AI Platform Training uses container images, based on runtime version designation, to configure cloud resources to service your training and prediction requests. This page lists the runtime versions and their constituent packages. For more information, see Managing runtime versions.
Supported AI Platform Training runtime versions
The following versions are supported in AI Platform Training:
Version | Package | Released On | Last Updated | End of availability | ||||
---|---|---|---|---|---|---|---|---|
2.11 | TensorFlow
2.11.x (latest patch) scikit-learn 1.0.2 XGBoost 1.6.1 GPUs are supported for online prediction in this runtime version. This runtime version also supports TPUs for training. Runtime version 2.11 does not support batch prediction. Python 3.7 is the only version of Python available for online prediction with runtime version 2.11. You cannot use Python 2 with runtime version 2.11. Package lists
|
January 23, 2023 | January 23, 2023 |
Starting on January 23, 2024, you can no longer create training jobs, batch prediction jobs, or model versions that use this runtime version. On January 23, 2025, AI Platform Prediction deletes all your model versions that use this runtime version. |
||||
2.9 | TensorFlow
2.9.x (latest patch) scikit-learn 1.0.2 XGBoost 1.6.1 GPUs are supported for online prediction in this runtime version. This runtime version also supports TPUs for training. Runtime version 2.9 does not support batch prediction. Python 3.7 is the only version of Python available for online prediction with runtime version 2.9. You cannot use Python 2 with runtime version 2.9. Package lists
|
October 26, 2022 | October 26, 2022 |
Starting on October 26, 2023, you can no longer create training jobs, batch prediction jobs, or model versions that use this runtime version. On October 26, 2024, AI Platform Prediction deletes all your model versions that use this runtime version. |
||||
2.8 | TensorFlow
2.8.x (latest patch) scikit-learn 1.0.2 XGBoost 1.5.2 GPUs are supported for online prediction in this runtime version. This runtime version also supports TPUs for training. Runtime version 2.8 does not support batch prediction. Python 3.7 is the only version of Python available for online prediction with runtime version 2.8. You cannot use Python 2 with runtime version 2.8. Package lists
|
February 15, 2022 | February 15, 2022 |
Starting on March 16, 2023, you can no longer create training jobs, batch prediction jobs, or model versions that use this runtime version. On March 16, 2024, AI Platform Prediction deletes all your model versions that use this runtime version. |
||||
2.7 | TensorFlow
2.7.x (latest patch) scikit-learn 1.0.1 XGBoost 1.5.0 GPUs are supported for online prediction in this runtime version. This runtime version also supports TPUs for training. Runtime version 2.7 does not support batch prediction. Python 3.7 is the only version of Python available for online prediction with runtime version 2.7. You cannot use Python 2 with runtime version 2.7. Package lists
|
December 8, 2021 | December 8, 2021 |
Starting on December 8, 2022, you can no longer create training jobs, batch prediction jobs, or model versions that use this runtime version. On December 8, 2023, AI Platform Prediction deletes all your model versions that use this runtime version. |
||||
2.6 | TensorFlow
2.6.x (latest patch) scikit-learn 0.24.2 XGBoost 1.4.2 GPUs are supported for online prediction in this runtime version. This runtime version also supports TPUs for training. Runtime version 2.6 does not support batch prediction. Python 3.7 is the only version of Python available for online prediction with runtime version 2.6. You cannot use Python 2 with runtime version 2.6. Package lists
|
October 6, 2021 | October 6, 2021 |
Starting on October 6, 2022, you can no longer create training jobs, batch prediction jobs, or model versions that use this runtime version. On October 6, 2023, AI Platform Prediction deletes all your model versions that use this runtime version. |
||||
2.5 | TensorFlow
2.5.x (latest patch) scikit-learn 0.24.2 XGBoost 1.4.0 GPUs are supported for training and online prediction in this runtime version. This runtime version also supports TPUs for training. Runtime version 2.5 does not support batch prediction. Python 3.7 is the only version of Python available for online prediction with runtime version 2.5. You cannot use Python 2 with runtime version 2.5. Package lists
|
August 13, 2021 | August 13, 2021 |
