This page documents production updates to all AI Platform products. You can periodically check this page for announcements about new or updated features, bug fixes, known issues, and deprecated functionality.
For more detailed information, read the documentation for each product.
You can see the latest product updates for all of Google Cloud on the Google Cloud page, browse and filter all release notes in the Google Cloud console, or programmatically access release notes in BigQuery.
To get the latest product updates delivered to you, add the URL of this page to your feed reader, or add the feed URL directly.
September 26, 2024
Deep Learning ContainersM125 release
- TensorFlow 2.17 container images are now available.
M125 release
- TensorFlow 2.17 VM images are now available.
August 20, 2024
Deep Learning ContainersM124 release
- Pytorch 2.3.0 with CUDA 12.1 and Python 3.10 container images are now available.
M124 release
- Pytorch 2.3.0 with CUDA 12.1 and Python 3.10 VM images are now available.
July 16, 2024
Deep Learning ContainersM123 release
- Hugging Face Text Generation Inference 2.1 GPU container images are now available.
M123 release
- TensorFlow 2.16 images are now available.
June 21, 2024
Deep Learning ContainersM122 release
- TensorFlow 2.16 container images are now available.
- PyTorch Inference 2.2 GPU container images are now available.
- PyTorch Inference 2.2 CPU container images are now available.
M122 release
- Updated Nvidia drivers to version 550.90.07 to fix vulnerabilities.
May 17, 2024
Deep Learning ContainersM121 release
- Updated the R CPU container image from R 4.3 to R 4.4. The R 4.3 container image is deprecated. There will be no further updates to this image in future releases.
M121 release
- CUDA 12.2 images are now available.
- Updated TensorFlow 2.15 images from CUDA 12.1 to CUDA 12.2.
- Re-enabled
common-gpu
Deep Learning VM releases that were erroneously deactivated in M117. - Updated Nvidia drivers to 550.54.15 to fix an issue where Nvidia drivers failed to install on startup after Debian 11 images upgraded kernel to
linux-image-5.10.0-29-cloud-amd64
. - The
linux-headers-cloud-amd64
metapackage is now installed for faster driver recompiling on kernel upgrades. - TensorFlow 2.6 CPU and GPU images are deprecated. There will be no further updates to these images in future releases.
April 25, 2024
Deep Learning ContainersM120 release
- Upgraded TensorFlow 2.15 container images to TensorFlow 2.15.1.
- Added CUDA-specific release tags for all TensorFlow and PyTorch container images, for example,
us-docker.pkg.dev/deeplearning-platform-release/gcr.io/tf-cu121.2-15
.
M120 release
- Upgraded TensorFlow 2.15 images to TensorFlow 2.15.1.
- Added Ubuntu 22.04 support for CPU images, and for GPU images using CUDA 12.1 or higher.
March 29, 2024
Deep Learning ContainersM119 release
- Fixed an issue wherein Dataproc extensions caused JupyterLab to crash when remote kernels weren't available.
M119 release
- Fixed an issue wherein Dataproc extensions caused JupyterLab to crash when remote kernels weren't available.
March 18, 2024
Deep Learning ContainersM118 release
- PyTorch 2.1.0 with CUDA 12.1 and Python 3.10 container images are now available.
- PyTorch 2.2.0 with CUDA 12.1 and Python 3.10 container images are now available.
M118 release
- Restored legacy gpu image families for TensorFlow 2.12 through 2.14, and for PyTorch 2.0.
- PyTorch 2.1.0 with CUDA 12.1 and Python 3.10 VM images are now available.
- PyTorch 2.2.0 with CUDA 12.1 and Python 3.10 VM images are now available.
- R images (Experimental) updated to R 4.3.3.
- Updated Nvidia drivers of older Deep Learning VM images to R535.
February 28, 2024
Deep Learning VM ImagesM117 release
- Added the CUDA version (CUDA 11.8) to the TensorFlow 2.12, 2.13, and 2.14 image names and image family names. For example,
tf-2-12-gpu
is renamedtf-2-12-cu118
.
February 20, 2024
Deep Learning ContainersM117 release
- Fixed an issue wherein the
latest
container had adeprecation-public-image
tag. In this release and future releases, this tag will only be on the deprecated containers. - Fixed a problem wherein the user couldn't access the vulnerabilities result of each container.
February 08, 2024
Deep Learning VM ImagesM116 release
- Added the CUDA version to the TensorFlow 2.15 image family name, for this release and future releases. For example,
tf-2-15-gpu
is renamed totf-2-15-cu121
. - Deprecated the
tf-2-15-gpu
image family in favor oftf-2-15-cu121
.
January 19, 2024
Deep Learning ContainersM115 release
- TensorFlow 2.15 with CUDA 12.1 and Python 3.10 container images are now available.
