Ollama Operator

Yet another operator for running large language models on Kubernetes with ease. 🙀

Powered by Ollama! 🐫

While Ollama is a powerful tool for running large language models locally, and the user experience of CLI is just the same as using Docker CLI, it's not possible yet to replicate the same user experience on Kubernetes, especially when it comes to running multiple models on the same cluster with loads of resources and configurations.

That's where the Ollama Operator kicks in:

• Install the operator on your Kubernetes cluster
• Apply the needed CRDs
• Create your models
• Wait for the models to be fetched and loaded, that's it!

Thanks to the great works of lama.cpp, no more worries about Python environment, CUDA drivers.

The journey to large language models, AIGC, localized agents, 🦜🔗 Langchain and more is just a few steps away!

Features

• ✅ Abilities to run multiple models on the same cluster.
• ✅ Compatible with all Ollama models, APIs, and CLI.
• ✅ Able to run on general Kubernetes clusters, K3s clusters (Respberry Pi, TrueNAS SCALE, etc.), kind, minikube, etc. You name it!
• ✅ Easy to install, uninstall, and upgrade.
• ✅ Pull image once, share across the entire node (just like normal images).
• ✅ Easy to expose with existing Kubernetes services, ingress, etc.
• ✅ Doesn't require any additional dependencies, just Kubernetes

Getting started

Install operator

kubectl apply \
  --server-side=true \
  -f https://raw.githubusercontent.com/nekomeowww/ollama-operator/v0.10.1/dist/install.yaml

Wait for the operator to be ready

kubectl wait \
  -n ollama-operator-system \
  --for=jsonpath='{.status.readyReplicas}'=1 \
  deployment/ollama-operator-controller-manager

Deploy a model

Note

You can also use the kollama CLI natively shipped by Ollama Operator, and will be easier to interact with the operator.

Install kollama CLI:

go install github.com/nekomeowww/ollama-operator/cmd/kollama@latest

Deploy a model can be done with the following command:

kollama deploy phi --expose --node-port 30001

More information can be found at CLI

Important

Working with kind?

The default provisioned StorageClass in kind is standard, and will only work with ReadWriteOnce access mode, therefore if you would need to run the operator with kind, you should specify persistentVolume with accessMode: ReadWriteOnce in the Model CRD:

apiVersion: ollama.ayaka.io/v1
kind: Model
metadata:
  name: phi
spec:
  image: phi
  persistentVolume:
    accessMode: ReadWriteOnce

Let's create a Model CR for the model phi:

apiVersion: ollama.ayaka.io/v1
kind: Model
metadata:
  name: phi
spec:
  image: phi

Apply the Model CR to your Kubernetes cluster:

kubectl apply -f ollama-model-phi.yaml

Wait for the model to be ready:

kubectl wait --for=jsonpath='{.status.readyReplicas}'=1 deployment/ollama-model-phi

Access the model

1. Ready! Now let's forward the ports to access the model:

kubectl port-forward svc/ollama-model-phi ollama

2. Interact with the model:

ollama run phi

Full options

apiVersion: ollama.ayaka.io/v1
kind: Model
metadata:
  name: phi
spec:
  # Scale the model to 2 replicas
  replicas: 2
  # Use the model image `phi`
  image: phi
  imagePullPolicy: IfNotPresent
  storageClassName: local-path
  # If you have your own PersistentVolumeClaim created
  persistentVolumeClaim: your-pvc
  # If you need to specify the access mode for the PersistentVolume
  persistentVolume:
    accessMode: ReadWriteOnce

Supported models

Unlock the abilities to run the following models with the Ollama Operator over Kubernetes:

Tip

By the power of Modelfile backed by Ollama, you can create and bundle any of your own model. As long as it's a GGUF formatted model.

Full list of available images can be found at Ollama Library.

Warning

You should have at least 8 GB of RAM available on your node to run the 7B models, 16 GB to run the 13B models, and 32 GB to run the 33B models.

Warning

The actual size of downloaded large language models are huge by comparing to the size of general container images.

1. Fast and stable network connection is recommended to download the models.
2. Efficient storage is required to store the models if you want to run models larger than 13B.

Architecture Overview

There are two major components that the Ollama Operator will create for:

1. Model Inferencing Server: The model inferencing server is a gRPC server that runs the model and serves the model's API. It is created as a Deployment in the Kubernetes cluster.
2. Model Image Storage: The model image storage is a PersistentVolume that stores the model image. It is created as a StatefulSet along with a PersistentVolumeClaim in the Kubernetes cluster.

Note

The image that created by Modelfile of Ollama is a valid OCI format image, however, due to the incompatible contentType value, and the overall structure of the Modelfile image to the general container image, it's not possible to run the model directly with the general container runtime. Therefore a standalone service/deployment of Model Image Storage is required to be persisted on the Kubernetes cluster in order to hold and cache the previously downloaded model image.

The detailed resources it creates, and the relationships between them are shown in the following diagram:

Contributing

• Refer to the CONTRIBUTING.md for more information.
• More information can be found via the Kubebuilder Documentation

Acknowledgements

Gratefully thanks to the following projects and their authors, contributors:

• Ollama
• llama.cpp
• Kubebuilder

It is because of their hard work and contributions that this program exists.