Train a model with GPUs on GKE Autopilot mode

Clone the sample repository

In Cloud Shell, run the following command:

 git  
clone  
https://github.com/GoogleCloudPlatform/ai-on-gke && 
 \ 
 cd 
  
ai-on-gke/tutorials-and-examples/gpu-examples/training-single-gpu 

Create a cluster

1. In the Google Cloud console, go to the Create an Autopilot clusterpage:

   Go to Create an Autopilot cluster

2. In the Namefield, enter gke-gpu-cluster .

3. In the Regionlist, select us-central1 .

4. Click Create.

Create a Cloud Storage bucket

1. In the Google Cloud console, go to the Create a bucketpage:

   Go to Create a bucket

2. In the Name your bucketfield, enter the following name:

     PROJECT_ID 
   -gke-gpu-bucket 
   

   Replace PROJECT_ID with your Google Cloud project ID.

3. Click Continue.

4. For Location type, select Region.

5. In the Regionlist, select us-central1 (Iowa) and click Continue.

6. In the Choose a storage class for your datasection, click Continue.

7. In the Choose how to control access to objectssection, for Access control, select Uniform.

8. Click Create.

9. In the Public access will be preventeddialog, ensure that the Enforce public access prevention on this bucketcheckbox is selected, and click Confirm.

Configure your cluster to access the bucket using Workload Identity Federation for GKE

To let your cluster access the Cloud Storage bucket, you do the following:

1. Create a Kubernetes ServiceAccount in your cluster.
2. Create an IAM allow policy that lets the ServiceAccount access the bucket.

Create a Kubernetes ServiceAccount in your cluster

In Cloud Shell, do the following:

1. Connect to your cluster:

    gcloud  
   container  
   clusters  
   get-credentials  
   gke-gpu-cluster  
    \ 
     
   --location = 
   us-central1 
   

2. Create a Kubernetes namespace:

    kubectl  
   create  
   namespace  
   gke-gpu-namespace 
   

3. Create a Kubernetes ServiceAccount in the namespace:

    kubectl  
   create  
   serviceaccount  
   gpu-k8s-sa  
   --namespace = 
   gke-gpu-namespace 
   

Create an IAM allow policy on the bucket

Grant the Storage Object Admin ( roles/storage.objectAdmin ) role on the bucket to the Kubernetes ServiceAccount:

 gcloud  
storage  
buckets  
add-iam-policy-binding  
gs:// PROJECT_ID 
  
 \ 
  
--member = 
principal://iam.googleapis.com/ PROJECT_NUMBER 
/locations/global/workloadIdentityPools/ PROJECT_ID 
.svc.id.goog/subject/ns/gke-gpu-namespace/sa/gpu-k8s-sa  
 \ 
  
--role = 
roles/storage.objectAdmin  
 \ 
  
--condition = 
None 

Replace PROJECT_NUMBER with your Google Cloud project number.

Verify that Pods can access the Cloud Storage bucket

1. In Cloud Shell, create the following environment variables:

     export 
     
    K8S_SA_NAME 
    = 
   gpu-k8s-sa export 
     
    BUCKET_NAME 
    = 
    PROJECT_ID 
   -gke-gpu-bucket 
   

   Replace PROJECT_ID with your Google Cloud project ID.

2. Create a Pod that has a TensorFlow container:

    envsubst < 
   src/gke-config/standard-tensorflow-bash.yaml  
    | 
     
   kubectl  
   --namespace = 
   gke-gpu-namespace  
   apply  
   -f  
   - 
   

   This command inserts the environment variables that you created into the corresponding references in the manifest. You can also open the manifest in a text editor and replace $K8S_SA_NAME and $BUCKET_NAME with the corresponding values.

3. Create a sample file in the bucket:

    touch  
   sample-file
   gsutil  
   cp  
   sample-file  
   gs:// PROJECT_ID 
   -gke-gpu-bucket 
   

4. Wait for your Pod to become ready:

    kubectl  
    wait 
     
   --for = 
    condition 
    = 
   Ready  
   pod/test-tensorflow-pod  
   -n = 
   gke-gpu-namespace  
   --timeout = 
   180s 
   

   When the Pod is ready, the output is the following:

    pod/test-tensorflow-pod condition met 
   

   If the command times out, GKE might still be creating new nodes to run the Pods. Run the command again and wait for the Pod to become ready.