- TensorFlow 2.14 with CUDA 11.8 and Python 3.10 container images are now available.
M115 release
- TensorFlow 2.15 with CUDA 12.1 and Python 3.10 images are now available.
- TensorFlow 2.14 with CUDA 11.8 and Python 3.10 images are now available.
December 14, 2023
Deep Learning ContainersM114 release
- Starting with this release, Python 3.7 is no longer available.
- Upgraded R to 4.3 on Python 3.10 containers.
- Fixed an issue where the PySpark-BigQuery connector didn't work properly on Python 3.10 PySpark container.
M114 release
- Starting with this release, Debian 10 Python 3.7 images are no longer available.
- Upgraded R to 4.3 on Debian 11 Python 3.10 images.
November 16, 2023
Deep Learning ContainersM113 release
- Miscellaneous bug fixes and improvements in Python 3.10 container images.
M113 release
- Miscellaneous bug fixes and improvements in Python 3.10 images.
October 10, 2023
Deep Learning ContainersM112 release
- Miscellaneous bug fixes and improvements.
M112 release
- CUDA 12.1 VM images are available with the following image names:
common-cu121-debian-11-py310
common-cu121-ubuntu-2004-py310
- Miscellaneous bug fixes and improvements.
September 14, 2023
Deep Learning ContainersM111 release
- PyTorch 2.0 container images now include PyTorch XLA 2.0.
- Miscellaneous software updates.
M111 release
- PyTorch 2.0 images now include PyTorch XLA 2.0.
- Miscellaneous software updates.
August 10, 2023
Deep Learning ContainersM110 release
- Added support for TensorFlow 2.13 with Python 3.10 on Debian 11.
- Added support for TensorFlow 2.8 with Python 3.10 on Debian 11.
- Miscellaneous software updates.
TensorFlow 2.9 container images are deprecated.
M110 release
- Added support for TensorFlow 2.13 with Python 3.10 on Debian 11.
- Added support for TensorFlow 2.8 with Python 3.10 on Debian 11.
- Miscellaneous software updates.
TensorFlow 2.9 images are deprecated.
July 31, 2023
AI Platform PipelinesThis legacy version of AI Platform Pipelines is deprecated and will no longer be available on Google Cloud after July 31, 2024. All the functionality of legacy AI Platform Pipelines and new features are available on the Vertex AI platform. Migrate your resources to Vertex AI Pipelines to get a managed orchestration engine for Kubeflow Pipelines (KFP) and many additional features.
This legacy version of AI Platform Prediction is deprecated and will no longer be available on Google Cloud after January 31, 2025. Migrate your resources to Vertex AI to get new machine learning features that are unavailable in AI Platform.
This legacy version of AI Platform Training is deprecated and will no longer be available on Google Cloud after January 31, 2025. Migrate your resources to Vertex AI custom training to get new machine learning features that are unavailable in AI Platform.
June 26, 2023
Deep Learning ContainersM109 release
- PyTorch 2.0 with Python 3.10 and CUDA 11.8 container images are now available.
- Miscellaneous software updates.
M109 release
- PyTorch 2.0 on Debian 11 with Python 3.10 and CUDA 11.8 images are now available.
- GPU-based Deep Learning VM Images now installs Nvidia drivers with the new open kernel modules if started on an A2 or G2 machine instead of the proprietary kernel modules.
- Miscellaneous software updates.
May 09, 2023
Deep Learning ContainersM108 update
This update of the M108 release includes the following:
- The following Deep Learning Containers images are now available:
- Tensorflow 2.12 CPU with CUDA 11.8 and Python 3.10 (
us-docker.pkg.dev/deeplearning-platform-release/gcr.io/tf-cpu.2-12.py310:latest
) - Tensorflow 2.12 GPU with CUDA 11.8 and Python 3.10 (
us-docker.pkg.dev/deeplearning-platform-release/gcr.io/tf-gpu.2-12.py310:latest
)
- Tensorflow 2.12 CPU with CUDA 11.8 and Python 3.10 (
M108 update
This update of the M108 release includes the following:
- The following Deep Learning VM images are now available:
- Tensorflow 2.12 CPU with CUDA 11.8 and Python 3.10 (
tf-2-12-cpu-debian-11-py310
) - Tensorflow 2.12 GPU with CUDA 11.8 and Python 3.10 (
tf-2-12-gpu-debian-11-py310
)
- Tensorflow 2.12 CPU with CUDA 11.8 and Python 3.10 (
May 04, 2023
Deep Learning ContainersM108 release
- Miscellaneous software updates.
M108 release
- The image name
common-container-experimental
was changed tocommon-container
. The related image family name wasn't changed. - Miscellaneous software updates.
