Error: mkdir /root/.ollama: read-only file system

[nostrdev]

Getting the following in a fresh install after downloading the model:

Setting the HOME variable does not help.

root@4772c55c-cc53-4e23-b90e-d3d8b5687bab:/app/code# ollama run qwen3.5:35b
Error: mkdir /root/.ollama: read-only file system

[james]

Hello @nostrdev

Please use ollama pull $MODELL_NAME as documented https://docs.cloudron.io/packages/ollama#downloading-models
This correctly downloads the model to /app/data/ollama-home/.

Cloudron sets the following environment variable in the start.sh#L23:

export OLLAMA_MODELS=/app/data/ollama-home/models

Ollama is started with a supervisor service https://git.cloudron.io/packages/ollama-app/-/blob/master/supervisor/ollama.conf in which the HOME is set:

environment=HOME=/app/data/ollama-home

Thus, when running ollama from the web terminal, you are running Ollama as user root with its own environment variables.
In order to use ollama run $MODELL_NAME you will need to run it the following way:

gosu cloudron:cloudron bash -c 'HOME=/app/data/ollama-home ollama run $MODELL_NAME'

That ollama run and pull behave so different is an upstream issue.
IMHO, ollama run should also respect the set environment, but it does not and instead uses os.UserHomeDir()

[james]

Now that I thought about it, it makes sense that ollama run behaves in such a way when you look at Ollama outside the Cloudron scope.

Example, you have an office AI-Server that is running ollama serve and every user can use ollama run on their local shell connecting to the server.
So user tina will write her own ollama run chat history to /home/tina/.ollama and user tim would write to /home/tim/.ollama.
If it was unified to a server side HOME you would have to do the chat separation on the server level.
And that is what the WebUI is for.

This does also imply that when using:

gosu cloudron:cloudron bash -c 'HOME=/app/data/ollama-home ollama run $MODELL_NAME'

you will share the chat history with anyone who is also might have access to the web terminal.

[nebulon]

Yes so I think the issue stems from the fact that ollama has a client/server architecture, where the Cloudron ollama app is acting as the server component (ollama serve) while if you run ollama run for example, it starts a client talking to that ollama server. The client here tries to save some info in the user's HOME which is /root here for the client.

At least so far the package was thought of only acting as the server. This means one uses ollama in client mode on the dev environment or via agentic tools like opencode. Those would not run inside that ollama app instance, but would just connect to the ollama web api and run only the models in the app.