Serve a LLM on GKE with Single-Host TPUs#
This guide showcases how to serve LLMs using Tensor Processing Units (TPUs) with the vLLM serving framework. A single-host TPU slice is a node pool that contains one or more independent TPU VMs. This tutorial uses one TPU Trillium node, which is Google’s sixth generation TPU. For more information about TPUs, see Use TPUs with KubeRay. This example showcases serving and inference with Llama-3.1-70B on a 2x4 Trillium (v6e) TPU node, or with 8 TPU chips total.
This example uses vLLM with KubeRay. However, Ray provides a native Serve LLM API to enable users to deploy multiple LLM models together with the Ray Serve API. This API is the recommended path for running LLMs through Ray Serve.
Step 1: Create a Kubernetes Cluster with TPUs and the Ray Operator Enabled#
First, set environment variables to be used in the following commands:
export PROJECT_ID=$(gcloud config get project)
export CLUSTER_NAME=vllm-tpu
export COMPUTE_REGION=REGION
export COMPUTE_ZONE=ZONE
Replace the following:
REGION: the region where you have TPU quota. Ensure that the TPU version that you want to use is available in this region. To learn more, see TPU availability in GKE.
ZONE: the zone with available TPU quota.
GKE Autopilot#
Create an Autopilot cluster with the Ray operator enabled:
gcloud container clusters create-auto $CLUSTER_NAME \
--enable-ray-operator \
--release-channel=rapid \
--location=$COMPUTE_REGION
GKE Standard#
Create a Standard cluster with the Ray operator enabled:
gcloud container clusters create $CLUSTER_NAME \
--release-channel=rapid \
--location=$COMPUTE_ZONE \
--workload-pool=$PROJECT_ID.svc.id.goog \
--machine-type="n1-standard-4" \
--addons=RayOperator,GcsFuseCsiDriver
Create a single-host TPU slice node pool:
gcloud container node-pools create v6e-8 \
--location=$COMPUTE_ZONE \
--cluster=$CLUSTER_NAME \
--machine-type=ct6e-standard-8t \
--num-nodes=1
Connect to the GKE cluster#
To allow kubectl to communicate with your cluster, run the following command:
gcloud container clusters get-credentials $CLUSTER_NAME \
--location=LOCATION
In the above command, replace LOCATION with \(COMPUTE_ZONE or \)COMPUTE_REGION if you created a Standard or Autopilot GKE cluster respectively.
Step 2: [Optional] Install the KubeRay operator#
Skip this step if the Ray Operator Addon is enabled in your GKE cluster. Follow Deploy a KubeRay operator instructions to install the latest stable KubeRay operator from the Helm repository. Note that the YAML file in this example uses serveConfigV2
, which KubeRay supports starting from v0.6.0.
Step 3: Build a vLLM image with TPU dependencies#
vLLM supports TPUs using PyTorch XLA, providing a Dockerfile.tpu for users to build their own vLLM image with TPU dependencies. This image includes ray[default] as a dependency and will serve as the image on our Ray head and workers.
Create a Docker repository to store the container images for this tutorial:
gcloud artifacts repositories create vllm-tpu \
--repository-format=docker \
--location=us-central1
Clone the vLLM repository:
git clone https://github.com/vllm-project/vllm.git && \
cd vllm
Build the TPU image:
docker build -f Dockerfile.tpu . -t vllm-tpu
Set environment variables to be used for your Docker image:
export TAG="latest"
export PROJECT_NUMBER=$(gcloud projects describe $PROJECT_ID --format="value(projectNumber)")
export VLLM_IMAGE=us-central1-docker.pkg.dev/$PROJECT_ID/vllm-tpu/vllm-tpu:$TAG
Tag the image with your Artifact Registry name:
docker tag vllm-tpu $VLLM_IMAGE
Push the vLLM image to your Artifact registry:
docker push $VLLM_IMAGE
Create a Cloud Storage bucket and setup access#
To accelerate the vLLM deployment startup time and minimize required disk space per node, use the Cloud Storage FUSE CSI driver to mount the downloaded model and compilation cache to the Ray nodes.
