README.md

vLLM Deployment

Example deployment of the vLLM server for OpenShift.

This deployment uses a UBI9 image defined in Containerfile and accessible at quay.io/rh-aiservices-bu/vllm-openai-ubi9 (check for the latest version available).

This image implement the OpenAI API interface for maximum compatibility with other tools. See here for more information.

A notebook example using Langchain is available here.

Installation

The default installation deploys the Mistral-7B-Instruct-v0.2 model. Although on the smaller side of LLMs, it will still require a GPU to work properly in a fast enough manner. See Advanced installation for instructions on how to change the model as well as various settings.

Automated Deployment

• Use the OpenShift GitOps/ArgoCD Application definition at gitops/vllm-app.yaml

Manual Deployment using Kustomize

After logging in to OpenShift, run the following commands:

$ oc new-project vllm
$ kustomize build https://github.com/rh-aiservices-bu/llm-on-openshift.git/llm-servers/vllm/gpu/gitops | oc apply -f -

You can also replace the github.com URL in the kustomize command with a local path instead, if this repository has been locally cloned.

Manual Deployment

Using the contents of the gitops folder, perform the following steps:

• Create PVC named vllm-models-cache with enough space to hold all the models you want to try.
• Create the Deployment using the file deployment.yaml.
• Create the Service using file service.yaml.
• If you want to expose the server outside of your OpenShift cluster, create the Route with the file route.yaml

Post Installation Tasks

As of April 18, 2024 some of the available models on Hugging Face are now gated - meaning that you will require a user access token in order for these to be downloaded at pod startup. Perform these steps to implement this:

• Sign up for an account at huggingface.co, and login
• Navigate to your user profile, settings, Access Tokens
• Create a new access token, and copy the details
• In OpenShift web console, navigate to the deployment, vllm, environment variables
• Add your token value to the HUGGING_FACE_HUB_TOKEN environment variable.
• Restart the deployment to roll out an updated pod.

The API is now accessible at the endpoints:

• defined by your Service, accessible internally on port 8000 using http. E.g. http://vllm.your-project.svc.cluster.local:8000/
• defined by your Route, accessible externally through https, e.g. https://vllm.your-project.your-cluster/

vLLM also includes an OpenAPI web interface that can be access from a browser by adding docs to the path, e.g. https://vllm.your-project.your-cluster/docs

Usage

You can directly query the model using the OpenAI API protocol. The server currently hosts one model at a time, and currently implements list models, create chat completion, and create completion endpoints.

Example of calling the /v1/models endpoint to list the available models:

curl http://service:8000//v1/models \
  -H 'accept: application/json'

Example of calling the /v1/completions endpoint to complete a sentence:

curl http://service:8000/v1/completions \
      -H "Content-Type: application/json" \
      -d '{
          "model": "mistralai/Mistral-7B-Instruct-v0.2",
          "prompt": "San Francisco is a",
          "max_tokens": 7,
          "temperature": 0
      }'

Example of calling the /v1/chat/completions endpoint demonstrating a conversation:

curl http://service:8000/v1/chat/completions \
      -H "Content-Type: application/json" \
      -d '{
            "model": "mistralai/Mistral-7B-Instruct-v0.2",
            "messages": [
              {"role": "user", "content": "Who won the world series in 2020?"},
              {"role": "assistant", "content": "You are a helpful assistant."}
            ]
          }'

or from Python:

pip install text-generation

from openai import OpenAI
# Set OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://service:8000/v1"

client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)

chat_response = client.chat.completions.create(
    model="Mistral-7B-Instruct-v0.2",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Tell me a joke."},
    ]
)
print("Chat response:", chat_response)

You can also find a notebook example using Langchain to query vLLM in this repo here.

Advanced installation

Parameters

All the parameters to start the model are passed as arguments (see L52 in the Deployment file).

The full list of available parameters is available here.

Node affinity

As an inference server is something that will run more or less permanently, you may want to dedicate a node to it, or at least set some affinity so that the Pod is scheduled on the right node. To do that, you can modify the default deployment with some taints and affinities, like in the following config. In this example, dedicated nodes are tainted with the key shared-tgi. In case of multiple nodes matching this taint, you can also set an affinity to be sure of being scheduled on the right node in the group.

kind: Deployment
apiVersion: apps/v1
metadata:
  ...
spec:
  ...
  template:
    metadata:
      labels:
        ...
        placement: a10g
    spec:
      restartPolicy: Always
      schedulerName: default-scheduler
      affinity:
        nodeAffinity:
          requiredDuringSchedulingIgnoredDuringExecution:
            nodeSelectorTerms:
              - matchExpressions:
                  - key: nvidia.com/gpu.product
                    operator: In
                    values:
                      - NVIDIA-A10G-SHARED
        podAffinity:
          preferredDuringSchedulingIgnoredDuringExecution:
            - weight: 100
              podAffinityTerm:
                labelSelector:
                  matchExpressions:
                    - key: placement
                      operator: In
                      values:
                        - a10g
                topologyKey: kubernetes.io/hostname
      ...
      tolerations:
        - key: nvidia.com/gpu
          operator: Exists
          effect: NoSchedule
        - key: shared-tgi
          operator: Exists
          effect: NoSchedule