Run TPU workloads in a Docker container

TPU device

1. Create Cloud TPU VM

   gcloud compute tpus tpu-vm create $TPU_NAME \
   --zone=$ZONE \
   --accelerator-type=$ACCELERATOR_TYPE \
   --version=$RUNTIME_VERSION
   

2. Connect to the TPU VM using SSH

   gcloud compute tpus tpu-vm ssh $TPU_NAME --zone=$ZONE
   

3. Ensure your Google Cloud user has been granted the Artifact Registry Reader role. For more information, see Granting Artifact Registry roles.

4. Start a container in the TPU VM using the nightly PyTorch/XLA image

   sudo docker run --net=host -ti --rm --name your-container-name --privileged \
   us-central1-docker.pkg.dev/tpu-pytorch-releases/docker/xla:r2.6.0_3.10_tpuvm_cxx11 \
   bash
   

   Note: After running this command, your command prompt will change indicating the terminal is connected to the running container.

5. Configure TPU runtime

   There are two PyTorch/XLA runtime options: PJRT and XRT. We recommend you use PJRT unless you have a reason to use XRT. To learn more about the different runtime configurations, see the PJRT runtime documentation.

   PJRT

   export PJRT_DEVICE=TPU
   

   XRT

   export XRT_TPU_CONFIG="localservice;0;localhost:51011"
   

6. Clone the PyTorch XLA repository

   git clone --recursive https://github.com/pytorch/xla.git
   

7. Train ResNet50

   python3 xla/test/test_train_mp_imagenet.py \
   --fake_data \
   --model=resnet50 \
   --num_epochs=1
   

When the training script completes, clean up the resources.

1. Type exit to exit from the Docker container
2. Type exit to exit from the TPU VM

3. Delete the TPU VM

   gcloud compute tpus tpu-vm delete $TPU_NAME --zone=$ZONE
   

TPU slice

When you run PyTorch code on a TPU slice, you must run your code on all TPU workers at the same time. One way to do this is to use the gcloud compute tpus tpu-vm ssh command with the --worker=all and --command flags. The following procedure shows you how to create a Docker image to make setting up each TPU worker easier.

1. Create a TPU VM

   gcloud compute tpus tpu-vm create $TPU_NAME \
   --zone=$ZONE \
   --accelerator-type=$ACCELERATOR_TYPE \
   --version=$RUNTIME_VERSION
   

2. Add the current user to the Docker group

   gcloud compute tpus tpu-vm ssh $TPU_NAME \
   --zone=$ZONE \
   --worker=all \
   --command='sudo usermod -a -G docker $USER'
   

3. Clone the PyTorch XLA repository

   gcloud compute tpus tpu-vm ssh $TPU_NAME --worker=all \
   --zone=$ZONE \
   --command="git clone --recursive https://github.com/pytorch/xla.git"
   

4. Run the training script in a container on all TPU workers

   gcloud compute tpus tpu-vm ssh $TPU_NAME --worker=all \
   --zone=$ZONE \
   --command="docker run --rm --privileged --net=host  -v ~/xla:/xla -e PJRT_DEVICE=TPU us-central1-docker.pkg.dev/tpu-pytorch-releases/docker/xla:r2.6.0_3.10_tpuvm_cxx11 python /xla/test/test_train_mp_imagenet.py --fake_data --model=resnet50 --num_epochs=1"
   

   Docker command flags:

   • --rm removes the container after its process terminates.
   • --privileged exposes the TPU device to the container.
   • --net=host binds all of the container's ports to the TPU VM to allow communication between the hosts in the pod.
   • -e sets environment variables.

When the training script completes, clean up the resources.

Delete the TPU VM using the following command:

gcloud compute tpus tpu-vm delete $TPU_NAME \
--zone=$ZONE

TPU device

1. Create the TPU VM

   gcloud compute tpus tpu-vm create $TPU_NAME \
   --zone=$ZONE \
   --accelerator-type=$ACCELERATOR_TYPE \
   --version=$RUNTIME_VERSION
   

2. Connect to the TPU VM using SSH

   gcloud compute tpus tpu-vm ssh $TPU_NAME --zone=$ZONE
   

3. Start Docker daemon in TPU VM

   sudo systemctl start docker
   

4. Start Docker container

   sudo docker run --net=host -ti --rm --name your-container-name \
   --privileged us-central1-docker.pkg.dev/tpu-pytorch-releases/docker/xla:r2.6.0_3.10_tpuvm_cxx11 \
   bash
   