Set environment variables to be used in the following commands:
export GSBUCKET=vllm-tpu-bucket
export KSA_NAME=vllm-tpu-sa
export NAMESPACE=default
Create the bucket:
gcloud storage buckets create gs://$GSBUCKET \
--uniform-bucket-level-access
Create a Kubernetes service account to access the bucket:
kubectl create serviceaccount $KSA_NAME \
--namespace $NAMESPACE
Grant the service account read-write access to the Cloud Storage bucket:
gcloud storage buckets add-iam-policy-binding gs://$GSBUCKET \
--member "principal://iam.googleapis.com/projects/$PROJECT_NUMBER/locations/global/workloadIdentityPools/$PROJECT_ID.svc.id.goog/subject/ns/$NAMESPACE/sa/$KSA_NAME" \
--role "roles/storage.objectUser"
Step 4: Create a Kubernetes Secret for Hugging Face credentials#
This example uses meta-llama/Llama-3.1-70B, a gated Hugging Face model that requires access to be granted before use. Create a Hugging Face account, if you don’t already have one, and follow the steps on the model page to request access to the model. Save your Hugging Face token for the following steps.
Set environment variable:
export HF_TOKEN=HUGGING_FACE_TOKEN
Replace HUGGING_FACE_TOKEN with your Hugging Face access token.
Create a Kubernetes Secret with your Hugging Face credentials:
kubectl create secret generic hf-secret \
--from-literal=hf_api_token=$HF_TOKEN \
--dry-run=client -o yaml | kubectl --namespace $NAMESPACE apply -f -
Step 5: Install the RayService CR#
Clone the sample repository:
git clone https://github.com/GoogleCloudPlatform/kubernetes-engine-samples.git
cd kubernetes-engine-samples
Navigate to the working directory:
cd ai-ml/gke-ray/rayserve/llm
Set the following environment vars for the model to serve:
export MODEL_ID="meta-llama/Llama-3.1-70B"
export MAX_MODEL_LEN=8192
In the next step, we’ll deploy a RayService CR with the following:
apiVersion: ray.io/v1
kind: RayService
metadata:
name: vllm-tpu
spec:
serveConfigV2: |
applications:
- name: llm
import_path: ai-ml.gke-ray.rayserve.llm.tpu.serve_tpu:model
deployments:
- name: VLLMDeployment
num_replicas: 1
runtime_env:
working_dir: "https://github.com/GoogleCloudPlatform/kubernetes-engine-samples/archive/main.zip"
env_vars:
MODEL_ID: "$MODEL_ID"
MAX_MODEL_LEN: "$MAX_MODEL_LEN"
DTYPE: "$DTYPE"
TOKENIZER_MODE: "$TOKENIZER_MODE"
TPU_CHIPS: "8"
rayClusterConfig:
headGroupSpec:
rayStartParams: {}
template:
metadata:
annotations:
gke-gcsfuse/volumes: "true"
gke-gcsfuse/cpu-limit: "0"
gke-gcsfuse/memory-limit: "0"
gke-gcsfuse/ephemeral-storage-limit: "0"
spec:
serviceAccountName: $KSA_NAME
containers:
- name: ray-head
image: $VLLM_IMAGE
imagePullPolicy: IfNotPresent
ports:
- containerPort: 6379
name: gcs
- containerPort: 8265
name: dashboard
- containerPort: 10001
name: client
- containerPort: 8000
name: serve
env:
- name: HUGGING_FACE_HUB_TOKEN
valueFrom:
secretKeyRef:
name: hf-secret
key: hf_api_token
- name: VLLM_XLA_CACHE_PATH
value: "/data"
resources:
limits:
cpu: "2"
memory: 8G
requests:
cpu: "2"
memory: 8G
volumeMounts:
- name: gcs-fuse-csi-ephemeral
mountPath: /data
- name: dshm
mountPath: /dev/shm
volumes:
- name: gke-gcsfuse-cache
emptyDir:
medium: Memory
- name: dshm
emptyDir:
medium: Memory
- name: gcs-fuse-csi-ephemeral
csi:
driver: gcsfuse.csi.storage.gke.io
volumeAttributes:
bucketName: $GSBUCKET
mountOptions: "implicit-dirs,file-cache:enable-parallel-downloads:true,file-cache:parallel-downloads-per-file:100,file-cache:max-parallel-downloads:-1,file-cache:download-chunk-size-mb:10,file-cache:max-size-mb:-1"
