Try AI training with ROCm and LoRA by bmwiedemann

Description

I want to setup a Radeon RX 9600 XT 16 GB at home with ROCm on Slowroll.

Goals

I want to test how fast AI inference can get with the GPU and if I can use LoRA to re-train an existing free model for some task.

Resources

• https://rocm.docs.amd.com/en/latest/compatibility/compatibility-matrix.html
• https://build.opensuse.org/project/show/science:GPU:ROCm
• https://src.opensuse.org/ROCm/
• https://www.suse.com/c/lora-fine-tuning-llms-for-text-classification/

Results

got inference working with llama.cpp:

export LLAMACPP_ROCM_ARCH=gfx1200
HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$LLAMACPP_ROCM_ARCH \
-DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON \
-Dhipblas_DIR=/usr/lib64/cmake/hipblaslt/ \
&& cmake --build build --config Release -j8
m=models/gpt-oss-20b-mxfp4.gguf
cd $P/llama.cpp && build/bin/llama-server --model $m --threads 8 --port 8005 --host 0.0.0.0 --device ROCm0 --n-gpu-layers 999

Without the --device option it faulted. Maybe because my APU also appears there?

I updated/fixed various related packages: https://src.opensuse.org/ROCm/rocm-examples/pulls/1 https://src.opensuse.org/ROCm/hipblaslt/pulls/1 SR 1320959

benchmark

I benchmarked inference with llama.cpp + gpt-oss-20b-mxfp4.gguf and ROCm offloading to a Radeon RX 9060 XT 16GB. I varied the number of layers that went to the GPU:

• 0 layers 14.49 tokens/s (8 CPU cores)
• 9 layers 17.79 tokens/s 34% VRAM
• 15 layers 22.39 tokens/s 51% VRAM
• 20 layers 27.49 tokens/s 64% VRAM
• 24 layers 41.18 tokens/s 74% VRAM
• 25+ layers 86.63 tokens/s 75% VRAM (only 200% CPU load)

So there is a significant performance-boost if the whole model fits into the GPU's VRAM.

---

issuefs: FUSE filesystem representing issues (e.g. JIRA) for the use with AI agents code-assistants by llansky3

Description

Creating a FUSE filesystem (issuefs) that mounts issues from various ticketing systems (Github, Jira, Bugzilla, Redmine) as files to your local file system.

And why this is good idea?

• User can use favorite command line tools to view and search the tickets from various sources
• User can use AI agents capabilities from your favorite IDE or cli to ask question about the issues, project or functionality while providing relevant tickets as context without extra work.
• User can use it during development of the new features when you let the AI agent to jump start the solution. The issuefs will give the AI agent the context (AI agents just read few more files) about the bug or requested features. No need for copying and pasting issues to user prompt or by using extra MCP tools to access the issues. These you can still do but this approach is on purpose different.

Goals

1. Add Github issue support
2. Proof the concept/approach by apply the approach on itself using Github issues for tracking and development of new features
3. Add support for Bugzilla and Redmine using this approach in the process of doing it. Record a video of it.
4. Clean-up and test the implementation and create some documentation
5. Create a blog post about this approach

Resources

There is a prototype implementation here. This currently sort of works with JIRA only.

---

Song Search with CLAP by gcolangiuli

Description

Contrastive Language-Audio Pretraining (CLAP) is an open-source library that enables the training of a neural network on both Audio and Text descriptions, making it possible to search for Audio using a Text input. Several pre-trained models for song search are already available on huggingface

Goals

Evaluate how CLAP can be used for song searching and determine which types of queries yield the best results by developing a Minimum Viable Product (MVP) in Python. Based on the results of this MVP, future steps could include:

• Music Tagging;
• Free text search;
• Integration with an LLM (for example, with MCP or the OpenAI API) for music suggestions based on your own library.

The code for this project will be entirely written using AI to better explore and demonstrate AI capabilities.

Result

In this MVP we implemented:

• Async Song Analysis with Clap model
• Free Text Search of the songs
• Similar song search based on vector representation
• Containerised version with web interface

We also documented what went well and what can be improved in the use of AI.

