Deploying Furiosa-LLM on Kubernetes

A guide to deploying Furiosa-LLM—an OpenAI-compatible server optimized for Furiosa AI RNGDs—on Kubernetes.

This guide describes how to deploy Furiosa-LLM—an OpenAI-compatible server optimized for Furiosa AI RNGDs—on Kubernetes.

Prerequisites

A Kubernetes cluster equipped with Furiosa RNGD devices.
Hugging Face account and access token.
A Kubernetes storage class which supports dynamic volume provisioning.

For detailed instructions on setting up an RNGD cluster, please refer to Installing Prerequisites and Kubernetes Plugins.

Step 1: Prepare Kubernetes Secret

Create a Kubernetes secret for your Hugging Face token:

yaml

apiVersion: v1
kind: Secret
metadata:
  name: hf-token-secret
type: Opaque
data:
  token: <your_base64_encoded_hf_token>

Encode your token using the following command:

bash

echo -n '<your_HF_TOKEN>' | base64

Then apply the secret:

bash

kubectl apply -f secret.yaml

Step 2: Create Persistent Volume Claim (PVC)

NOTE

Using ephemeral storage inside a Pod for large files, such as LLM model caches, may result in eviction due to disk pressure. Thus, using a Persistent Volume Claim (PVC) is highly recommended to ensure data durability and stability.

Create a PVC manifest to store models. The storage class specified below ("default") is an example, and the actual available storage class in your Kubernetes environment may vary. Please verify and specify the appropriate storage class for your cluster:

yaml

apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: llama-storage
spec:
  accessModes:
    - ReadWriteOnce
  resources:
    requests:
      storage: 10Gi
  # The storage class "default" is an example, use an appropriate one for your cluster.
  storageClassName: default

Apply the PVC:

bash

kubectl apply -f pvc.yaml

Step 3: Deploy Furiosa-LLM Server

Create a Deployment manifest for serving the furiosa-ai/Llama-3.1-8B-Instruct-FP8 model with Furiosa-LLM. For detailed information about the furiosa-ai/Llama-3.1-8B-Instruct-FP8 model, please see the Hugging Face model page.

yaml

apiVersion: apps/v1
kind: Deployment
metadata:
  name: llama-3-8b
  labels:
    app: llama-3-8b
spec:
  replicas: 1
  selector:
    matchLabels:
      app: llama-3-8b
  template:
    metadata:
      labels:
        app: llama-3-8b
    spec:
      containers:
        - name: llama-3-8b
          image: furiosaai/furiosa-llm:latest
          args:
            - "serve"
            - "furiosa-ai/Llama-3.1-8B-Instruct-FP8"
          ports:
            - containerPort: 8000
          resources:
            # Recommended resources for one RNGD card: 10 CPU cores and 100GB memory
            limits:
              cpu: 10
              memory: 100Gi
              furiosa.ai/rngd: "1"
            requests:
              cpu: 10
              memory: 100Gi
              furiosa.ai/rngd: "1"
          env:
            - name: HF_TOKEN
              valueFrom:
                secretKeyRef:
                  name: hf-token-secret
                  key: token
          volumeMounts:
            - name: model-storage
              mountPath: /root/.cache/huggingface
          securityContext:
            capabilities:
              drop:
                - ALL
            seccompProfile:
              type: Unconfined
          # Increase initialDelaySeconds of livenessProbe and readinessProbe if larger models take longer to load
          livenessProbe:
            httpGet:
              path: /health
              port: 8000
            initialDelaySeconds: 180
            periodSeconds: 10
          readinessProbe:
            httpGet:
              path: /health
              port: 8000
            initialDelaySeconds: 180
            periodSeconds: 5
      volumes:
        # If dynamic PVC provisioning isn’t possible in your cluster, consider using a hostPath volume.
        - name: model-storage
          persistentVolumeClaim:
            claimName: llama-storage

Apply the deployment:

bash

kubectl apply -f deployment.yaml

Step 4: Expose the Deployment as a Service

Expose the Furiosa-LLM server using a Kubernetes Service:

yaml

apiVersion: v1
kind: Service
metadata:
  name: llama-3-8b
spec:
  selector:
    app: llama-3-8b
  ports:
    - port: 8000
      targetPort: 8000
  type: ClusterIP

Apply the service:

bash

kubectl apply -f service.yaml

Step 5: Test the Deployment

Confirm the server is running by inspecting the logs:

bash

kubectl logs deployment/llama-3-8b

You should see output similar to:

plaintext

INFO:     Started server process [1]
INFO:     Waiting for application startup.
INFO:     Application startup complete.
INFO:     Uvicorn running on http://0.0.0.0:8000 (Press CTRL+C to quit)

Test the inference endpoint:

bash

kubectl port-forward svc/llama-3-8b 8000:8000

Then, issue a test request:

bash

curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{
        "model": "furiosa-ai/Llama-3.1-8B-Instruct-FP8",
        "prompt": "San Francisco is a",
        "max_tokens": 7,
        "temperature": 0
      }'

You should receive a valid response from the Furiosa-LLM server, similar to:

json

{
  "id": "cmpl-69102e5d78c94e29b74660eaadbe39db",
  "object": "text_completion",
  "created": 1748675715,
  "model": "meta-llama/Llama-3.1-8B-Instruct",
  "choices": [
    {
      "index": 0,
      "text": " city that is known for its vibrant",
      "logprobs": null,
      "finish_reason": "length",
      "prompt_logprobs": null
    }
  ],
  "usage": {
    "prompt_tokens": 5,
    "total_tokens": 12,
    "completion_tokens": 7,
    "completion_tokens_details": null
  }
}

Conclusion

Deploying Furiosa-LLM with Kubernetes leverages the efficiency and scalability of Furiosa RNGD accelerators. This guide provides a straightforward approach to setting up your Kubernetes cluster to serve inference workloads efficiently.

Deploying Furiosa-LLM on Kubernetes

Prerequisites

Step 1: Prepare Kubernetes Secret

Step 2: Create Persistent Volume Claim (PVC)

Step 3: Deploy Furiosa-LLM Server

Step 4: Expose the Deployment as a Service

Step 5: Test the Deployment

Conclusion

On this page