Try AI training with ROCm and LoRA
a project by bmwiedemann
Updated 14 days ago. 3 hacker ♥️. 1 follower.

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.

training data

I collected training data from my bugzilla archive: https://www.zq1.de/~bernhard/linux/opensuse/bugzilla/bugzillapkgdata.json.xz This contains one JSON per line for summary, description(body) and packages(pkgs) that got fixes. The extraction code lives in https://github.com/bmwiedemann/bugzillai

Join this project

Looking for hackers with the skills:

ai training rocm

This project is part of:

Hack Week 25

Activity

  • 18 days ago: pgonin liked this project.
  • 18 days ago: mkoutny liked this project.
  • 19 days ago: bmwiedemann started this project.
  • 19 days ago: bmwiedemann liked this project.
  • 20 days ago: bmwiedemann added keyword "rocm" to this project.
  • 20 days ago: bmwiedemann added keyword "ai" to this project.
  • 20 days ago: bmwiedemann added keyword "training" to this project.
  • 20 days ago: bmwiedemann originated this project.


    • Comments

    • bmwiedemann
      7 days ago by bmwiedemann | Reply

      collected training data: https://www.zq1.de/~bernhard/linux/opensuse/bugzilla/

    Similar Projects

    AI-Powered Unit Test Automation for Agama by joseivanlopez

    The Agama project is a multi-language Linux installer that leverages the distinct strengths of several key technologies:

    • Rust: Used for the back-end services and the core HTTP API, providing performance and safety.
    • TypeScript (React/PatternFly): Powers the modern web user interface (UI), ensuring a consistent and responsive user experience.
    • Ruby: Integrates existing, robust YaST libraries (e.g., yast-storage-ng) to reuse established functionality.

    The Problem: Testing Overhead

    Developing and maintaining code across these three languages requires a significant, tedious effort in writing, reviewing, and updating unit tests for each component. This high cost of testing is a drain on developer resources and can slow down the project's evolution.

    The Solution: AI-Driven Automation

    This project aims to eliminate the manual overhead of unit testing by exploring and integrating AI-driven code generation tools. We will investigate how AI can:

    1. Automatically generate new unit tests as code is developed.
    2. Intelligently correct and update existing unit tests when the application code changes.

    By automating this crucial but monotonous task, we can free developers to focus on feature implementation and significantly improve the speed and maintainability of the Agama codebase.

    Goals

    • Proof of Concept: Successfully integrate and demonstrate an authorized AI tool (e.g., gemini-cli) to automatically generate unit tests.
    • Workflow Integration: Define and document a new unit test automation workflow that seamlessly integrates the selected AI tool into the existing Agama development pipeline.
    • Knowledge Sharing: Establish a set of best practices for using AI in code generation, sharing the learned expertise with the broader team.

    Contribution & Resources

    We are seeking contributors interested in AI-powered development and improving developer efficiency. Whether you have previous experience with code generation tools or are eager to learn, your participation is highly valuable.

    If you want to dive deep into AI for software quality, please reach out and join the effort!

    • Authorized AI Tools: Tools supported by SUSE (e.g., gemini-cli)
    • Focus Areas: Rust, TypeScript, and Ruby components within the Agama project.

    Interesting Links

    Local AI assistant with optional integrations and mobile companion by livdywan

    Description

    Setup a local AI assistant for research, brainstorming and proof reading. Look into SurfSense, Open WebUI and possibly alternatives. Explore integration with services like openQA. There should be no cloud dependencies. Mobile phone support or an additional companion app would be a bonus. The goal is not to develop everything from scratch.

    User Story

    • Allison Average wants a one-click local AI assistent on their openSUSE laptop.
    • Ash Awesome wants AI on their phone without an expensive subscription.

    Goals

    • Evaluate a local SurfSense setup for day to day productivity
    • Test opencode for vibe coding and tool calling

    Timeline

    Day 1

    • Took a look at SurfSense and started setting up a local instance.
    • Unfortunately the container setup did not work well. Tho this was a great opportunity to learn some new podman commands and refresh my memory on how to recover a corrupted btrfs filesystem.

    Day 2

    • Due to its sheer size and complexity SurfSense seems to have triggered btrfs fragmentation. Naturally this was not visible in any podman-related errors or in the journal. So this took up much of my second day.

    Day 3

    Day 4

    • Context size is a thing, and models are not equally usable for vibe coding.
    • Through arduous browsing for ollama models I did find some like myaniu/qwen2.5-1m:7b with 1m but even then it is not obvious if they are meant for tool calls.

    Day 5

    • Whilst trying to make opencode usable I discovered ramalama which worked instantly and very well.