You can have a look at the result here:

• Result Presentation
• hw25-song-search Github repo

Future implementation can be related to performance improvement and stability of the analysis.

References

• CLAP: The main model being researched;
• huggingface: Pre-trained models for CLAP;
• Free Music Archive: Creative Commons songs that can be used for testing;

---

Multi-agent AI assistant for Linux troubleshooting by doreilly

Description

Explore multi-agent architecture as a way to avoid MCP context rot.

Having one agent with many tools bloats the context with low-level details about tool descriptions, parameter schemas etc which hurts LLM performance. Instead have many specialised agents, each with just the tools it needs for its role. A top level supervisor agent takes the user prompt and delegates to appropriate sub-agents.

Goals

Create an AI assistant with some sub-agents that are specialists at troubleshooting Linux subsystems, e.g. systemd, selinux, firewalld etc. The agents can get information from the system by implementing their own tools with simple function calls, or use tools from MCP servers, e.g. a systemd-agent can use tools from systemd-mcp.

Example prompts/responses:

user$ the system seems slow
assistant$ process foo with pid 12345 is using 1000% cpu ...

user$ I can't connect to the apache webserver
assistant$ the firewall is blocking http ... you can open the port with firewall-cmd --add-port ...

Resources

Language Python. The Python ADK is more mature than Golang.

https://google.github.io/adk-docs/

https://github.com/djoreilly/linux-helper

---

MCP Server for SCC by digitaltomm

Description

Provide an MCP Server implementation for customers to access data on scc.suse.com via MCP protocol. The core benefit of this MCP interface is that it has direct (read) access to customer data in SCC, so the AI agent gets enhanced knowledge about individual customer data, like subscriptions, orders and registered systems.

Architecture

Goals

We want to demonstrate a proof of concept to connect to the SCC MCP server with any AI agent, for example gemini-cli or codex. Enabling the user to ask questions regarding their SCC inventory.

For this Hackweek, we target that users get proper responses to these example questions:

• Which of my currently active systems are running products that are out of support?
• Do I have ready to use registration codes for SLES?
• What are the latest 5 released patches for SLES 15 SP6? Output as a list with release date, patch name, affected package names and fixed CVEs.
• Which versions of kernel-default are available on SLES 15 SP6?

Technical Notes

Similar to the organization APIs, this can expose to customers data about their subscriptions, orders, systems and products. Authentication should be done by organization credentials, similar to what needs to be provided to RMT/MLM. Customers can connect to the SCC MCP server from their own MCP-compatible client and Large Language Model (LLM), so no third party is involved.

Milestones

[x] Basic MCP API setup
  MCP endpoints
  [x] Products / Repositories
  [x] Subscriptions / Orders 
  [x] Systems
  [x] Packages
[x] Document usage with Gemini CLI, Codex

Resources

• fast_mcp library for Rails
• MCP test client
• SCC branch

Gemini CLI setup:

~/.gemini/settings.json:

---

Self-Scaling LLM Infrastructure Powered by Rancher by ademicev0

Self-Scaling LLM Infrastructure Powered by Rancher

---

Description

The Problem

Running LLMs can get expensive and complex pretty quickly.

Today there are typically two choices:

1. Use cloud APIs like OpenAI or Anthropic. Easy to start with, but costs add up at scale.
2. Self-host everything - set up Kubernetes, figure out GPU scheduling, handle scaling, manage model serving... it's a lot of work.

What if there was a middle ground?

What if infrastructure scaled itself instead of making you scale it?

Can we use existing Rancher capabilities like CAPI, autoscaling, and GitOps to make this simpler instead of building everything from scratch?

Project Repository: github.com/alexander-demicev/llmserverless

---

What This Project Does

A key feature is hybrid deployment: requests can be routed based on complexity or privacy needs. Simple or low-sensitivity queries can use public APIs (like OpenAI), while complex or private requests are handled in-house on local infrastructure. This flexibility allows balancing cost, privacy, and performance - using cloud for routine tasks and on-premises resources for sensitive or demanding workloads.