April 13, 2023
Deep Learning ContainersM107 release
- Miscellaneous software updates.
M107 release
- Miscellaneous software updates.
April 06, 2023
Deep Learning ContainersM106 release
- Miscellaneous software updates.
M106 release
- Rolled back a previous change in which Jupyter dependencies were located in a separate Conda environment.
- Miscellaneous software updates.
March 31, 2023
Deep Learning ContainersM105 release
The following Deep Learning Containers images are now available with Python 3.10 on Debian 11:
- TensorFlow 2.11 CPU (
us-docker.pkg.dev/deeplearning-platform-release/gcr.io/tf-cpu.2-11.py310:latest
) - TensorFlow 2.11 GPU with Cuda 11.3 (
us-docker.pkg.dev/deeplearning-platform-release/gcr.io/tf-gpu.2-11.py310:latest
) - PyTorch 1.13 with Cuda 11.3 (
us-docker.pkg.dev/deeplearning-platform-release/gcr.io/pytorch-gpu.1-13.py310:latest
) - Base CPU (
us-docker.pkg.dev/deeplearning-platform-release/gcr.io/base-cpu.py310:latest
) - Base GPU with Cuda 11.3 (
us-docker.pkg.dev/deeplearning-platform-release/gcr.io/base-cu113.py310:latest
)
- TensorFlow 2.11 CPU (
The following Deep Learning Containers images are now available with Python 3.9 on Debian 11:
- TensorFlow 2.6 CPU (
us-docker.pkg.dev/deeplearning-platform-release/gcr.io/tf-cpu.2-6.py39:latest
) - TensorFlow 2.6 GPU with Cuda 11.3 (
us-docker.pkg.dev/deeplearning-platform-release/gcr.io/tf-gpu.2-6.py39:latest
)
- TensorFlow 2.6 CPU (
Miscellaneous bug fixes and improvements.
M105 release
The following Deep Learning VM images are now available with Python 3.10 on Debian 11:
- TensorFlow 2.11 CPU (
tf-2-11-cpu-debian-11-py310
) - TensorFlow 2.11 GPU with Cuda 11.3 (
tf-2-11-cu113-debian-11-py310
) - PyTorch 1.13 with Cuda 11.3 (
pytorch-1-13-cu113-debian-11-py310
) - Base CPU (
common-cpu-debian-11-py310
) - Base GPU with Cuda 11.3 (
common-cu113-debian-11-py310
)
- TensorFlow 2.11 CPU (
The following Deep Learning VM images are now available with Python 3.9 on Debian 11:
- TensorFlow 2.6 CPU (
tf-2-6-cpu-debian-11-py39
) - TensorFlow 2.6 GPU with Cuda 11.3 (
tf-2-6-cu113-debian-11-py39
)
- TensorFlow 2.6 CPU (
Jupyter-related libraries have been moved to a different Conda environment, separate from the one containing machine learning frameworks and base software libraries.
Miscellaneous bug fixes and improvements.
March 16, 2023
Deep Learning ContainersM104 release
- Added the following packages:
- google-cloud-artifact-registry
- google-cloud-bigquery-storage
- google-cloud-language
- keyring
- keyrings.google-artifactregistry-auth
- Fixed a bug in which curl could not find the right SSL certificate path by default.
TensorFlow Enterprise 2.1 has reached the end of its support period. See Version details.
M104 release
- Added the following packages:
- google-cloud-artifact-registry
- google-cloud-bigquery-storage
- google-cloud-language
- keyring
- keyrings.google-artifactregistry-auth
- Fixed a bug in which curl could not find the right SSL certificate path by default.
TensorFlow Enterprise 2.1 has reached the end of its support period. See Version details.
January 30, 2023
Deep Learning ContainersM103 release
- Upgraded PyTorch to 1.13.1.
- Minor bug fixes and improvements.
M103 release
- Upgraded PyTorch to 1.13.1.
- Minor bug fixes and improvements.
January 24, 2023
AI Platform TrainingRuntime version 2.11 is available. You can use runtime version 2.11 to train with TensorFlow 2.11, scikit-learn 1.0.2, or XGBoost 1.6.1. Runtime version 2.11 supports training with CPUs, GPUs, or TPUs.
See the full list of updated dependencies in runtime version 2.11.
January 23, 2023
AI Platform Data Labeling ServiceAI Hub and the legacy versions of the following products are deprecated and will no longer be available on Google Cloud after January 17, 2024:
All the functionality of these legacy versions and new features are available on the Vertex AI platform. See Migrate to Vertex AI to learn how to migrate your resources.
December 15, 2022
Deep Learning ContainersM102 release
- TensorFlow 2.11 is now available.
