Serve an LLM using TPUs on GKE with JetStream and PyTorch

This guide shows you how to serve a large language model (LLM) using Tensor Processing Units (TPUs) on Google Kubernetes Engine (GKE) with JetStream through PyTorch. In this guide, you download model weights to Cloud Storage and deploy them on a GKE Autopilot or Standard cluster using a container that runs JetStream.

If you need the scalability, resilience, and cost-effectiveness offered by Kubernetes features when deploying your model on JetStream, this guide is a good starting point.

This guide is intended for Generative AI customers who use PyTorch, new or existing users of GKE, ML Engineers, MLOps (DevOps) engineers, or platform administrators who are interested in using Kubernetes container orchestration capabilities for serving LLMs.

Background

By serving an LLM using TPUs on GKE with JetStream, you can build a robust, production-ready serving solution with all the benefits of managed Kubernetes, including cost-efficiency, scalability and higher availability. This section describes the key technologies used in this tutorial.

About TPUs

TPUs are Google's custom-developed application-specific integrated circuits (ASICs) used to accelerate machine learning and AI models built using frameworks such as TensorFlow, PyTorch, and JAX.

Before you use TPUs in GKE, we recommend that you complete the following learning path:

1. Learn about current TPU version availability with the Cloud TPU system architecture.
2. Learn about TPUs in GKE.

This tutorial covers serving various LLM models. GKE deploys the model on single-host TPUv5e nodes with TPU topologies configured based on the model requirements for serving prompts with low latency.

About JetStream

JetStream is an open source inference serving framework developed by Google. JetStream enables high-performance, high-throughput, and memory-optimized inference on TPUs and GPUs. JetStream provides advanced performance optimizations, including continuous batching, KV cache optimizations, and quantization techniques, to facilitate LLM deployment. JetStream enables PyTorch/XLA and JAX TPU serving to achieve optimal performance.

Continuous Batching

Continuous batching is a technique that dynamically groups incoming inference requests into batches, reducing latency and increasing throughput.

KV cache quantization

KV cache quantization involves compressing the key-value cache used in attention mechanisms, reducing memory requirements.

Int8 weight quantization

Int8 weight quantization reduces the precision of model weights from 32-bit floating point to 8-bit integers, leading to faster computation and reduced memory usage.

To learn more about these optimizations, refer to the JetStream PyTorch and JetStream MaxText project repositories.

About PyTorch

PyTorch is an open source machine learning framework developed by Meta and now part of the Linux Foundation umbrella. PyTorch provides high-level features such as tensor computation and deep neural networks.

Objectives

1. Prepare a GKE Autopilot or Standard cluster with the recommended TPU topology based on the model characteristics.
2. Deploy JetStream components on GKE.
3. Get and publish your model.
4. Serve and interact with the published model.

Architecture

This section describes the GKE architecture used in this tutorial. The architecture includes a GKE Autopilot or Standard cluster that provisions TPUs and hosts JetStream components to deploy and serve the models.

The following diagram shows you the components of this architecture:

This architecture includes the following components:

• A GKE Autopilot or Standard regional cluster.
• Two single-host TPU slice node pools that host the JetStream deployment.
• The Service component spreads inbound traffic to all JetStream HTTP replicas.
• JetStream HTTP is an HTTP server which accepts requests as a wrapper to JetStream's required format and sends it to JetStream's GRPC client.
• JetStream-PyTorch is a JetStream server that performs inferencing with continuous batching.

Before you begin

• Sign in to your Google Cloud account. If you're new to Google Cloud, create an account to evaluate how our products perform in real-world scenarios. New customers also get $300 in free credits to run, test, and deploy workloads.

• In the Google Cloud console, on the project selector page, select or create a Google Cloud project.

  Go to project selector

• Make sure that billing is enabled for your Google Cloud project.

• Enable the required API.

  Enable the API

• In the Google Cloud console, on the project selector page, select or create a Google Cloud project.

  Go to project selector

• Make sure that billing is enabled for your Google Cloud project.