A complete, self-scaling LLM infrastructure that:

• Scales to zero when idle (no idle costs)
• Scales up automatically when requests come in
• Adds more nodes when needed, removes them when demand drops
• Runs on any infrastructure - laptop, bare metal, or cloud

Think of it as "serverless for LLMs" - focus on building, the infrastructure handles itself.

How It Works

A combination of open source tools working together:

Flow:

• Users interact with OpenWebUI (chat interface)
• Requests go to LiteLLM Gateway
• LiteLLM routes requests to:
  • Ollama (Knative) for local model inference (auto-scales pods)
  • Or cloud APIs for fallback

---

issuefs: FUSE filesystem representing issues (e.g. JIRA) for the use with AI agents code-assistants by llansky3

Description

Creating a FUSE filesystem (issuefs) that mounts issues from various ticketing systems (Github, Jira, Bugzilla, Redmine) as files to your local file system.

And why this is good idea?

• User can use favorite command line tools to view and search the tickets from various sources
• User can use AI agents capabilities from your favorite IDE or cli to ask question about the issues, project or functionality while providing relevant tickets as context without extra work.
• User can use it during development of the new features when you let the AI agent to jump start the solution. The issuefs will give the AI agent the context (AI agents just read few more files) about the bug or requested features. No need for copying and pasting issues to user prompt or by using extra MCP tools to access the issues. These you can still do but this approach is on purpose different.

Goals

1. Add Github issue support
2. Proof the concept/approach by apply the approach on itself using Github issues for tracking and development of new features
3. Add support for Bugzilla and Redmine using this approach in the process of doing it. Record a video of it.
4. Clean-up and test the implementation and create some documentation
5. Create a blog post about this approach

Resources

There is a prototype implementation here. This currently sort of works with JIRA only.

---

Bugzilla goes AI - Phase 1 by nwalter

Description

This project, Bugzilla goes AI, aims to boost developer productivity by creating an autonomous AI bug agent during Hackweek. The primary goal is to reduce the time employees spend triaging bugs by integrating Ollama to summarize issues, recommend next steps, and push focused daily reports to a Web Interface.

Goals

To reduce employee time spent on Bugzilla by implementing an AI tool that triages and summarizes bug reports, providing actionable recommendations to the team via Web Interface.

Project Charter

Bugzilla goes AI Phase 1

Description

Project Achievements during Hackweek

In this file you can read about what we achieved during Hackweek.

Project Achievements

---

Extended private brain - RAG my own scripts and data into offline LLM AI by tjyrinki_suse

Description

For purely studying purposes, I'd like to find out if I could teach an LLM some of my own accumulated knowledge, to use it as a sort of extended brain.

I might use qwen3-coder or something similar as a starting point.

Everything would be done 100% offline without network available to the container, since I prefer to see when network is needed, and make it so it's never needed (other than initial downloads).

Goals

1. Learn something about RAG, LLM, AI.
2. Find out if everything works offline as intended.
3. As an end result have a new way to access my own existing know-how, but so that I can query the wisdom in them.
4. Be flexible to pivot in any direction, as long as there are new things learned.

Resources

To be found on the fly.

Timeline

Day 1 (of 4)

• Tried out a RAG demo, expanded on feeding it my own data
• Experimented with qwen3-coder to add a persistent chat functionality, and keeping vectors in a pickle file
• Optimizations to keep everything within context window
• Learn and add a bit of PyTest

Day 2

• More experimenting and more data
• Study ChromaDB
• Add a Web UI that works from another computer even though the container sees network is down

Day 3

• The above RAG is working well enough for demonstration purposes.
• Pivot to trying out OpenCode, configuring local Ollama qwen3-coder there, to analyze the RAG demo.
• Figured out how to configure Ollama template to be usable under OpenCode. OpenCode locally is super slow to just running qwen3-coder alone.

Day 4 (final day)

• Battle with OpenCode that was both slow and kept on piling up broken things.
• Call it success as after all the agentic AI was working locally.
• Clean up the mess left behind a bit.

