tpu

TPU User Guide

This page contains instructions for how to set up Ray on GKE with TPUs.

Prerequisites

Please follow the official Google Cloud documentation for an introduction to TPUs. In partiuclar, please ensure that your GCP project has sufficient quotas to provision the cluster, see this link for details.

For addition useful information about TPUs on GKE (such as topology configurations and availability), see this page.

In addition, please ensure the following are installed on your local development environment:

• Helm (v3.9.3)
• Terraform (v1.7.4)
• Kubectl

Provisioning a GKE Cluster with Terraform (Optional)

Skip this section if you already have a GKE cluster with TPUs (cluster version should be 1.28 or later).

1. git clone https://github.com/GoogleCloudPlatform/ai-on-gke

2. cd ai-on-gke/infrastructure

3. Edit platform.tfvars with your GCP settings.

4. Change the region or zone to one where TPUs are available (see this link for details. For v4 TPUs (the default type), the region should be set to us-central2 or us-central2-b.

5. Set the following flags (note that TPUs are currently only supported on GKE standard):

autopilot_cluster = false
...
enable_tpu = true

6. Change the following lines in the tpu_pools configuration to match your desired TPU accelerator.

accelerator_count      = 2
accelerator_type       = "nvidia-tesla-t4"

7. Run terraform init && terraform apply -var-file platform.tfvars

Manually Installing the TPU Initialization Webhook

The TPU Initialization Webhook automatically bootstraps the TPU environment for TPU clusters. The webhook needs to be installed once per GKE cluster and requires a Kuberay Operator running v1.1+ and GKE cluster version of 1.28+. The webhook requires cert-manager to be installed in-cluster to handle TLS certificate injection. cert-manager can be installed in both GKE standard and autopilot clusters using the following helm commands:

helm repo add jetstack https://charts.jetstack.io
helm repo update
helm install --create-namespace --namespace cert-manager --set installCRDs=true --set global.leaderElection.namespace=cert-manager cert-manager jetstack/cert-manager

After installing cert-manager, it may take up to two minutes for the certificate to become ready.

Installing the webhook:

1. git clone https://github.com/GoogleCloudPlatform/ai-on-gke
2. cd ai-on-gke/ray-on-gke/tpu/kuberay-tpu-webhook
3. make deploy
   • this will create the webhook deployment, configs, and service in the "ray-system" namespace
   • to change the namespace, edit the "namespace" value in each .yaml in deployments/ and certs/
4. make deploy-cert

For common errors encountered when deploying the webhook, see the Troubleshooting guide.

Creating the Kuberay Cluster

You can find sample TPU cluster manifests for single-host and multi-host here.

If you are using Terraform:

1. Get the GKE cluster name and location/region from infrastructure/platform.tfvars. Run gcloud container clusters get-credentials %gke_cluster_name% --location=%location%. Configuring gcloud instructions

2. cd ../applications/ray

3. Edit workloads.tfvars with your GCP settings. Replace <your project ID> and <your cluster name> with the names you used in platform.tfvars.

4. Run terraform init && terraform apply -var-file workloads.tfvars

This should deploy a Kuberay cluster with a single TPU worker node (v4 TPU with 2x2x1 topology).

To deploy a multi-host Ray Cluster, modify the worker spec here by changing the cloud.google.com/gke-tpu-topology nodeSelector to a multi-host topology. Set the numOfHosts field in the worker spec to the number of hosts specified by your chosen topology. For v4 TPUs, each TPU VM has access to 4 TPU chips. Therefore, you can calculate the number of TPU VM hosts by taking the product of the topology and dividing by 4 (i.e. a 2x2x4 TPU podslice will have 4 TPU VM hosts).

Running Sample Workloads

1. Save the following to a local file (e.g. test_tpu.py):

import ray

ray.init(
    address="ray://ray-cluster-kuberay-head-svc:10001",
    runtime_env={
        "pip": [
            "jax[tpu]==0.4.12",
            "-f https://storage.googleapis.com/jax-releases/libtpu_releases.html",
        ]
    }
)


@ray.remote(resources={"TPU": 4})
def tpu_cores():
    import jax
    return "TPU cores:" + str(jax.device_count())

num_workers = 4
result = [tpu_cores.remote() for _ in range(num_workers)]
print(ray.get(result))

2. kubectl port-forward svc/ray-cluster-kuberay-head-svc 8265:8265 &
3. export RAY_ADDRESS=http://localhost:8265
4. ray job submit --runtime-env-json='{"working_dir": "."}' -- python test_tpu.py

For a more advanced workload running Stable Diffusion on TPUs, see here.