5. Open a shell in the TensorFlow container:

    kubectl  
   -n  
   gke-gpu-namespace  
    exec 
     
   --stdin  
   --tty  
   test-tensorflow-pod  
   --container  
   tensorflow  
   --  
   /bin/bash 
   

6. Try to read the sample file that you created:

    ls  
   /data 
   

   The output shows the sample file.

7. Check the logs to identify the GPU attached to the Pod:

    python  
   -c  
    "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))" 
    
   

   The output shows the GPU attached to the Pod, similar to the following:

    ...
   PhysicalDevice(name='/physical_device:GPU:0',device_type='GPU') 
   

8. Exit the container:

     exit 
    
   

9. Delete the sample Pod:

    kubectl  
   delete  
   -f  
   src/gke-config/standard-tensorflow-bash.yaml  
    \ 
     
   --namespace = 
   gke-gpu-namespace 
   

Train and predict using the MNIST dataset

In this section, you run a training workload on the MNIST example dataset.

1. Copy the example data to the Cloud Storage bucket:

    gsutil  
   -m  
   cp  
   -R  
   src/tensorflow-mnist-example  
   gs:// PROJECT_ID 
   -gke-gpu-bucket/ 
   

2. Create the following environment variables:

     export 
     
    K8S_SA_NAME 
    = 
   gpu-k8s-sa export 
     
    BUCKET_NAME 
    = 
    PROJECT_ID 
   -gke-gpu-bucket 
   

3. Review the training Job:

     # Copyright 2023 Google LLC 
    # 
    # Licensed under the Apache License, Version 2.0 (the "License"); 
    # you may not use this file except in compliance with the License. 
    # You may obtain a copy of the License at 
    # 
    #      http://www.apache.org/licenses/LICENSE-2.0 
    # 
    # Unless required by applicable law or agreed to in writing, software 
    # distributed under the License is distributed on an "AS IS" BASIS, 
    # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 
    # See the License for the specific language governing permissions and 
    # limitations under the License. 
    apiVersion 
    : 
     
    batch/v1 
    kind 
    : 
     
    Job 
    metadata 
    : 
     
    name 
    : 
     
    mnist-training-job 
    spec 
    : 
     
    template 
    : 
     
    metadata 
    : 
     
    name 
    : 
     
    mnist 
     
    annotations 
    : 
     
    gke-gcsfuse/volumes 
    : 
     
    "true" 
     
    spec 
    : 
     
    nodeSelector 
    : 
     
    cloud.google.com/gke-accelerator 
    : 
     
    nvidia-tesla-t4 
     
    tolerations 
    : 
     
    - 
     
    key 
    : 
     
    "nvidia.com/gpu" 
     
    operator 
    : 
     
    "Exists" 
     
    effect 
    : 
     
    "NoSchedule" 
      
    containers 
    : 
     
    - 
     
    name 
    : 
     
    tensorflow 
     
    image 
    : 
     
    tensorflow/tensorflow:latest-gpu 
     
     
    command 
    : 
     
    [ 
    "/bin/bash" 
    , 
     
    "-c" 
    , 
     
    "--" 
    ] 
     
    args 
    : 
     
    [ 
    "cd 
     
    /data/tensorflow-mnist-example; 
     
    pip 
     
    install 
     
    -r 
     
    requirements.txt; 
     
    python 
     
    tensorflow_mnist_train_distributed.py" 
    ] 
     
    resources 
    : 
     
    limits 
    : 
      
    nvidia.com/gpu 
    : 
     
    1 
     
    cpu 
    : 
     
    1 
     
    memory 
    : 
     
    3Gi 
     
    volumeMounts 
    : 
     
    - 
     
    name 
    : 
     
    gcs-fuse-csi-vol 
     
    mountPath 
    : 
     