Blog Post

Summarized the findings at blog post.

---

SUSE Observability MCP server by drutigliano

Description

The idea is to implement the SUSE Observability Model Context Protocol (MCP) Server as a specialized, middle-tier API designed to translate the complex, high-cardinality observability data from StackState (topology, metrics, and events) into highly structured, contextually rich, and LLM-ready snippets.

This MCP Server abstract the StackState APIs. Its primary function is to serve as a Tool/Function Calling target for AI agents. When an AI receives an alert or a user query (e.g., "What caused the outage?"), the AI calls an MCP Server endpoint. The server then fetches the relevant operational facts, summarizes them, normalizes technical identifiers (like URNs and raw metric names) into natural language concepts, and returns a concise JSON or YAML payload. This payload is then injected directly into the LLM's prompt, ensuring the final diagnosis or action is grounded in real-time, accurate SUSE Observability data, effectively minimizing hallucinations.

Goals

• Grounding AI Responses: Ensure that all AI diagnoses, root cause analyses, and action recommendations are strictly based on verifiable, real-time data retrieved from the SUSE Observability StackState platform.
• Simplifying Data Access: Abstract the complexity of StackState's native APIs (e.g., Time Travel, 4T Data Model) into simple, semantic functions that can be easily invoked by LLM tool-calling mechanisms.
• Data Normalization: Convert complex, technical identifiers (like component URNs, raw metric names, and proprietary health states) into standardized, natural language terms that an LLM can easily reason over.
• Enabling Automated Remediation: Define clear, action-oriented MCP endpoints (e.g., execute_runbook) that allow the AI agent to initiate automated operational workflows (e.g., restarts, scaling) after a diagnosis, closing the loop on observability.

 Hackweek STEP

• Create a functional MCP endpoint exposing one (or more) tool(s) to answer queries like "What is the health of service X?") by fetching, normalizing, and returning live StackState data in an LLM-ready format.

 Scope

• Implement read-only MCP server that can:
  • Connect to a live SUSE Observability instance and authenticate (with API token)
  • Use tools to fetch data for a specific component URN (e.g., current health state, metrics, possibly topology neighbors, ...).
  • Normalize response fields (e.g., URN to "Service Name," health state DEVIATING to "Unhealthy", raw metrics).
  • Return the data as a structured JSON payload compliant with the MCP specification.

Deliverables

• MCP Server v0.1 A running Golang MCP server with at least one tool.
• A README.md and a test script (e.g., curl commands or a simple notebook) showing how an AI agent would call the endpoint and the resulting JSON payload.

Outcome A functional and testable API endpoint that proves the core concept: translating complex StackState data into a simple, LLM-ready format. This provides the foundation for developing AI-driven diagnostics and automated remediation.

Resources

• https://www.honeycomb.io/blog/its-the-end-of-observability-as-we-know-it-and-i-feel-fine
• https://www.datadoghq.com/blog/datadog-remote-mcp-server
• https://modelcontextprotocol.io/specification/2025-06-18/index
• https://modelcontextprotocol.io/docs/develop/build-server

 Basic implementation

• https://github.com/drutigliano19/suse-observability-mcp-server

---

Results

Successfully developed and delivered a fully functional SUSE Observability MCP Server that bridges language models with SUSE Observability's operational data. This project demonstrates how AI agents can perform intelligent troubleshooting and root cause analysis using structured access to real-time infrastructure data.

Example execution

---

Improve UML page fault handler by ptesarik

Description

Improve UML handling of segmentation faults in kernel mode. Although such page faults are generally caused by a kernel bug, it is annoying if they cause an infinite loop, or panic the kernel. More importantly, a robust implementation allows to write KUnit tests for various guard pages, preventing potential kernel self-protection regressions.

Goals

Convert the UML page fault handler to use oops_* helpers, go through a few review rounds and finally get my patch series merged in 6.14.

