Running Ollama as a k8s STS with external script as entrypoint to load models

At a mechanical level, the backslashes inside the for loop are causing problems. This causes the shell to combine the lines together, so you get a single command echo model: $i ollama pull $i done, but there's not a standalone done command to terminate the loop.

The next problem you'll run into is that this entrypoint script is the only thing the container runs, and when this script exits, the container will exit as well. It doesn't matter that you've started the Ollama server in the background. If you wanted to run the container this way, you need to wait for the server to exit. That would look something like

#!/bin/bash
ollama serve &
pid=$!                       # ADD: save the process ID of the server
sleep 10
models="${MODELS//,/ }"
for i in "${models[@]}"; do  # FIX: remove backslashes
  echo model: "$i"
  ollama pull "$i"
done
wait "$pid"                  # ADD: keep the script running as long as the server is too

However, this model of starting a background process and then waiting for it often isn't the best approach. If the Pod gets shut down, for example, the termination signal will go to the wrapper script and not the Ollama server, and you won'd be able to have a clean shutdown.

In a Kubernetes context (you say you're running this in a StatefulSet) a PostStart hook fits here. This will let you run an unmodified image, but add your own script that runs at about the same time as the container startup. In a Kubernetes manifest this might look like:

spec:
  template:
    spec:
      containers:
        - name: ollama
          image: ollama/ollama  # the unmodified upstream image
          lifecycle:
            postStart:
              exec:
                command:
                  - /bin/sh
                  - -c
                  - |
                      for i in $(seq 10); do
                        ollama ps && break
                        sleep 1
                      done
                      for model in llama3.2; do
                        ollama pull "$model"
                      done

This setup writes a shell script inline in the Kubernetes manifest. It wraps it in /bin/sh -c to it can be run this way. This uses an "exec" mechanism, so the script runs as a secondary process in the same container. The first fragment waits up to 10 seconds for the server to be running, and the second is the loop to load the models.