5. Install JAX

   pip install jax[tpu]
   

6. Install FLAX

   pip install --upgrade clu
   git clone https://github.com/google/flax.git
   pip install --user -e flax
   

7. Install tensorflow and tensorflow-datasets packages

   pip install tensorflow
   pip install tensorflow-datasets
   

8. Run the FLAX MNIST training script

   cd flax/examples/mnist
   python3 main.py --workdir=/tmp/mnist \
   --config=configs/default.py \
   --config.learning_rate=0.05 \
   --config.num_epochs=5
   

When the training script completes, clean up the resources.

1. Type exit to exit from the Docker container
2. Type exit to exit from the TPU VM

3. Delete the TPU VM

   gcloud compute tpus tpu-vm delete $TPU_NAME --zone=$ZONE
   

TPU slice

When you run JAX code on a TPU slice, you must run your JAX code on all TPU workers at the same time. One way to do this is to use the gcloud compute tpus tpu-vm ssh command with the --worker=all and --command flags. The following procedure shows you how to create a Docker image to make setting up each TPU worker easier.

1. Create a file named Dockerfile in your current directory and paste the following text

   FROM python:3.10
   RUN pip install jax[tpu]
   RUN pip install --upgrade clu
   RUN git clone https://github.com/google/flax.git
   RUN pip install --user -e flax
   RUN pip install tensorflow
   RUN pip install tensorflow-datasets
   WORKDIR ./flax/examples/mnist
   

2. Prepare an Artifact Registry

   gcloud artifacts repositories create your-repo \
   --repository-format=docker \
   --location=europe-west4 --description="Docker repository" \
   --project=$PROJECT_ID
   
   gcloud artifacts repositories list \
   --project=$PROJECT_ID
   
   gcloud auth configure-docker europe-west4-docker.pkg.dev
   

3. Build the Docker image

   docker build -t your-image-name .
   

4. Add a tag to your Docker image before pushing it to the Artifact Registry. For more information on working with Artifact Registry, see Work with container images.

   docker tag your-image-name europe-west4-docker.pkg.dev/$PROJECT_ID/your-repo/your-image-name:your-tag
   

5. Push your Docker image to the Artifact Registry

   docker push europe-west4-docker.pkg.dev/$PROJECT_ID/your-repo/your-image-name:your-tag
   

6. Create a TPU VM

   gcloud compute tpus tpu-vm create $TPU_NAME \
   --zone=$ZONE \
   --accelerator-type=$ACCELERATOR_TYPE \
   --version=$RUNTIME_VERSION
   

7. Pull the Docker image from the Artifact Registry on all TPU workers

   gcloud compute tpus tpu-vm ssh $TPU_NAME --worker=all \
   --zone=$ZONE \
   --command='sudo usermod -a -G docker ${USER}'
   

   gcloud compute tpus tpu-vm ssh $TPU_NAME --worker=all \
   --zone=$ZONE \
   --command="gcloud auth configure-docker europe-west4-docker.pkg.dev --quiet"
   

   gcloud compute tpus tpu-vm ssh $TPU_NAME --worker=all \
   --zone=$ZONE \
   --command="docker pull europe-west4-docker.pkg.dev/$PROJECT_ID/your-repo/your-image-name:your-tag"
   

8. Run the container on all TPU workers

   gcloud compute tpus tpu-vm ssh $TPU_NAME --worker=all \
   --zone=$ZONE \
   --command="docker run -ti -d --privileged --net=host --name your-container-name europe-west4-docker.pkg.dev/$PROJECT_ID/your-repo/your-image-name:your-tag bash"
   

9. Run the training script on all TPU workers

   gcloud compute tpus tpu-vm ssh $TPU_NAME --worker=all \
   --zone=$ZONE \
   --command="docker exec --privileged your-container-name python3 main.py --workdir=/tmp/mnist \
   --config=configs/default.py \
   --config.learning_rate=0.05 \
   --config.num_epochs=5"
   

When the training script completes, clean up the resources.

1. Shut down the container on all workers

   gcloud compute tpus tpu-vm ssh $TPU_NAME --worker=all \
   --zone=$ZONE \
   --command="docker kill your-container-name"
   

2. Delete the TPU VM

   gcloud compute tpus tpu-vm delete $TPU_NAME \
   --zone=$ZONE