Resources

Wrong initial attempt: https://lore.kernel.org/lkml/20231215121431.680-1-petrtesarik@huaweicloud.com/T/

---

early stage kdump support by mbrugger

Project Description

When we experience a early boot crash, we are not able to analyze the kernel dump, as user-space wasn't able to load the crash system. The idea is to make the crash system compiled into the host kernel (think of initramfs) so that we can create a kernel dump really early in the boot process.

Goal for the Hackweeks

1. Investigate if this is possible and the implications it would have (done in HW21)
2. Hack up a PoC (done in HW22 and HW23)
3. Prepare RFC series (giving it's only one week, we are entering wishful thinking territory here).

update HW23

• I was able to include the crash kernel into the kernel Image.
• I'll need to find a way to load that from init/main.c:start_kernel() probably after kcsan_init()
• I workaround for a smoke test was to hack kexec_file_load() systemcall which has two problems:
  1. My initramfs in the porduction kernel does not have a new enough kexec version, that's not a blocker but where the week ended
  2. As the crash kernel is part of init.data it will be already stale once I can call kexec_file_load() from user-space.

The solution is probably to rewrite the POC so that the invocation can be done from init.text (that's my theory) but I'm not sure if I can reuse the kexec infrastructure in the kernel from there, which I rely on heavily.

update HW24

• Day1
  • rebased on v6.12 with no problems others then me breaking the config
  • setting up a new compilation and qemu/virtme env
  • getting desperate as nothing works that used to work
• Day 2
  • getting to call the invocation of loading the early kernel from __init after kcsan_init()

• Day 3

  • fix problem of memdup not being able to alloc so much memory... use 64K page sizes for now
  • code refactoring
  • I'm now able to load the crash kernel
  • When using virtme I can boot into the crash kernel, also it doesn't boot completely (major milestone!), crash in elfcorehdr_read_notes()

• Day 4

  • crash systems crashes (no pun intended) in copy_old_mempage() link; will need to understand elfcorehdr...
  • call path vmcore_init() -> parse_crash_elf_headers() -> elfcorehdr_read() -> read_from_oldmem() -> copy_oldmem_page() -> copy_to_iter()

• Day 5

  • hacking arch/arm64/kernel/crash_dump.c:copy_old_mempage() to see if crash system really starts. It does.
  • fun fact: retested with more reserved memory and with UEFI FW, host kernel crashes in init but directly starts the crash kernel, so it works (somehow) \o/

update HW25

• Day 1
  • rebased crash-kernel on v6.12.59 (for now), still crashing

---

Add Qualcomm Snapdragon 765G (SM7250) basic device tree to mainline linux kernel by pvorel

Qualcomm Snapdragon 765G (SM7250) (smartphone SoC) has no support in the linux kernel, nor in u-boot. Try to add basic device tree support. The hardest part will be to create boot.img which will be accepted by phone.

UART is available for smartphone :).

---

bpftrace contribution by mkoutny

Description

bpftrace is a great tool, no need to sing odes to it here. It can access any kernel data and process them in real time. It provides helpers for some common Linux kernel structures but not all.

Goals

• set up bpftrace toolchain
• learn about bpftrace implementation and internals
• implement support for percpu_counters
• look into some of the first issues
• send a refined PR (on Thu)

Resources

• https://bpftrace.org/
• https://docs.kernel.org/

---

pudc - A PID 1 process that barks to the internet by mssola

Description

As a fun exercise in order to dig deeper into the Linux kernel, its interfaces, the RISC-V architecture, and all the dragons in between; I'm building a blog site cooked like this:

• The backend is written in a mixture of C and RISC-V assembly.
• The backend is actually PID1 (for real, not within a container).
• We poll and parse incoming HTTP requests ourselves.
• The frontend is a mere HTML page with htmx.

The project is meant to be Linux-specific, so I'm going to use io_uring, pidfs, namespaces, and Linux-specific features in order to drive all of this.

I'm open for suggestions and so on, but this is meant to be a solo project, as this is more of a learning exercise for me than anything else.

Goals

• Have a better understanding of different Linux features from user space down to the kernel internals.
• Most importantly: have fun.

Resources

• https://github.com/mssola/pudc

---