- PyTorch 1.13 is now available.
- Regular security patches and package upgrades.
M102 release
- TensorFlow 2.11 is now available.
- PyTorch 1.13 is now available.
- Added support for Jupyter[Lab] Language Server Protocol.
- Regular security patches and package upgrades.
December 09, 2022
Deep Learning ContainersM101 release
- TensorFlow patch version upgrades:
- From 2.8.3 to 2.8.4.
- From 2.9.2 to 2.9.3.
- From 2.10.0 to 2.10.1.
- TensorFlow 1.15 Deep Learning Containers images are now deprecated.
- Regular security patches and package upgrades.
M101 release
- TensorFlow patch version upgrades:
- From 2.8.3 to 2.8.4.
- From 2.9.2 to 2.9.3.
- From 2.10.0 to 2.10.1.
- TensorFlow 1.15 Deep Learning VM images are now deprecated.
- Regular security patches and package upgrades.
November 08, 2022
Deep Learning ContainersM100 release
- Regular package updates.
M100 release
- Migrated the Docker proxy agent to use a systemctl service.
- Regular package updates.
November 02, 2022
Deep Learning ContainersM99 release
- Fixed a bug where Jupyter widgets through
ipywidgets
were causing errors and not displaying. - Regular package updates.
M99 release
- Fixed a bug where Jupyter widgets through
ipywidgets
were causing errors and not displaying. - Updated TPU versions for TensorFlow 2.8, 2.9, and 2.10 Deep Learning VMs.
- Improved error messages for debugging custom container Deep Learning VMs that were instantiated with a GPU but without installing NVIDIA drivers.
- Regular package updates.
October 18, 2022
Deep Learning ContainersM98 release
- Upgraded JupyterLab from 3.2 to 3.4.
- Upgraded R from 4.1 to 4.2.
- Miscellaneous bug and display fixes.
- Regular package updates.
M98 release
- Upgraded JupyterLab from 3.2 to 3.4.
- Upgraded R from 4.1 to 4.2.
- Removed the requirement to have the
compute.instances.get
permission in the Service Account attached to the VM introduced in m97. - Added support for the
notebook-enable-debug
metadata flag for JupyterLab low level debugging, which sets:c.Application.log_level = 0
. The default value is 30. - Added support for the
disable-check-xsrf
metadata flag, which sets:c.ServerApp.disable_check_xsrf = True
. The default value is false. - Fixed a bug in which Cloud Marketplace was deploying an older version of Deep Learning VM images.
- Miscellaneous bug and display fixes.
- Regular package updates.
September 29, 2022
Deep Learning ContainersM97 release
- Regular package updates.
M97 release
- Improved the startup time for Ubuntu GPU images.
- Regular package updates.
Proxy registration fails if the Service Account attached to the VM does not have the compute.instances.get
permission
September 20, 2022
Deep Learning ContainersM96 release
- TensorFlow 2.10.0 is now available.
- TensorFlow patch updates for 2.9.2 and 2.8.3 are now available.
- The PyTorch patch update for 1.12.1 is now available.
- Miscellaneous bug fixes.
M96 release
- TensorFlow 2.10.0 is now available.
- TensorFlow patch updates for 2.9.2 and 2.8.3 are now available.
- The PyTorch patch update for 1.12.1 is now available.
- The Diagnostic tool supports DNS resolution check.
- Docker is updated to 20.10.
- Miscellaneous bug fixes.
August 12, 2022
Deep Learning ContainersM95 release
- Tensorflow has been updated to 2.9.1, 2.8.1, and 2.6.5 to include upstream changes.
- Regular package refreshment and bug fixes.
M95 release
- Tensorflow has been updated to 2.9.1, 2.8.1, and 2.6.5 to include upstream changes.
- Updated to the latest NVIDIA driver version: 510.47.03.
- The latest NVIDIA driver version does not support K80 GPUs. To use K80 GPUs, you must use an M94 or earlier environment.
- Fixed bug in which the user is prompted with the warning
JupyterLab build is suggested
on startup for TensorFlow Deep Learning VMs. - Regular package refreshment and bug fixes.
n1-standard-1
Compute instances that use the tensorflow-gpu
family fail to boot if they were created with a single disk and no accelerator.
Please use the tf-latest-cpu
image family for instances without accelerators, or increase the machine type to at least n1-standard-2
.
July 06, 2022
Deep Learning ContainersM94 release
- Added support for PyTorch 1.12.
- Added more system libraries to the R Deep Learning Containers image.
M94 release
- Added support for PyTorch 1.12.
- Added more system libraries to the R Deep Learning VM image.
May 27, 2022
Deep Learning ContainersM93 release
- Starting with this release, information on known vulnerabilities for Deep Learning Containers images is now available on Cloud Storage.
