AMD BC-250: Ollama falls back to CPU unless OLLAMA_IGPU_ENABLE=1 is added to compose.yaml

[SwiatLinuksa]

SkillFishOS version

26.06 Aetherium

Hardware

AMD BC-250

Kernel flavor

7.0.11-skillfishos (BC-250 main)

Affected area

AI panel / Ollama

What happened?

After installing SkillFish AI using the default installation wizard, Ollama started correctly and Open WebUI worked as expected.

However, on my AMD BC-250 test system, inference was executed on the CPU instead of the integrated GPU.

I verified this using:

docker exec skillfish-ollama ollama ps

which reported:

PROCESSOR 100% CPU

At the same time:

• GPU frequency stayed around 350 MHz (idle)
• CPU usage was very high
• Responses were extremely slow.

I then edited the generated compose.yaml and added:

environment:

• OLLAMA_VULKAN=1
• OLLAMA_IGPU_ENABLE=1

to the Ollama service.

After recreating the containers:

docker compose down
docker compose up -d

the behaviour changed immediately.

Now:

docker exec skillfish-ollama ollama ps

reports:

PROCESSOR 100% GPU

and during inference:

• GPU frequency reaches approximately 2000 MHz
• CPU usage drops significantly
• inference speed improves dramatically (around 75–100 tokens/sec with Qwen3 4B).

Additional information

I reproduced this multiple times by removing and adding only:

OLLAMA_IGPU_ENABLE=1

while keeping the same hardware, model and SkillFishOS installation.

The behaviour was consistent in my tests.

This suggests that the generated compose configuration may be missing a required environment variable for AMD BC-250 GPU acceleration.

Steps to reproduce

1. Install SkillFish AI using the default installation wizard.
2. Download a model (tested with Qwen3 4B).
3. Ask any question in Open WebUI.
4. Execute:

docker exec skillfish-ollama ollama ps

5. Observe:

PROCESSOR 100% CPU

GPU frequency remains around 350 MHz.

6. Edit compose.yaml and add:

environment:

• OLLAMA_VULKAN=1
• OLLAMA_IGPU_ENABLE=1

7. Restart the stack:

docker compose down
docker compose up -d

8. Repeat the same test.

Now:

PROCESSOR 100% GPU

GPU frequency reaches approximately 2000 MHz.

Logs / output

Before:

$ docker exec skillfish-ollama ollama ps

NAME        ID              SIZE      PROCESSOR    CONTEXT
qwen3:4b    359d7dd4bcda    3.9 GB    100% CPU     8192


After:

$ docker exec skillfish-ollama ollama ps

NAME        ID              SIZE      PROCESSOR    CONTEXT
qwen3:4b    359d7dd4bcda    3.9 GB    100% GPU     8192