GETTING ODYSSEUS RUNNING WITH GPU ACCELERATION ON THE MINISFORUM AI X1 PRO (RADEON 890M)

[Ojisan1]

GETTING ODYSSEUS RUNNING WITH GPU ACCELERATION ON THE MINISFORUM AI X1 PRO (RADEON 890M)

Community guide - June 2026

The Radeon 890M integrated GPU (found in the Minisforum AI X1 Pro and similar Strix Point-based machines) currently has incomplete ROCm support. Many users have found that llama.cpp + Vulkan provides more reliable GPU acceleration than trying to force ROCm or Ollama on this hardware.

This post outlines a working Docker-based setup that gets Odysseus running with proper iGPU acceleration using models like Qwen3-Coder-Next with GPU layers set to 99.

A more detailed technical spec with exact commands, verification steps, and a deviations log is attached to this post as a separate Markdown file.
Odysseus_Radeon_890M_Vulkan_Detailed_Setup_Spec.md

================================================================================
HARDWARE & BIOS RECOMMENDATIONS

• Machine: Minisforum AI X1 Pro (Ryzen AI 9 HX 370 + Radeon 890M iGPU)
• RAM: 96 GB recommended (other amounts should work)
• BIOS Setting: Set UMA Frame Buffer Size to 8 GB or 16 GB (do not go to 32 GB+) VERY IMPORTANT

This shared memory configuration gives the iGPU enough VRAM while leaving plenty for the system.

================================================================================
HIGH-LEVEL APPROACH

• Run Odysseus in Docker (recommended for reproducibility)
• Install Vulkan development tools + Mesa drivers inside the container
• Manually build llama.cpp with GGML_VULKAN=ON (the Cookbook "Rebuild" button currently does not target Vulkan)
• Pass through /dev/dri and the render group so Vulkan can see the iGPU
• Use the resulting llama-server binary for serving models in the Cookbook

================================================================================
STEP-BY-STEP OUTLINE

--- STEP 1: PREPARE THE HOST ---

Make sure your user is in the render group:

sudo usermod -aG render $USER
newgrp render

Check your render GID (you will need this value later):

getent group render | cut -d: -f3

--- STEP 2: SET UP ODYSSEUS ---

Clone the repository and enter the directory:

git clone https://github.com/pewdiepie-archdaemon/odysseus.git
cd odysseus

--- STEP 3: EDIT THE DOCKERFILE (CRITICAL STEP) ---

Add the following Vulkan-related packages to the existing apt-get install block. Merge them into the existing line - do not create a duplicate RUN instruction:

libvulkan-dev \
vulkan-tools \
libshaderc-dev \
glslang-tools \
spirv-tools \
glslc \
mesa-vulkan-drivers \

These packages ensure glslc and proper AMD RADV Vulkan drivers are available inside the container.

--- STEP 4: CONFIGURE DOCKER COMPOSE ---

You have two good options:

Option A (Recommended): Use COMPOSE_FILE in .env

echo 'COMPOSE_FILE=docker-compose.yml:docker/gpu.amd.yml' >> .env

Option B: Use the standalone file directly

docker compose -f docker-compose.gpu-amd.yml up -d

Make sure the searxng service uses host port 9000 (to avoid conflicts with other containers).

Also set your render GID in .env:

RENDER_GID=992     # Change this to match your host's render group GID

--- STEP 5: BUILD THE IMAGE (CLEAN) ---

docker compose build --no-cache odysseus

--- STEP 6: START THE STACK ---

docker compose up -d

--- STEP 7: MANUALLY BUILD LLAMA.CPP WITH VULKAN SUPPORT ---

This is the most important step for getting GPU acceleration:

docker compose exec -u 1000:1000 odysseus bash -lc '
set -euo pipefail
export HOME=/app
export PATH="/app/bin:/app/.local/bin:$PATH"
mkdir -p /app/bin /app/.local/bin

if [ ! -d /app/llama.cpp ]; then
  git clone --depth 1 https://github.com/ggml-org/llama.cpp /app/llama.cpp
fi

cd /app/llama.cpp
rm -rf build
cmake -B build \
  -DGGML_VULKAN=ON \
  -DLLAMA_BUILD_SERVER=ON \
  -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release -j"$(nproc)" --target llama-server

ln -sf /app/llama.cpp/build/bin/llama-server /app/bin/llama-server
ln -sf /app/llama.cpp/build/bin/llama-server /app/.local/bin/llama-server

echo "llama-server built with Vulkan support"
'

--- STEP 8: VERIFY THE BUILD ---

Check that Vulkan is properly linked:

docker compose exec odysseus bash -lc \
  'ldd /app/llama.cpp/build/bin/llama-server | grep -E "vulkan|libvulkan"'

You should see libggml-vulkan.so and libvulkan.so.

Also verify the iGPU is visible inside the container:

docker compose exec odysseus bash -lc \
  'vulkaninfo --summary 2>&1 | grep -E "deviceName|RADV"'

You should see something like: AMD Radeon Graphics (RADV ...)

--- STEP 9: LAUNCH A MODEL IN THE UI ---

1. Open http://localhost:7000
2. Go to Cookbook -> Serve
3. Select your GGUF model (for example Qwen3-Coder-Next)
4. Set:
   • GPU Layers: 99
   • Context: 16384-32768 (adjust based on your available RAM)
5. Launch the model

While generating, you can monitor real GPU usage on the host with:

radeontop

================================================================================
IMPORTANT CAVEATS

• The Cookbook "Rebuild" button currently rebuilds using HIP/CUDA/CPU logic and will overwrite your Vulkan build. Always re-run the manual build in Step 7 after using Rebuild.

• The first manual llama.cpp build can take 1 to 3 minutes.

• Make sure you pass both /dev/dri and the render group correctly (this is handled by the docker/gpu.amd.yml overlay).

• mesa-vulkan-drivers MUST be installed inside the container. Without it, vulkaninfo will report no drivers even if the device is passed through.

• This setup is specific to llama.cpp + Vulkan. It does not use ROCm or Ollama's backend.

================================================================================
VERIFICATION CHECKLIST

After launching a model with GPU layers enabled, check the following:

• vulkaninfo --summary inside the container shows "AMD Radeon Graphics (RADV ...)"
• ldd on llama-server shows libggml-vulkan.so
• No warning message saying "no usable GPU found"
• Reasonable tokens per second for the model size (typically 15-40+ t/s depending on quantization and context)
• radeontop shows GPU activity during generation
• The model responds correctly in chat

================================================================================
FULL DETAILED SPEC

A more comprehensive technical document (including exact file edits, full verification commands, and a deviations log from a successful run) is attached to this post.

The attached file is named:
Odysseus_Radeon_890M_Vulkan_Detailed_Setup_Spec.md

That document is the single source of truth for this configuration.

================================================================================
QUESTIONS & FEEDBACK

If you try this setup on a Minisforum AI X1 Pro or similar machine, feel free to reply with your results. I hope others find this useful!

Improvements to the upstream Odysseus Dockerfile (better Vulkan support) or Cookbook build logic would be very welcome contributions to the project.

This is a community-tested configuration, not official documentation. Always review changes before applying them to your system.

[ktzee]

Thank you for this. Would this approach also allow me to use my 7900XTX on my desktop? I'm having issues getting GPU acceleration on my AMD machine too. Ollama-rocm outside of Odysseus accelerates correctly, but I can't get Odysseus to do the same thing.
Also, what's the difference between using vulkan and using rocm as a backend?