    /data 
     
    readOnly 
    : 
     
    false 
     
    serviceAccountName 
    : 
     
    $K8S_SA_NAME 
     
    volumes 
    : 
     
    - 
     
    name 
    : 
     
    gcs-fuse-csi-vol 
     
    csi 
    : 
     
    driver 
    : 
     
    gcsfuse.csi.storage.gke.io 
     
    readOnly 
    : 
     
    false 
     
    volumeAttributes 
    : 
     
    bucketName 
    : 
     
    $BUCKET_NAME 
     
    mountOptions 
    : 
     
    "implicit-dirs" 
     
    restartPolicy 
    : 
     
    "Never" 
    
   

4. Deploy the training Job:

    envsubst < 
   src/gke-config/standard-tf-mnist-train.yaml  
    | 
     
   kubectl  
   -n  
   gke-gpu-namespace  
   apply  
   -f  
   - 
   

   This command replaces the environment variables that you created into the corresponding references in the manifest. You can also open the manifest in a text editor and replace $K8S_SA_NAME and $BUCKET_NAME with the corresponding values.

5. Wait until the Job has the Completed status:

    kubectl  
    wait 
     
   -n  
   gke-gpu-namespace  
   --for = 
    condition 
    = 
   Complete  
   job/mnist-training-job  
   --timeout = 
   180s 
   

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

    job.batch/mnist-training-job condition met 
   

   If the command times out, GKE might still be creating new nodes to run the Pods. Run the command again and wait for the Job to become ready.

6. Check the logs from the TensorFlow container:

    kubectl  
   logs  
   -f  
   jobs/mnist-training-job  
   -c  
   tensorflow  
   -n  
   gke-gpu-namespace 
   

   The output shows the following events occur:

   • Install required Python packages
   • Download the MNIST dataset
   • Train the model using a GPU
   • Save the model
   • Evaluate the model

    ...
   Epoch 12/12
   927/938 [============================>.] - ETA: 0s - loss: 0.0188 - accuracy: 0.9954
   Learning rate for epoch 12 is 9.999999747378752e-06
   938/938 [==============================] - 5s 6ms/step - loss: 0.0187 - accuracy: 0.9954 - lr: 1.0000e-05
   157/157 [==============================] - 1s 4ms/step - loss: 0.0424 - accuracy: 0.9861
   Eval loss: 0.04236088693141937, Eval accuracy: 0.9861000180244446 Training finished. Model saved 
   

7. Delete the training workload:

    kubectl  
   -n  
   gke-gpu-namespace  
   delete  
   -f  
   src/gke-config/standard-tf-mnist-train.yaml 
   

Deploy an inference workload

In this section, you deploy an inference workload that takes a sample dataset as input and returns predictions.

1. Copy the images for prediction to the bucket:

    gsutil  
   -m  
   cp  
   -R  
   data/mnist_predict  
   gs:// PROJECT_ID 
   -gke-gpu-bucket/ 
   

2. Review the inference workload:

     # Copyright 2023 Google LLC 
    # 
    # Licensed under the Apache License, Version 2.0 (the "License"); 
    # you may not use this file except in compliance with the License. 
    # You may obtain a copy of the License at 
    # 
    #      http://www.apache.org/licenses/LICENSE-2.0 
    # 
    # Unless required by applicable law or agreed to in writing, software 
    # distributed under the License is distributed on an "AS IS" BASIS, 
    # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 
    # See the License for the specific language governing permissions and 
    # limitations under the License. 
    apiVersion 
    : 
     
    batch/v1 
    kind 
    : 
     
    Job 
    metadata 
    : 
     
    name 
    : 
     
    mnist-batch-prediction-job 
    spec 
    : 
     
    template 
    : 
     
    metadata 
    : 
     
    name 
    : 
     
    mnist 
     
    annotations 
    : 
     
    gke-gcsfuse/volumes 
    : 
     
