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 cluster page:

   Go to Create an Autopilot cluster

2. In the Name field, enter gke-gpu-cluster.

3. In the Region list, select us-central1.

4. Click Create.

Create a Cloud Storage bucket

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

   Go to Create a bucket

2. In the Name your bucket field, 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 Region list, select us-central1 (Iowa) and click Continue.

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

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

8. Click Create.

9. In the Public access will be prevented dialog, ensure that the Enforce public access prevention on this bucket checkbox 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.