware procurement and ownership. The SIG agreed that the available GPU hardware is in a “transitionary/exploration” phase focused on CI for the F45 release cycle. Any interim access requests will be closed as wontfix with that explanation. A follow-up is planned for the August 13th meeting, by which time @tflink hopes there will be more clarity. Gordon Messmer (@gordonmessmer) noted that an Acceptable Use Policy will be needed regardless of the access mechanism, since “CI” describes a mechanism, not a purpose. @jflory7 will open a separate ticket for that AUP discussion, targeting Q4.
Worth noting, we also welcomed two newcomers: @kannon92, who works on Kubernetes and has been experimenting with llama.cpp and pi-coding-agent, and @davdunc, who works primarily on the Cloud SIG and is interested in building images for cloud hardware and AI workloads.
Action Items
- @kannon92: Review and edit the existing
pi-coding-agentdocs on the AI/ML SIG Fedora Wiki page, and coordinate in #ai-ml:fedoraproject.org on changes as needed. - @jflory7: Create skills-reviewers team on Forgejo, then close issue ai-ml/tickets#31. Continue planning and working on Agent Skills in the
ai-ml/skills-libraryrepo. - @jflory7: Open a new ticket for Acceptable Use Policy discussion, likely targeting Q4.
- @jflory7:
Post a meeting summary of today’s meeting on Fedora Discussion.
Detailed Meeting Log
Roll Call and Introductions
@jflory7 opened the meeting and called roll. Attendees: @jflory7, @trix, @gordonmessmer, @tflink, @davdunc, and Kevin Hannon (@kannon92, attending via @kph1992:matrix.org). @kannon92 introduced himself as working on Kubernetes with interest in llama.cpp and pi-coding-agent. @davdunc introduced himself as primarily working on the Cloud SIG, with extensive packaging experience and interest in cloud hardware images for AI workloads. @jflory7 welcomed both newcomers and directed them to the ticket tracker and the #ai-ml:fedoraproject.org Matrix room.
llama.cpp F45/F46 (@trix)
@trix outlined the forward-looking needs for llama.cpp in F45/F46. The package needs maintainers for new backends, with Vulkan being the highest priority to keep Vulkan support in Ollama. CI is also needed to catch stability issues; @trix noted that Vulkan support was broken at one point but may now be working again, and CI would help prevent regressions.
@jflory7 asked whether @trix had posted to the devel@ mailing list and what the urgency level was. @trix clarified this is an outreach effort, not a blocker: he wants people who don’t see their favorite backend or CPU architecture on the list to step up. As an AMD employee, his testing and development focus is AMD hardware, leaving gaps for Vulkan, Intel, Nvidia, PowerPC, and s390x.
@kannon92 expressed interest in Vulkan, BLAS, and CPU builds but noted he lacks GPU hardware. @davdunc asked whether functional testing could run on AMD instances without GPUs. @tflink noted that llama.cpp does work on CPU but currently pulls in all of ROCm as a dependency. @kannon92 confirmed he hit a failure when trying to use the package without a GPU.
@jflory7 briefly raised whether Fedora Packaging Guidelines should address AI-related packages. @tflink observed that the backend pattern is not unique to AI. @trix agreed, and @jflory7 deferred it as a future idea.
pi-coding-agent in F45/rawhide (@trix)
@trix announced that pi-coding-agent is now available in F45 and Rawhide, giving Fedora users access to agentic coding tools. He is looking for help with non-x86 architecture testing (priority on aarch64) and for contributors to hook up local LLMs other than Ollama.
@trix also mentioned that lemonade-server has landed, which works with NPUs, and shared a pi skill package he is exploring: pi-graphify, which could make understanding large codebases like the Linux kernel more tractable.
@tflink asked whether pi-coding-agent would be available on older Fedora releases. @trix acknowledged the tension between fast-moving AI tooling and stable release support.[1]
@kannon92 volunteered to work on the pi-coding-agent wiki documentation. @trix shared the existing wiki section and noted he is looking for people to contribute hookups for LLMs beyond Ollama.
Ticket: ai-ml/tickets#31 (Establish a Skills Reviewers Sub-team)
@jflory7 reported that since the ticket was introduced two weeks ago, five people volunteered to join the skills-reviewers team: @cverna, @davdunc, @lbarczio, @cybette, and @ttomecek. @lbarczio also proposed packaging-focused skills (scratch build fixing, patch backporting) and volunteered to draft a human reviewer guide. @cverna opened a CI/linting discussion in ai-ml/skills-library#13.
@jflory7 called a vote to create the team, seed it with six members (himself included), and post an open call on Matrix and Discourse. @davdunc, the only volunteer present, expressed full support. No objections were raised.
@jflory7 encouraged @davdunc to support @lbarczio on drafting reviewer guidelines as a first step. @trix noted there is “a lot of energy” around skills work generally, and shared a Red Hat Bugzilla entry (BZ#2468413) showing RPM-build-related skills in ClawHub, suggesting Fedora doesn’t need to reinvent the wheel.
@jflory7 mentioned he has been exploring agr, which is already packaged in Fedora. The package turns out to be maintained by Ankur Sinha (@ankursinha).
Ticket: ai-ml/tickets#34 (Scope gpu01 Access)
@jflory7 flagged that he had significantly edited #34 since the previous meeting, changing both the title and body, and asked for feedback rather than a decision.
@tflink and @trix both stated that it is too early for this discussion. Both had also left comments on the ticket to this effect. @tflink said there are too many moving parts to make decisions. @trix noted that basic questions around machine ownership and whether the hardware is a loaner remain unresolved.
@jflory7 acknowledged the timing issue but pointed out that ticket #33 (a general GPU access request that the reporter closed on their own) showed that people are already assuming the hardware is available through the SIG. He proposed framing the situation as a “transitionary/exploration” phase to set expectations.
@gordonmessmer made a notable point: the GPU systems need an Acceptable Use Policy regardless of how jobs are run. “CI” describes a mechanism, not a purpose; batch jobs can do anything, so the policy needs to define acceptable purposes.
@trix raised trust as a concern and said open access should not be on the table. @tflink mentioned that Fedora Infrastructure has run packager-checkout-able systems for exotic hardware (s390x, PPC) in the past, so this is not entirely new ground.
The group agreed on two short-term resolutions:
-
Treat the available hardware as being in a transitionary/exploration phase focused on CI for the F45 release cycle, closing interim access ticket requests as
state/wontfix -
Follow up at the August 13th meeting if there is no new clarity by then. @jflory7 will open a separate ticket for the AUP discussion.
Open Floor
No additional topics were raised. @jflory7 thanked attendees and closed the meeting.
Wow, this is a rare time when I hear how Fedora is too slow with dependencies in our releases to keep up! ↩︎