Starting on August 13, 2022, you can no longer create training jobs, batch prediction jobs, or model versions that use this runtime version. On August 13, 2023, AI Platform Prediction deletes all your model versions that use this runtime version. |
||||
2.4 | TensorFlow
2.4.x (latest patch) scikit-learn 0.24.0 XGBoost 1.3.1 GPUs are supported for training and online prediction in this runtime version. This runtime version also supports TPUs for training. Runtime version 2.4 does not support batch prediction. Python 3.7 is the only version of Python available for training and online prediction with runtime version 2.4. You cannot use Python 2 with runtime version 2.4. Package lists
|
February 10, 2021 | February 10, 2021 |
Starting on April 16, 2022, you can no longer create training jobs, batch prediction jobs, or model versions that use this runtime version. On April 16, 2023, AI Platform Prediction deletes all your model versions that use this runtime version. |
||||
2.3 | TensorFlow
2.3.x (latest patch) scikit-learn 0.23.2 XGBoost 1.2.1 GPUs are supported for training and online prediction in this runtime version. This runtime version also supports TPUs for training. Runtime version 2.3 does not support batch prediction. Python 3.7 is the only version of Python available for training and online prediction with runtime version 2.3. You cannot use Python 2 with runtime version 2.3. Package lists
|
December 9, 2020 | December 9, 2020 |
Starting on December 9, 2021, you can no longer create training jobs, batch prediction jobs, or model versions that use this runtime version. On December 9, 2022, AI Platform Prediction deletes all your model versions that use this runtime version. |
||||
2.2 | TensorFlow
2.2.x (latest patch) scikit-learn 0.23.1 XGBoost 1.1.1 GPUs are supported for training and online prediction in this runtime version. This runtime version also supports TPUs for training. Runtime version 2.2 does not support batch prediction. Python 3.7 is the only version of Python available for training and online prediction with runtime version 2.2. You cannot use Python 2 with runtime version 2.2. Package lists
|
August 28, 2020 | August 28, 2020 |
Starting on August 28, 2021, you can no longer create training jobs, batch prediction jobs, or model versions that use this runtime version. On August 28, 2022, AI Platform Prediction deletes all your model versions that use this runtime version. |
||||
2.1 | TensorFlow 2.1.0 scikit-learn 0.22.1 XGBoost 0.90 Runtime version 2.1 supports TensorFlow 2.1.0 for CPU and GPU. GPUs are supported for training and online prediction in this runtime version. This runtime version also supports TPUs for training. Python 3.7 is the only version of Python available for training and online prediction with runtime version 2.1. You cannot use Python 2 with runtime version 2.1. Package lists
|
March 9, 2020 | March 9, 2020 |
Starting on March 9, 2021, you can no longer create training jobs that use this runtime version or model versions that use this runtime version. Batch Prediction will continue to be supported. On January 31, 2023, AI Platform Prediction deletes all your model versions that use this runtime version. |
||||
1.15 | TensorFlow 1.15.0 scikit-learn 0.20.4 XGBoost 0.82 Runtime version 1.15 supports TensorFlow 1.15.0 for CPU and GPU. GPUs are supported for training and online prediction in this runtime version. This runtime version also supports TPUs for training. Python 3.7 is available for training and online prediction with runtime version 1.15. The Ubuntu packages for Python 3 (indicated in bold) are installed when running Python 3. Package lists
|
December 19, 2019 | December 19, 2019 |
Starting on September 30, 2022, you can no longer create training jobs, batch prediction jobs, or model versions that use this runtime version. On September 30, 2023, AI Platform Prediction deletes all your model versions that use this runtime version. |
||||
1.14 | TensorFlow 1.14.0 scikit-learn 0.20.2 XGBoost 0.81 Runtime version 1.14 supports TensorFlow 1.14.0 for CPU and GPU. GPUs are supported for training and online prediction in this runtime version. This runtime version also supports TPUs for training. Python 3.5 is available for training and online prediction with runtime version 1.14. The Ubuntu packages for Python 3 (indicated in bold) are installed when running Python 3. Package lists
|
July 19, 2019 | August 16, 2019 |
Starting on July 19, 2020, you can no longer create training jobs, batch prediction jobs, or model versions that use this runtime version. On July 19, 2021, AI Platform Prediction deletes all your model versions that use this runtime version. |
||||
