Warning
XPK is officially deprecated and in maintenance mode.
We are transitioning all infrastructure provisioning and workload management to Cluster Toolkit (CT).
XPK will receive critical bug fixes and maintenance support through Q3 2026. After this transition period, the repository will be archived. Please note that support for new hardware accelerators (upcoming TPU/GPU generations) will only be added to Cluster Toolkit.
Clusters previously provisioned via XPK will continue to run normally, but we strongly recommend planning a maintenance window to recreate your infrastructure using Cluster Toolkit.
Why the change? We are evolving from XPK's imperative CLI to Cluster Toolkit's declarative blueprint model. This provides a more robust, customizable, and standardized approach to defining AI/ML clusters using Terraform and Packer. Major frameworks like MaxText and Orbax are actively transitioning their deployment standards to CT.
- Migration guide: XPK to Cluster Toolkit Migration Guide
- Migration Q&A: Have questions about mapping XPK commands to CT? Join the discussion in GitHub Issues
- Repository: GoogleCloudPlatform/cluster-toolkit
- Examples: Check the CT Examples Directory for tested GKE and storage blueprints
- Documentation: Read the Cluster Toolkit Overview for guided deployments.
- Workload Submission: Check the official Job Guide for details on workload submission and supported features.
XPK (Accelerated Processing Kit, pronounced x-p-k) is a command line interface that simplifies cluster creation and workload execution on Google Kubernetes Engine (GKE). XPK generates preconfigured, training-optimized clusters and allows easy workload scheduling without any Kubernetes expertise.
XPK is recommended for quick creation of GKE clusters for proofs of concepts and testing.
XPK decouples provisioning capacity from running jobs. There are two structures: clusters (provisioned VMs) and workloads (training jobs). Clusters represent the physical resources you have available. Workloads represent training jobs -- at any time some of these will be completed, others will be running and some will be queued, waiting for cluster resources to become available.
The ideal workflow starts by provisioning the clusters for all of the ML hardware you have reserved. Then, without re-provisioning, submit jobs as needed. By eliminating the need for re-provisioning between jobs, using Docker containers with pre-installed dependencies and cross-ahead of time compilation, these queued jobs run with minimal start times. Further, because workloads return the hardware back to the shared pool when they complete, developers can achieve better use of finite hardware resources. And automated tests can run overnight while resources tend to be underutilized.
XPK supports a variety of hardware accelerators.
| Accelerator | Type | Recipes |
|---|---|---|
| Ironwood | tpu7x | Run training workload with Ironwood and regular/gSC/DWS Calendar reservations using GCS Bucket storage Run training workload with Ironwood with flex-start using Filestore storage Run training workload with Ironwood and flex-start using Lustre storage |
| Trillium | v6e | Create Cluster Create Workload |
| TPU v5p | v5p | Create Cluster Create Workload |
| TPU v5e | v5e | Create Cluster Create Workload |
| TPU v4 | v4 | Create Cluster Create Workload |
| GPU A4X | gb200 | Create Cluster Create Workload |
| GPU A4 | b200 | Create Cluster Create Workload |
| GPU A3 Ultra | h200 | Create Cluster Create Workload |
| GPU A3 Mega | h100-mega | Create Cluster Create Workload |
| GPU A3 High | h100 | Create Cluster Create Workload |
| GPU A100 | A100 | Create Cluster Create Workload |
| CPU | n2-standard-32 | Create Cluster Create Workload |
XPK also supports the following Google Cloud Storage solutions:
| Storage Type | Documentation |
|---|---|
| Cloud Storage FUSE | docs |
| Filestore | docs |
| Parallelstore | docs |
| Block storage (Persistent Disk, Hyperdisk) | docs |
| Dependency | When used |
|---|---|
| Google Cloud SDK (gcloud) | always |
| kubectl | always (Auto-installed) |
| ClusterToolkit | Provisioning GPU clusters (Auto-installed) |
| Kueue | Scheduling workloads (Auto-installed) |
| JobSet | Workload creation (Auto-installed) |
| Crane | Building workload container (Auto-installed) |
| CoreDNS | Cluster set up (Auto-installed) |
To help improve XPK, feature usage statistics are collected and sent to Google. You can opt-out at any time by executing the following shell command:
xpk config set send-telemetry <true/false>XPK telemetry overall is handled in accordance with the Google Privacy Policy. When you use XPK to interact with or utilize GCP Services, your information is handled in accordance with the Google Cloud Privacy Notice.
Please read contributing.md for details on our code of conduct, and the process for submitting pull requests to us.
We'd love to hear from you! If you have questions or want to discuss ideas, join us on GitHub Discussions. Found a bug or have a feature request? Please let us know on GitHub Issues.
This project is licensed under the Apache License 2.0 - see the LICENSE file for details