• Enable the required API.

  Enable the API

• Make sure that you have the following role or roles on the project: roles/container.admin, roles/iam.serviceAccountAdmin

  Check for the roles

  1. In the Google Cloud console, go to the IAM page.

     Go to IAM
  2. Select the project.

  3. In the Principal column, find all rows that identify you or a group that you're included in. To learn which groups you're included in, contact your administrator.

  4. For all rows that specify or include you, check the Role colunn to see whether the list of roles includes the required roles.

  Grant the roles

  1. In the Google Cloud console, go to the IAM page.

     Go to IAM
  2. Select the project.
  3. Click person_add Grant access.

  4. In the New principals field, enter your user identifier. This is typically the email address for a Google Account.

  5. In the Select a role list, select a role.
  6. To grant additional roles, click add Add another role and add each additional role.
  7. Click Save.

• Ensure that you have sufficient quota for eight TPU v5e PodSlice Lite chips. In this tutorial, you use on-demand instances.
• Create a Hugging Face token, if you don't already have one.

Get access to the model

Get access to various models on Hugging Face for deployment to GKE

Prepare the environment

In this tutorial, you use Cloud Shell to manage resources hosted on Google Cloud. Cloud Shell comes preinstalled with the software you'll need for this tutorial, including kubectl and gcloud CLI.

To set up your environment with Cloud Shell, follow these steps:

1. In the Google Cloud console, launch a Cloud Shell session by clicking Activate Cloud Shell in the Google Cloud console. This launches a session in the bottom pane of Google Cloud console.

2. Set the default environment variables:

   gcloud config set project PROJECT_ID
   export PROJECT_ID=$(gcloud config get project)
   export CLUSTER_NAME=CLUSTER_NAME
   export BUCKET_NAME=BUCKET_NAME
   export REGION=REGION
   export LOCATION=LOCATION
   export CLUSTER_VERSION=CLUSTER_VERSION
   

   Replace the following values:

   • PROJECT_ID: your Google Cloud project ID.
   • CLUSTER_NAME: the name of your GKE cluster.
   • BUCKET_NAME: the name of your Cloud Storage bucket. You don't need to specify the gs:// prefix.
   • REGION: the region where your GKE cluster, Cloud Storage bucket, and TPU nodes are located. The region contains zones where TPU v5e machine types are available (for example, us-west1, us-west4, us-central1, us-east1, us-east5, or europe-west4). For Autopilot clusters, ensure that you have sufficient TPU v5e zonal resources for your region of choice.
   • (Standard cluster only) LOCATION: the zone where the TPU resources are available (for example, us-west4-a). For Autopilot clusters, you don't need to specify the zone, only the region.
   • CLUSTER_VERSION: the GKE version, which must support the machine type that you want to use. Note that the default GKE version might not have availability for your target TPU. For a list of minimum GKE versions available by TPU machine type, see TPU availability in GKE.

Create and configure Google Cloud resources

Follow these instructions to create the required resources.

Create a GKE cluster

You can serve Gemma on TPUs in a GKE Autopilot or Standard cluster. We recommend that you use a Autopilot cluster for a fully managed Kubernetes experience. To choose the GKE mode of operation that's the best fit for your workloads, see Choose a GKE mode of operation.

gcloud container clusters create-auto CLUSTER_NAME \
    --project=PROJECT_ID \
    --region=REGION \
    --cluster-version=CLUSTER_VERSION

Create a Cloud Storage bucket

Create a Cloud Storage bucket to store the converted checkpoint:

gcloud storage buckets create gs://BUCKET_NAME --location=REGION

Generate your Hugging Face CLI token in Cloud Shell

Generate a new Hugging Face token if you don't already have one:

1. Click Your Profile > Settings > Access Tokens.
2. Click New Token.
3. Specify a Name of your choice and a Role of at least Read.
4. Click Generate a token.
5. Edit permissions to your access token to have read access to your model's Hugging Face repository.
6. Copy the generated token to your clipboard.