M93 release
- Fixed a bug that prevented kernels from shutting down properly in Vertex AI Workbench managed notebooks.
May 16, 2022
Deep Learning ContainersM92 release
- TensorFlow Enterprise 2.9 is now available. Note that this TensorFlow Enterprise version does not include Long Term Version Support.
- Starting with PyTorch 1.11, PyTorch environments now support XLA by default.
- TensorFlow Enterprise patch releases: 2.6.4 and 2.8.1.
- Deep Learning Containers are now available on Artifact Registry.
M92 release
- TensorFlow Enterprise 2.9 is now available. Note that this TensorFlow Enterprise version does not include Long Term Version Support.
- Starting with PyTorch 1.11, PyTorch environments now support XLA by default.
- TensorFlow Enterprise patch releases: 2.6.4 and 2.8.1.
- Fixed an issue in the Cloud Storage backup and restore feature. This fix helps prevent the deletion of local files after a reboot when the VM loses connectivity to the configured Cloud Storage backup bucket.
April 19, 2022
AI Platform TrainingPre-built PyTorch containers for PyTorch 1.11 are available for training. You can use these containers to train with CPUs, GPUs, or TPUs.
April 14, 2022
Cloud Optimization AICloud Fleet Routing is now generally available (GA).
You can now grant access to Cloud Optimization IAM roles to a user, a group, or a service account to perform create or get operations in the context of a batchOptimizeTours
request.
You can now create models that can solve up to 120 mns (instead of 60 mns) using automatic checkpoints in a batchOptimizeTour
request to solve complex problems.
You can now set a soft_max_load
with related costs on your vehicles to balance the load limit across your fleet.
Optimization AI now has a concurrent batch solve quota of 60000 timeout seconds per project.
March 21, 2022
Deep Learning ContainersM91 release
- PyTorch 1.11 and PyTorch XLA 1.11 are now available in both Deep Learning Containers and Deep Learning VM Images.
- Fixed an R package installation issue for R Deep Learning Containers and Vertex AI Workbench.
M91 release
- PyTorch 1.11 and PyTorch XLA 1.11 are now available in both Deep Learning VM Images and Deep Learning Containers.
- Updated NVIDIA drivers to 470.57.02.
- Upgraded Compute Engine Virtual Ethernet (GVE) to 1.3.0.
February 28, 2022
Deep Learning ContainersM90 release
- CUDA has been upgraded from 11.3.0 to 11.3.1 to address some NCCL issues.
- VSlim GPU TensorFlow containers are available and have a significantly smaller size.
- TensorFlow 2.7 containers are re-released.
M90 release
- Vertex AI sample notebooks are now included in the
/usr/share/tutorials
folder. - Instances now allow the Jupyter options for disabling terminals and deleting files instead of sending them to the trash or recycling bin.
In M90 release instances, gRPC 1.44.0 can generate spurious error logs, though this doesn't affect the VM's ability to boot up. A fix is planned for the next release.
February 15, 2022
AI Platform TrainingRuntime version 2.8 is available. You can use runtime version 2.8 to train with TensorFlow 2.8, scikit-learn 1.0.2, or XGBoost 1.5.2. Runtime version 2.8 supports training with CPUs, GPUs, or TPUs.
See the full list of updated dependencies in runtime version 2.8.
February 02, 2022
Deep Learning ContainersM89 release
- TensorFlow Enterprise 2.8 is now available and includes Long Term Version Support. TensorFlow Enterprise 2.8 is available in both Deep Learning Containers and Deep Learning VM Images.
- Upgraded TensorFlow Enterprise 2.6.2 to 2.6.3.
M89 release
- TensorFlow Enterprise 2.8 is now available and includes Long Term Version Support. TensorFlow Enterprise 2.8 is available in both Deep Learning VM Images and Deep Learning Containers.
- Upgraded TensorFlow Enterprise 2.6.2 to 2.6.3.
December 20, 2021
Deep Learning ContainersM88 release
- To help address the Apache Log4j 2 vulnerability, H2O has been updated to 3.34.0.6 in the R container images. This addresses CVE-2021-44228 and CVE-2021-45046. See the official H2O update.
- Starting with this release, the Python packages that are installed on each container image are listed in files that are available on Cloud Storage. For example, the lists for this release are available at
gs://deeplearning-platform-release/installed-dependencies/containers/m88/
.
M88 release
- To help address the Apache Log4j 2 vulnerability, H2O has been updated to 3.34.0.6 in the R and Theia images. This addresses CVE-2021-44228 and CVE-2021-45046. See the official H2O update.