    Outcomes

    surfsense

    I could not easily set this up completely. Maybe in part due to my filesystem issues. Was expecting this to be less of an effort.

    opencode

    Installing opencode and ollama in my distrobox container along with the following configs worked for me.

    When preparing a new project from scratch it is a good idea to start out with a template.

    opencode.json

    ``` {

    SUSE Edge Image Builder MCP by eminguez

    Description

    Based on my other hackweek project, SUSE Edge Image Builder's Json Schema I would like to build also a MCP to be able to generate EIB config files the AI way.

    Realistically I don't think I'll be able to have something consumable at the end of this hackweek but at least I would like to start exploring MCPs, the difference between an API and MCP, etc.

    Goals

    • Familiarize myself with MCPs
    • Unrealistic: Have an MCP that can generate an EIB config file

    Resources

    Result

    https://github.com/e-minguez/eib-mcp

    I've extensively used antigravity and its agent mode to code this. This heavily uses https://hackweek.opensuse.org/25/projects/suse-edge-image-builder-json-schema for the MCP to be built.

    I've ended up learning a lot of things about "prompting", json schemas in general, some golang, MCPs and AI in general :)

    Example:

    Generate an Edge Image Builder configuration for an ISO image based on slmicro-6.2.iso, targeting x86_64 architecture. The output name should be 'my-edge-image' and it should install to /dev/sda. It should deploy a 3 nodes kubernetes cluster with nodes names "node1", "node2" and "node3" as: * hostname: node1, IP: 1.1.1.1, role: initializer * hostname: node2, IP: 1.1.1.2, role: agent * hostname: node3, IP: 1.1.1.3, role: agent The kubernetes version should be k3s 1.33.4-k3s1 and it should deploy a cert-manager helm chart (the latest one available according to https://cert-manager.io/docs/installation/helm/). It should create a user called "suse" with password "suse" and set ntp to "foo.ntp.org". The VIP address for the API should be 1.2.3.4

    Generates:

    ``` apiVersion: "1.0" image: arch: x86_64 baseImage: slmicro-6.2.iso imageType: iso outputImageName: my-edge-image kubernetes: helm: charts: - name: cert-manager repositoryName: jetstack

    Liz - Prompt autocomplete by ftorchia

    Description

    Liz is the Rancher AI assistant for cluster operations.

    Goals

    We want to help users when sending new messages to Liz, by adding an autocomplete feature to complete their requests based on the context.

    Example:

    • User prompt: "Can you show me the list of p"
    • Autocomplete suggestion: "Can you show me the list of p...od in local cluster?"

    Example:

    • User prompt: "Show me the logs of #rancher-"
    • Chat console: It shows a drop-down widget, next to the # character, with the list of available pod names starting with "rancher-".

    Technical Overview

    1. The AI agent should expose a new ws/autocomplete endpoint to proxy autocomplete messages to the LLM.
    2. The UI extension should be able to display prompt suggestions and allow users to apply the autocomplete to the Prompt via keyboard shortcuts.

    Resources

    GitHub repository

    The Agentic Rancher Experiment: Do Androids Dream of Electric Cattle? by moio

    Rancher is a beast of a codebase. Let's investigate if the new 2025 generation of GitHub Autonomous Coding Agents and Copilot Workspaces can actually tame it. A GitHub robot mascot trying to lasso a blue bull with a Kubernetes logo tatooed on it

    The Plan

    Create a sandbox GitHub Organization, clone in key Rancher repositories, and let the AI loose to see if it can handle real-world enterprise OSS maintenance - or if it just hallucinates new breeds of Kubernetes resources!

    Specifically, throw "Agentic Coders" some typical tasks in a complex, long-lived open-source project, such as:


    ❥ The Grunt Work: generate missing GoDocs, unit tests, and refactorings. Rebase PRs.

    ❥ The Complex Stuff: fix actual (historical) bugs and feature requests to see if they can traverse the complexity without (too much) human hand-holding.

    ❥ Hunting Down Gaps: find areas lacking in docs, areas of improvement in code, dependency bumps, and so on.


    If time allows, also experiment with Model Context Protocol (MCP) to give agents context on our specific build pipelines and CI/CD logs.

    Why?

    We know AI can write "Hello World." and also moderately complex programs from a green field. But can it rebase a 3-month-old PR with conflicts in rancher/rancher? I want to find the breaking point of current AI agents to determine if and how they can help us to reduce our technical debt, work faster and better. At the same time, find out about pitfalls and shortcomings.

    The CONCLUSION!!!

    A add-emoji State of the Union add-emoji document was compiled to summarize lessons learned this week. For more gory details, just read on the diary below! add-emoji