Create a Kubernetes Secret for Hugging Face credentials

In Cloud Shell, do the following:

1. Configure kubectl to communicate with your cluster:

   gcloud container clusters get-credentials CLUSTER_NAME --location=REGION
   

2. Create a Secret to store the Hugging Face credentials:

   kubectl create secret generic huggingface-secret \
       --from-literal=HUGGINGFACE_TOKEN=HUGGINGFACE_TOKEN
   

   Replace HUGGINGFACE_TOKEN with your Hugging Face token.

Configure your workloads access using Workload Identity Federation for GKE

Assign a Kubernetes ServiceAccount to the application and configure that Kubernetes ServiceAccount to act as an IAM service account.

1. Create an IAM service account for your application:

   gcloud iam service-accounts create wi-jetstream
   

2. Add an IAM policy binding for your IAM service account to manage Cloud Storage:

   gcloud projects add-iam-policy-binding PROJECT_ID \
       --member "serviceAccount:wi-jetstream@PROJECT_ID.iam.gserviceaccount.com" \
       --role roles/storage.objectUser
   
   gcloud projects add-iam-policy-binding PROJECT_ID \
       --member "serviceAccount:wi-jetstream@PROJECT_ID.iam.gserviceaccount.com" \
       --role roles/storage.insightsCollectorService
   

3. Allow the Kubernetes ServiceAccount to impersonate the IAM service account by adding an IAM policy binding between the two service accounts. This binding allows the Kubernetes ServiceAccount to act as the IAM service account:

   gcloud iam service-accounts add-iam-policy-binding wi-jetstream@PROJECT_ID.iam.gserviceaccount.com \
       --role roles/iam.workloadIdentityUser \
       --member "serviceAccount:PROJECT_ID.svc.id.goog[default/default]"
   

4. Annotate the Kubernetes service account with the email address of the IAM service account:

   kubectl annotate serviceaccount default \
       iam.gke.io/gcp-service-account=wi-jetstream@PROJECT_ID.iam.gserviceaccount.com
   

Convert the model checkpoints

In this section, you create a Job to do the following:

1. Download the base checkpoint from Hugging Face to the local directory.
2. Convert the checkpoint to a JetStream-Pytorch compatible checkpoint.
3. Upload the checkpoint to a Cloud Storage bucket.

Deploy the model checkpoint conversion Job

1. Replace BUCKET_NAME with your GSBucket created earlier:

   sed -i "s|BUCKET_NAME|BUCKET_NAME|g" job-checkpoint-converter.yaml
   

2. Apply the manifest:

   kubectl apply -f job-checkpoint-converter.yaml
   

3. Wait for the Pod scheduling the Job to begin running:

   kubectl get pod -w
   

   The output will be similar to the following, this may take a few minutes:

   NAME                        READY   STATUS              RESTARTS   AGE
   checkpoint-converter-abcd   0/1     ContainerCreating   0          28s
   checkpoint-converter-abcd   1/1     Running             0          51s
   

   For Autopilot clusters, it may take a few minutes to provision the required TPU resources.

4. Verify the Job is completed by viewing the logs for the Job:

   kubectl logs -f jobs/checkpoint-converter
   

   When the Job is completed, the output is similar to the following:

   Completed uploading converted checkpoint from local path /pt-ckpt/ to GSBucket gs://BUCKET_NAME/pytorch/<model_name>/final/bf16/"
   

Deploy JetStream

Deploy the JetStream container to serve your model:

Save the following manifest as jetstream-pytorch-deployment.yaml:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: jetstream-pytorch-server
spec:
  replicas: 2
  selector:
    matchLabels:
      app: jetstream-pytorch-server
  template:
    metadata:
      labels:
        app: jetstream-pytorch-server
      annotations:
        gke-gcsfuse/volumes: "true"
    spec:
      nodeSelector:
        cloud.google.com/gke-tpu-topology: 2x4
        cloud.google.com/gke-tpu-accelerator: tpu-v5-lite-podslice
      containers:
      - name: jetstream-pytorch-server
        image: us-docker.pkg.dev/cloud-tpu-images/inference/jetstream-pytorch-server:v0.2.3
        args:
        - --size=7b
        - --model_name=gemma
        - --batch_size=32
        - --max_cache_length=2048
        - --quantize_weights=False
        - --quantize_kv_cache=False
        - --tokenizer_path=/models/pytorch/gemma-7b-it/final/bf16/tokenizer.model
        - --checkpoint_path=/models/pytorch/gemma-7b-it/final/bf16/model.safetensors
        ports:
        - containerPort: 9000
        volumeMounts:
        - name: gcs-fuse-checkpoint
          mountPath: /models
          readOnly: true
        resources:
          requests:
            google.com/tpu: 8
          limits:
            google.com/tpu: 8
      - name: jetstream-http
        image: us-docker.pkg.dev/cloud-tpu-images/inference/jetstream-http:v0.2.2
        ports:
        - containerPort: 8000
      volumes:
      - name: gcs-fuse-checkpoint
        csi:
          driver: gcsfuse.csi.storage.gke.io
          readOnly: true
          volumeAttributes:
            bucketName: BUCKET_NAME
            mountOptions: "implicit-dirs"
---
apiVersion: v1
kind: Service
metadata:
  name: jetstream-svc
spec:
  selector:
    app: jetstream-pytorch-server
  ports:
  - protocol: TCP
    name: jetstream-http
    port: 8000
    targetPort: 8000

apiVersion: apps/v1
kind: Deployment
metadata:
  name: jetstream-pytorch-server
spec:
  replicas: 2
  selector:
    matchLabels:
      app: jetstream-pytorch-server
  template:
    metadata:
      labels:
        app: jetstream-pytorch-server
      annotations:
        gke-gcsfuse/volumes: "true"
    spec:
      nodeSelector:
        cloud.google.com/gke-tpu-topology: 2x4
        cloud.google.com/gke-tpu-accelerator: tpu-v5-lite-podslice
      containers:
      - name: jetstream-pytorch-server
        image: us-docker.pkg.dev/cloud-tpu-images/inference/jetstream-pytorch-server:v0.2.3
        args:
        - --size=8b
        - --model_name=llama-3
        - --batch_size=32
        - --max_cache_length=2048
        - --quantize_weights=False
        - --quantize_kv_cache=False
        - --tokenizer_path=/models/pytorch/llama-3-8b/final/bf16/tokenizer.model
        - --checkpoint_path=/models/pytorch/llama-3-8b/final/bf16/model.safetensors
        ports:
        - containerPort: 9000
        volumeMounts:
        - name: gcs-fuse-checkpoint
          mountPath: /models
          readOnly: true
        resources:
          requests:
            google.com/tpu: 8
          limits:
            google.com/tpu: 8
      - name: jetstream-http
        image: us-docker.pkg.dev/cloud-tpu-images/inference/jetstream-http:v0.2.2
        ports:
        - containerPort: 8000
      volumes:
      - name: gcs-fuse-checkpoint
        csi:
          driver: gcsfuse.csi.storage.gke.io
          readOnly: true
          volumeAttributes:
            bucketName: BUCKET_NAME
            mountOptions: "implicit-dirs"
---
apiVersion: v1
kind: Service
metadata:
  name: jetstream-svc
spec:
  selector:
    app: jetstream-pytorch-server
  ports:
  - protocol: TCP
    name: jetstream-http
    port: 8000
    targetPort: 8000

The manifest sets the following key properties:

• size: the size of your model.
• model_name: the model name (gemma, llama-3).
• batch_size: the decoding batch size per device, where one TPU chip equals one device.
• max_cache_length: the maximum length of kv cache.
• quantize_weights: whether the checkpoint is quantized.
• quantize_kv_cache: whether the kv cache is quantized.
• tokenizer_path: the path to the model tokenizer file.
• checkpoint_path: the path to the checkpoint.