- As previously announced in the M87 release and M71 release, the previous format of TensorFlow 2.x image names,
tf2-xxx-2-y-zzz
, is unavailable starting with this release. Please use the current format oftf-xxx-2-y-zzz
for image names.
- Starting with this release, the Python packages that are installed on each image are listed in files that are available on Cloud Storage. For example, the lists for this release are available at
gs://deeplearning-platform-release/installed-dependencies/images/m88/
.
- Images from the M88 release mistakenly have M87 metadata stored in the images. For example, the welcome message upon terminal login for the base CPU image shows "Version: common-cpu.m87". This mistaken metadata is also shown in the
version
field in notebookCustom metadata
. Users can verify they are actually using the M88 images by looking forv20211219
in the image name of the boot disk. After clicking the image, users can also verify if the image has the labelrelease : m88
. Other than the mistaken metadata, users can use the M88 images as normal.
December 08, 2021
AI Platform TrainingRuntime version 2.7 is available. You can use runtime version 2.7 to train with TensorFlow 2.7, scikit-learn 1.0.1, or XGBoost 1.5.0. Runtime version 2.7 supports training with CPUs, GPUs, or TPUs.
See the full list of updated dependencies in runtime version 2.7.
December 06, 2021
Deep Learning ContainersM87 release
- Added Artifact Registry's Python keyring authentication library to deep learning vm environments.
- TensorFlow 2.x container image names are available in two formats: the current standard, which includes a
tf-
prefix, and the previous standard, which includes atf2-
prefix. For example, bothgcr.io/deeplearning-platform-release/tf-gpu.2-7
andgcr.io/deeplearning-platform-release/tf2-gpu.2-7
are available although they are the same container images. Starting within approximately six months, releases of TensorFlow 2 container images will only be named with the current standard.
M87 release
- Added Artifact Registry's Python keyring authentication library to deep learning vm environments.
- The M87 release is the last release in which TensorFlow 2.x image names are available in two formats: the current standard,
tf-xxx-2-y-zzz
and the previous standard,tf2-xxx-2-y-zzz
. For example, bothtf-ent-2-7-cpu
andtf2-ent-2-7-cpu
are available although they are the same images. The next release will only contain TensorFlow 2 images named with the current standard, as originally announced in the M71 release notes from June 2021.
November 18, 2021
Deep Learning ContainersM86 release
- Upgraded all Ubuntu 18.04 LTS Deep Learning Container images to Ubuntu 20.04 LTS (see What is an Ubuntu LTS release?).
- Released PyTorch/XLA 1.10.
- Upgraded TensorFlow Enterprise image to the latest patch version: 2.6.2
- Deprecated CUDA 10.x environments.
- Locked JupyterLab version to 3.2.
M86 release
- Upgraded all Ubuntu 18.04 LTS Deep Learning VM images to Ubuntu 20.04 LTS (see What is an Ubuntu LTS release?).
- Released PyTorch/XLA 1.10.
- Upgraded TensorFlow Enterprise image to the latest patch version: 2.6.2
- Deprecated CUDA 10.x environments.
- Locked JupyterLab version to 3.2.
November 17, 2021
AI Platform PredictionRuntime version 2.7 is now available. You can use runtime version 2.7 to serve online predictions with TensorFlow 2.7.0, scikit-learn 1.0, or XGBoost 1.4.2. Runtime version 2.7 does not support batch prediction.
See the full list of updated dependencies in runtime version 2.7.
November 15, 2021
Cloud Optimization AICloud Fleet Routing is now available in General Availability GA.
You can now use Place ID in Cloud Fleet Routing as an alternative to latitude and longitude, which is still available.
You can now accept Google Map Platform's Terms of Service when using Distance Matrix API in Cloud Fleet Routing through Google Maps for Fleet Routing.
demands[]/capacities[]
fields are improved with Load Limit
: this feature will help you dynamically control how much shipment load will be added to each vehicle, pushing the solver to balance out the shipments more evenly.
The demands[]/capacities[]
related fields will be deprecated on November 15th, 2022.
googleMapsApiKey
field in optimizeTours
is deprecated and replaced with Google Maps for Fleet Routing. Update your code to remove the google_maps_api_key
field from sending requests. If not, it causes errors on the client side.
November 08, 2021
Deep Learning ContainersM85 release
- Regular package refreshment and bug fixes.
M85 release
- CUDA 11.3 Debian-10 image is available.
- Regular package refreshment and bug fixes.
November 05, 2021
Deep Learning ContainersM84 release
- TensorFlow Enterprise 2.7 is now available with CUDA 11.3 support. Note that this TensorFlow Enterprise version does not include Long Term Version Support.
M84 release
- TensorFlow Enterprise 2.7 is now available with CUDA 11.3 support. Note that this TensorFlow Enterprise version does not include Long Term Version Support.