1.13 | TensorFlow 1.13.1 scikit-learn 0.20.2 XGBoost 0.81 Runtime version 1.13 supports TensorFlow 1.13.1 for CPU and GPU. GPUs are supported for training and online prediction in this runtime version. This runtime version also supports TPUs for training. Python 3.5 is available for training and online prediction with runtime version 1.13. The Ubuntu packages for Python 3 (indicated in bold) are installed when running Python 3. Package lists
|
March 6, 2019 | August 16, 2019 |
Starting on July 19, 2020, you can no longer create training jobs, batch prediction jobs, or model versions that use this runtime version. On March 6, 2021, AI Platform Prediction deletes all your model versions that use this runtime version. |
||||
1.12 | TensorFlow 1.12.3 scikit-learn 0.20.0 XGBoost 0.81 Runtime version 1.12 supports TensorFlow 1.12.3 for CPU and GPU. GPUs are supported for training and online prediction in this runtime version. Python 3.5 is available for training and online prediction with runtime version 1.12. The Ubuntu packages for Python 3 (indicated in bold) are installed when running Python 3. Package lists
|
December 19, 2018 | July 19, 2019 |
Starting on July 19, 2020, you can no longer create training jobs, batch prediction jobs, or model versions that use this runtime version. On December 19, 2020, AI Platform Prediction deletes all your model versions that use this runtime version. |
||||
1.11 | TensorFlow 1.11 scikit-learn 0.19.2 XGBoost 0.80 Runtime version 1.11 supports TensorFlow 1.11.0 for CPU and GPU. GPUs are supported for training and online prediction in this runtime version. Python 3.5 is available for training and online prediction with runtime version 1.11. The Ubuntu packages for Python 3 (indicated in bold) are installed when running Python 3. Package lists
|
December 19, 2018 |
Starting on July 19, 2020, you can no longer create training jobs, batch prediction jobs, or model versions that use this runtime version. On December 19, 2020, AI Platform Prediction deletes all your model versions that use this runtime version. |
|||||
1.10 | TensorFlow 1.10 scikit-learn 0.19.2 XGBoost 0.72.1 Runtime version 1.10 supports TensorFlow 1.10.0 for CPU and GPU (GPUs are not supported for online prediction). Python 3.5 is available for training and online prediction with runtime version 1.10. The Ubuntu packages for Python 3 (indicated in bold) are installed when running Python 3. Package lists
|
August 31, 2018 | December 19, 2018 |
Starting on (date not set), you can no longer create training jobs, batch prediction jobs, or model versions that use this runtime version. On August 31, 2020, AI Platform Prediction deletes all your model versions that use this runtime version. |
||||
1.9 | TensorFlow 1.9 scikit-learn 0.19.1 XGBoost 0.72.1 Runtime version 1.9 supports TensorFlow 1.9.0 for CPU and GPU (GPUs are not supported for online prediction). Python 3.5 is available for training and online prediction with runtime version 1.9. The Ubuntu packages for Python 3 (indicated in bold) are installed when running Python 3. Package lists
|
June 27, 2018 | December 19, 2018 |
Starting on March 16, 2020, you can no longer create training jobs that use this runtime version. Starting on (date not set), you can no longer create batch prediction jobs or model versions that use this runtime version. On June 27, 2020, AI Platform Prediction deletes all your model versions that use this runtime version. |
||||
1.8 | TensorFlow 1.8 scikit-learn 0.19.1 XGBoost 0.71 Runtime version 1.8 supports TensorFlow 1.8.0 for CPU and GPU (GPUs are not supported for online prediction). Python 3.5 is available for training and online prediction with runtime version 1.8. The Ubuntu packages for Python 3 (indicated in bold) are installed when running Python 3. The Package lists
|
May 8, 2018 | December 19, 2018 |
Starting on March 16, 2020, you can no longer create training jobs that use this runtime version. Starting on (date not set), you can no longer create batch prediction jobs or model versions that use this runtime version. On May 8, 2020, AI Platform Prediction deletes all your model versions that use this runtime version. |
||||
1.7 | TensorFlow 1.7 scikit-learn 0.19.1 XGBoost 0.7.post3 Runtime version 1.7 supports TensorFlow 1.7.0 for CPU and GPU. Python 3.5 is available for training and online prediction with runtime version 1.7. The Ubuntu packages for Python 3 (indicated in bold) are installed when running Python 3. Package lists
|
April 26, 2018 | December 19, 2018 |