    "true" 
     
    spec 
    : 
     
    nodeSelector 
    : 
     
    cloud.google.com/gke-accelerator 
    : 
     
    nvidia-tesla-t4 
     
    tolerations 
    : 
     
    - 
     
    key 
    : 
     
    "nvidia.com/gpu" 
     
    operator 
    : 
     
    "Exists" 
     
    effect 
    : 
     
    "NoSchedule" 
      
    containers 
    : 
     
    - 
     
    name 
    : 
     
    tensorflow 
     
    image 
    : 
     
    tensorflow/tensorflow:latest-gpu 
     
     
    command 
    : 
     
    [ 
    "/bin/bash" 
    , 
     
    "-c" 
    , 
     
    "--" 
    ] 
     
    args 
    : 
     
    [ 
    "cd 
     
    /data/tensorflow-mnist-example; 
     
    pip 
     
    install 
     
    -r 
     
    requirements.txt; 
     
    python 
     
    tensorflow_mnist_batch_predict.py" 
    ] 
     
    resources 
    : 
     
    limits 
    : 
      
    nvidia.com/gpu 
    : 
     
    1 
     
    cpu 
    : 
     
    1 
     
    memory 
    : 
     
    3Gi 
     
    volumeMounts 
    : 
     
    - 
     
    name 
    : 
     
    gcs-fuse-csi-vol 
     
    mountPath 
    : 
     
    /data 
     
    readOnly 
    : 
     
    false 
     
    serviceAccountName 
    : 
     
    $K8S_SA_NAME 
     
    volumes 
    : 
     
    - 
     
    name 
    : 
     
    gcs-fuse-csi-vol 
     
    csi 
    : 
     
    driver 
    : 
     
    gcsfuse.csi.storage.gke.io 
     
    readOnly 
    : 
     
    false 
     
    volumeAttributes 
    : 
     
    bucketName 
    : 
     
    $BUCKET_NAME 
     
    mountOptions 
    : 
     
    "implicit-dirs" 
     
    restartPolicy 
    : 
     
    "Never" 
    
   

3. Deploy the inference workload:

    envsubst < 
   src/gke-config/standard-tf-mnist-batch-predict.yaml  
    | 
     
   kubectl  
   -n  
   gke-gpu-namespace  
   apply  
   -f  
   - 
   

   This command replaces the environment variables that you created into the corresponding references in the manifest. You can also open the manifest in a text editor and replace $K8S_SA_NAME and $BUCKET_NAME with the corresponding values.

4. Wait until the Job has the Completed status:

    kubectl  
    wait 
     
   -n  
   gke-gpu-namespace  
   --for = 
    condition 
    = 
   Complete  
   job/mnist-batch-prediction-job  
   --timeout = 
   180s 
   

   The output is similar to the following:

    job.batch/mnist-batch-prediction-job condition met 
   

5. Check the logs from the TensorFlow container:

    kubectl  
   logs  
   -f  
   jobs/mnist-batch-prediction-job  
   -c  
   tensorflow  
   -n  
   gke-gpu-namespace 
   

   The output is the prediction for each image and the model's confidence in the prediction, similar to the following:

    Found 10 files belonging to 1 classes.
   1/1 [==============================] - 2s 2s/step
   The image /data/mnist_predict/0.png is the number 0 with a 100.00 percent confidence.
   The image /data/mnist_predict/1.png is the number 1 with a 99.99 percent confidence.
   The image /data/mnist_predict/2.png is the number 2 with a 100.00 percent confidence.
   The image /data/mnist_predict/3.png is the number 3 with a 99.95 percent confidence.
   The image /data/mnist_predict/4.png is the number 4 with a 100.00 percent confidence.
   The image /data/mnist_predict/5.png is the number 5 with a 100.00 percent confidence.
   The image /data/mnist_predict/6.png is the number 6 with a 99.97 percent confidence.
   The image /data/mnist_predict/7.png is the number 7 with a 100.00 percent confidence.
   The image /data/mnist_predict/8.png is the number 8 with a 100.00 percent confidence.
   The image /data/mnist_predict/9.png is the number 9 with a 99.65 percent confidence.