November 02, 2021
AI Platform TrainingUsing interactive shells to inspect training jobs is generally available (GA).
You can use these interactive shells with VPC Service Controls.
October 28, 2021
Deep Learning ContainersM83 release
- PyTorch 1.10 is now available.
M83 release
- PyTorch 1.10 is now available.
October 26, 2021
Deep Learning ContainersM82 release
- Released CUDA11.3 container images.
- The Vertex SDK for Python is available across all deep learning environment products; it was previously available only in TensorFlow images.
- Theia IDE (experimental) images were refreshed. PyTorch has been removed from Theia IDE images.
M82 release
- The Vertex SDK for Python is available across all deep learning environment products; it was previously available only in TensorFlow images.
- Theia IDE (experimental) images were refreshed. PyTorch has been removed from Theia IDE images.
October 12, 2021
Deep Learning ContainersM81 release
- Upgraded R to 4.1.
- Fixed bug that prevented R kernels from working properly.
M81 release
- Upgraded R to 4.1.
- Improved Cloud Storage sync logic so that only newer files sync.
- Fixed bug that prevented R kernels from working properly.
October 06, 2021
AI Platform TrainingRuntime version 2.6 is available. You can use runtime version 2.6 to train with TensorFlow 2.6, scikit-learn 0.24.2, or XGBoost 1.4.2. Runtime version 2.6 supports training with CPUs, GPUs, or TPUs.
See the full list of updated dependencies in runtime version 2.6.
September 24, 2021
Deep Learning ContainersStarting with the M80 image release, all environments will include JupyterLab 3.x by default. To continue using an existing environment's JupyterLab 1.x version, disable auto-upgrade (if enabled) and do not manually upgrade the environment to a new environment version. To create new instances using older images that have JupyterLab 1.x installed, see creating specific versions of instances.
M80 release
- Updated JupyterLab from 1.x to 3.x.
- Added Jupytext.
Starting with the M80 image release, all environments will include JupyterLab 3.x by default. To continue using an existing environment's JupyterLab 1.x version, disable auto-upgrade (if enabled) and do not manually upgrade the environment to a new environment version. To create new instances using older images that have JupyterLab 1.x installed, see creating specific versions of instances.
M80 release
- Updated JupyterLab from 1.x to 3.x.
- Added Jupytext.
- Deep Learning VM Images in Cloud Marketplace have been updated. They were not updated in the last release.
September 23, 2021
AI Platform TrainingTraining with pre-built PyTorch containers is generally available (GA).
Pre-built PyTorch containers for PyTorch 1.9 are available for training. You can use these containers to train with CPUs, GPUs, or TPUs.
September 09, 2021
Deep Learning ContainersM79 release
- Updated Pytorch 1.9 containers (they were not refreshed in the last release).
- Updated Theia IDE (experimental) containers.
- Node.js is pinned to
>=12.14.1,<13
.
- M79 is the last release version that has JupyterLab 1.x installed. For the next release (M80), JupyterLab will be upgraded to 3.x for all Deep Learning VM Images, Deep Learning Containers, and Notebooks.
- Fixed a bug in which the home folder in custom container VMs was owned by the root instead of Jupyter.
M79 release
- Updated Pytorch 1.9 images (they were not refreshed in the last release).
- Updated Theia IDE (experimental) images.
- Node.js is pinned to
>=12.14.1,<13
.
- M79 is the last release version that has JupyterLab 1.x installed. For the next release (M80), JupyterLab will be upgraded to 3.x for all Deep Learning VM Images, Deep Learning Containers, and Notebooks.
- Deep Learning VM Images in Cloud Marketplace have not been updated. They are planned to be refreshed during the next release.
- Fixed a bug in which the home folder in custom container VMs was owned by the root instead of Jupyter.
September 01, 2021
AI Platform PredictionRuntime version 2.6 is now available. You can use runtime version 2.6 to serve online predictions with TensorFlow 2.6.0, scikit-learn 0.24.2, or XGBoost 1.4.2. Runtime version 2.6 does not support batch prediction.
See the full list of updated dependencies in runtime version 2.6.
August 18, 2021
Deep Learning ContainersM78 release
- Updated TensorFlow Enterprise patch version 2.3.3 to 2.3.4.
TensorFlow Enterprise 2.5
- TensorFlow Enterprise 2.5 Deep Learning Containers are now deprecated.
M78 release
- Updated TensorFlow Enterprise patch version 2.3.3 to 2.3.4.
- Fixed a bug that prevented users from exporting a notebook as a PDF.
- Fixed a bug that caused some users to be unable to SSH into their host machines.