1. Replace BUCKET_NAME with your GSBucket created earlier:

   sed -i "s|BUCKET_NAME|BUCKET_NAME|g" jetstream-pytorch-deployment.yaml
   

2. Apply the manifest:

   kubectl apply -f jetstream-pytorch-deployment.yaml
   

3. Verify the Deployment:

   kubectl get deployment
   

   The output is similar to the following:

   NAME                              READY   UP-TO-DATE   AVAILABLE   AGE
   jetstream-pytorch-server          2/2     2            2           ##s
   

   For Autopilot clusters, it may take a few minutes to provision the required TPU resources.

4. View the HTTP server logs to check that the model has been loaded and compiled. It might take the server a few minutes to complete this operation.

   kubectl logs deploy/jetstream-pytorch-server -f -c jetstream-http
   

   The output is similar to the following:

   kubectl logs deploy/jetstream-pytorch-server -f -c jetstream-http
   
   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)
   

5. View the JetStream-PyTorch server logs and verify that the compilation is done:

   kubectl logs deploy/jetstream-pytorch-server -f -c jetstream-pytorch-server
   

   The output is similar to the following:

   Started jetstream_server....
   2024-04-12 04:33:37,128 - root - INFO - ---------Generate params 0 loaded.---------
   

Serve the model

In this section, you interact with the model.

Set up port forwarding

You can access the JetStream Deployment through the ClusterIP Service that you created in the preceding step. The ClusterIP Services are only reachable from within the cluster. Therefore, to access the Service from outside the cluster, complete the following steps:

To establish a port forwarding session, run the following command:

kubectl port-forward svc/jetstream-svc 8000:8000

Interact with the model using curl

1. Verify that you can access the JetStream HTTP server by opening a new terminal and running the following command:

   curl --request POST \
   --header "Content-type: application/json" \
   -s \
   localhost:8000/generate \
   --data \
   '{
       "prompt": "What are the top 5 programming languages",
       "max_tokens": 200
   }'
   

   The initial request can take several seconds to complete due to model warmup. The output is similar to the following:

   {
       "response": " for data science in 2023?\n\n**1. Python:**\n- Widely used for data science due to its readability, extensive libraries (pandas, scikit-learn), and integration with other tools.\n- High demand for Python programmers in data science roles.\n\n**2. R:**\n- Popular choice for data analysis and visualization, particularly in academia and research.\n- Extensive libraries for statistical modeling and data wrangling.\n\n**3. Java:**\n- Enterprise-grade platform for data science, with strong performance and scalability.\n- Widely used in data mining and big data analytics.\n\n**4. SQL:**\n- Essential for data querying and manipulation, especially in relational databases.\n- Used for data analysis and visualization in various industries.\n\n**5. Scala:**\n- Scalable and efficient for big data processing and machine learning models.\n- Popular in data science for its parallelism and integration with Spark and Spark MLlib."
   }
   
   

Troubleshoot issues

• If you get the message Empty reply from server, it's possible the container has not finished downloading the model data. Check the Pod's logs again for the Connected message which indicates that the model is ready to serve.
• If you see Connection refused, verify that your port forwarding is active.

Clean up

To avoid incurring charges to your Google Cloud account for the resources used in this tutorial, either delete the project that contains the resources, or keep the project and delete the individual resources.

Delete the deployed resources

To avoid incurring charges to your Google Cloud account for the resources that you created in this guide, run the following commands and follow the prompts:

gcloud container clusters delete CLUSTER_NAME --region=REGION

gcloud iam service-accounts delete wi-jetstream@PROJECT_ID.iam.gserviceaccount.com

gcloud storage rm --recursive gs://BUCKET_NAME

What's next

• Discover how you can run Gemma models on GKE and how to run optimized AI/ML workloads with GKE platform orchestration capabilities.
• Learn more about TPUs in GKE.
• Explore the JetStream GitHub repository.
• Explore the Vertex AI Model Garden.