Starting on March 16, 2020, you can no longer create training jobs that use this runtime version. Starting on (date not set), you can no longer create batch prediction jobs or model versions that use this runtime version. On April 26, 2020, AI Platform Prediction deletes all your model versions that use this runtime version. |
||||
1.6 | TensorFlow 1.6 scikit-learn 0.19.1 XGBoost 0.7.post3 Runtime version 1.6 supports TensorFlow 1.6.0 for CPU and GPU. Python 3.5 is available for training and online prediction with runtime version 1.6. The Ubuntu packages for Python 3 (indicated in bold) are installed when running Python 3. Package lists
|
March 20, 2018 | December 19, 2018 |
Starting on March 16, 2020, you can no longer create training jobs that use this runtime version. Starting on (date not set), you can no longer create batch prediction jobs or model versions that use this runtime version. On April 13, 2020, AI Platform Prediction deletes all your model versions that use this runtime version. |
||||
1.5 | TensorFlow 1.5 scikit-learn 0.19.1 XGBoost 0.7.post3 Runtime version 1.5 supports TensorFlow 1.5.0 for CPU and GPU. Python 3.5 is available for training and online prediction with runtime version 1.5. The Ubuntu packages for Python 3 (indicated in bold) are installed when running Python 3. Package lists
|
March 13, 2018 | December 19, 2018 |
Starting on March 16, 2020, you can no longer create training jobs that use this runtime version. Starting on (date not set), you can no longer create batch prediction jobs or model versions that use this runtime version. On April 13, 2020, AI Platform Prediction deletes all your model versions that use this runtime version. |
||||
1.4 | TensorFlow 1.4.0 and 1.4.1 scikit-learn 0.18.1 XGBoost 0.6a2 Runtime version 1.4 uses TensorFlow 1.4.0 for online prediction, and 1.4.1 for batch prediction and training. Python 3.5 is available for training and online prediction with runtime version 1.4. The Ubuntu packages for Python 3 (indicated in bold) are installed when running Python 3. The earliest AI Platform Training runtime version that provides support for scikit-learn and XGBoost is version 1.4. Package lists
|
December 11, 2017 | December 19, 2018 |
Starting on March 16, 2020, you can no longer create training jobs that use this runtime version. Starting on (date not set), you can no longer create batch prediction jobs or model versions that use this runtime version. On April 13, 2020, AI Platform Prediction deletes all your model versions that use this runtime version. |
||||
1.2 | TensorFlow 1.2.0 Runtime version 1.2 uses an Ubuntu 16.04 OS base image instead of the Debian Jessie version used by 1.0. Package lists
|
June 27, 2017 |
Starting on March 16, 2020, you can no longer create training jobs that use this runtime version. Starting on (date not set), you can no longer create batch prediction jobs or model versions that use this runtime version. On April 13, 2020, AI Platform Prediction deletes all your model versions that use this runtime version. |
Policy for supporting older runtime versions
AI Platform Training and AI Platform Prediction support each runtime version for one year after its release date.
Support for each runtime version changes according to the following schedule:
Starting on the release date: You can create training jobs, batch prediction jobs, and model versions that use the runtime version.
Starting 12 months after the release date: You can no longer create training jobs, batch prediction jobs, or model versions that use the runtime version.
Existing model versions that have been deployed to AI Platform Prediction continue to function.
Starting 24 months after the release date: AI Platform Prediction automatically deletes all model versions that use the runtime version.
A modified version of this policy is being applied retroactively, over several stages, to runtime versions 1.13 and earlier. Refer to this document for the current availability of each runtime version.
Support for GPUs
GPU-enabled machines come pre-installed with tensorflow-gpu, the TensorFlow Python package with GPU support.
Other machines come pre-installed with the regular tensorflow package instead.
Support for Cloud TPU
AI Platform Training runtime version(s) 1.15, 2.1, 2.2, 2.3, 2.4, 2.5, 2.6, 2.7, 2.8, 2.9, and 2.11 are available for training your models on Cloud TPU. See how to request Cloud TPU quota and how to use TPUs to train your model.
Pre-built PyTorch Containers
In addition to runtime versions, AI Platform Training offers several pre-built PyTorch containers that you can use to train a PyTorch model. Learn about the available PyTorch containers and how to use them.