workerGroupSpecs:
- groupName: tpu-group
replicas: 1
minReplicas: 1
maxReplicas: 1
numOfHosts: 1
rayStartParams: {}
template:
metadata:
annotations:
gke-gcsfuse/volumes: "true"
gke-gcsfuse/cpu-limit: "0"
gke-gcsfuse/memory-limit: "0"
gke-gcsfuse/ephemeral-storage-limit: "0"
spec:
serviceAccountName: $KSA_NAME
containers:
- name: ray-worker
image: $VLLM_IMAGE
imagePullPolicy: IfNotPresent
resources:
limits:
cpu: "100"
google.com/tpu: "8"
ephemeral-storage: 40G
memory: 200G
requests:
cpu: "100"
google.com/tpu: "8"
ephemeral-storage: 40G
memory: 200G
env:
- name: JAX_PLATFORMS
value: "tpu"
- name: HUGGING_FACE_HUB_TOKEN
valueFrom:
secretKeyRef:
name: hf-secret
key: hf_api_token
- name: VLLM_XLA_CACHE_PATH
value: "/data"
volumeMounts:
- name: gcs-fuse-csi-ephemeral
mountPath: /data
- name: dshm
mountPath: /dev/shm
volumes:
- name: gke-gcsfuse-cache
emptyDir:
medium: Memory
- name: dshm
emptyDir:
medium: Memory
- name: gcs-fuse-csi-ephemeral
csi:
driver: gcsfuse.csi.storage.gke.io
volumeAttributes:
bucketName: $GSBUCKET
mountOptions: "implicit-dirs,file-cache:enable-parallel-downloads:true,file-cache:parallel-downloads-per-file:100,file-cache:max-parallel-downloads:-1,file-cache:download-chunk-size-mb:10,file-cache:max-size-mb:-1"
nodeSelector:
cloud.google.com/gke-tpu-accelerator: tpu-v6e-slice
cloud.google.com/gke-tpu-topology: 2x4
The above example RayService manifest deploys on Trillium (v6e) TPUs.
Create the RayService CR:
# For v6e TPUs
envsubst < tpu/ray-service.tpu-v6e-singlehost.yaml | kubectl apply -f -
The above instruction replaces the container images with the VLLM_IMAGE environment variable.
Step 6: View the Serve deployment in the Ray Dashboard#
Verify that you deployed the RayService CR and it’s running:
kubectl get rayservices vllm-tpu
Retrieve the name of the RayCluster head service:
SERVICE_NAME=$(kubectl --namespace=$NAMESPACE get rayservices/vllm-tpu \
--template={{.status.activeServiceStatus.rayClusterStatus.head.serviceName}})
Port-forward the Ray Dashboard from the Ray Head service. To view the dashboard, open http://localhost:8265/ on your local machine.
kubectl port-forward svc/$SERVICE_NAME 8265:8265 2>&1 >/dev/null &
Step 7: Send prompts to the model server#
Port-forward the model endpoint from Ray head:
kubectl port-forward svc/$SERVICE_NAME 8000:8000 2>&1 >/dev/null &
Send a text prompt to the Llama model:
curl -X POST http://localhost:8000/v1/generate -H "Content-Type: application/json" -d '{"prompt": "What are the top 5 most popular programming languages? Be brief.", "max_tokens": 1024}'
We can then view the output of the provided prompt:
{"prompt": "What are the top 5 most popular programming languages? Be brief.", "text": " The answer is based on the Tiobe Index, which is a programming language popularity index that has been ranking programming languages since 2001.\n\n1. **Java**: 14.63% of the index.\n2. **Python**: 11.52% of the index.\n3. **C**: 7.94% of the index.\n4. **C++**: 7.17% of the index.\n5. **JavaScript**: 6.86% of the index.\n\nThe Tiobe Index is based on searches on search engines like Google, as well as on forums and other online platforms. It does not necessarily reflect the actual usage of programming languages in the industry, but it is a good indicator of their popularity and usage in the programming community.
...}