TensorFlow Enterprise 2.5
- TensorFlow Enterprise 2.5 Deep Learning VM images are now deprecated.
August 11, 2021
Deep Learning ContainersM77 release
TensorFlow Enterprise 2.6.0 is now available and includes Long Term Version Support.
M77 release
TensorFlow Enterprise 2.6.0 is now available and includes Long Term Version Support.
August 09, 2021
AI Platform TrainingYou can use TPU Pods for training. This feature is available in Preview.
August 02, 2021
Deep Learning ContainersM76 release
- Added the Vertex SDK for Python.
- Regular package refreshment and bug fixes.
M76 release
- Added the Vertex SDK for Python.
- Regular package refreshment and bug fixes.
July 19, 2021
AI Platform TrainingYou can now use an interactive shell to inspect your training container while it runs. The interactive shell can be helpful for monitoring and debugging training jobs.
This feature is available in preview.
July 15, 2021
Deep Learning ContainersM75 release
- Enhanced environment configurations so it is easier to install additional frameworks in CUDA containers.
M75 release
- Improved the clarity of error messages for custom container users.
July 07, 2021
Deep Learning VM ImagesM74 release
- Re-released all CUDA 11 images to fix an issue in which VM instances are unable to detect GPU accelerators after a reboot. If you have not experienced this issue, no action is needed. If your GPU is disconnected from your instance after a reboot, see these resolution options.
- In Debian 10 GPU images, updated NVIDIA drivers to 460.73.01 and CUDA to 11.0.3.
- Added support for controlling the Cloud Storage backup synchronization time and reducing the output of synchronization.
- Preinstalled the table of contents extension in JupyterLab.
- Added fastai 2.4 to the PyTorch 1.9 GPU image.
July 01, 2021
Cloud Optimization AICloud Optimization AI rolled out updates to replace
ShipmentRoute.TravelStep
with ShipmentRoute.Transition
.
Transitions are meant to inform about everything that is happening before and/or after visits. Whereas a TravelStep only contained travel distance and duration information, a Transition also holds quantity loads and delay durations. More fields will be added to Transitions in the future.
This is a backwards-compatible change: the old fields are kept and maintained. No features or capabilities were removed. Details for this release are listed here:
ShipmentRoute.travel_steps
is changed toShipmentRoute.transitions
.ShipmentRoute.Visit.delay_before_start
andShipmentRoute.delay_before_vehicle_end
are changed toShipmentRoute.Transition.delay_duration
.ShipmentRoute.Visit.arrival_loads
andShipmentRoute.end_loads
are changed toShipmentRoute.Transition.loads
.ShipmentRoute.TravelStep.duration
is changed toShipmentRoute.Transition.travel_duration
.ShipmentRoute.TravelStep.distance_meters
is changed toShipmentRoute.Transition.travel_distance_meters
.
June 22, 2021
Deep Learning ContainersM73 release
- Upgraded TensorFlow Enterprise 2.1.3 to 2.1.4.
- Upgraded TensorFlow Enterprise 2.3.2 to 2.3.3.
- Miscellaneous bug fixes and updates.
M73 release
- Upgraded TensorFlow Enterprise 2.1.3 to 2.1.4.
- Upgraded TensorFlow Enterprise 2.3.2 to 2.3.3.
- Disabled automatic updates for Ubuntu to be in line with the behavior in Debian images.
- Miscellaneous bug fixes and updates.
June 17, 2021
Deep Learning ContainersM72 release
- Added PyTorch 1.9 and PyTorch/XLA 1.9 containers.
M72 release
- Added PyTorch 1.9 and PyTorch/XLA 1.9 images.
June 11, 2021
AI Platform TrainingNVIDIA A100 GPUs and accelerator-optimized (A2) machine types are generally available (GA).
June 08, 2021
AI Platform PredictionRuntime version 2.5 is now available. You can use runtime version 2.5 to serve online predictions with TensorFlow 2.5.1, scikit-learn 0.24.1, or XGBoost 1.4.0. Runtime version 2.5 does not support batch prediction.
See the full list of updated dependencies in runtime version 2.5.
June 02, 2021
Deep Learning ContainersM71 release
- Upgraded TensorFlow Probability, TensorFlow I/O, and TensorFlow Estimator in TensorFlow 2.5 containers.
M71 release
- Refreshed the Debian-10 images (Ubuntu images not refreshed in this release).
- Upgraded TensorFlow Probability, TensorFlow I/O, and TensorFlow Estimator in TensorFlow 2.5 images.
- Added support for a Post Startup script and provided status in guest attributes.
- TensorFlow 2.x image names are now available in two formats:
tf-xxx-2-y-zzz
(the new standard format)tf2-xxx-2-y-zzz
(the previous standard format). Image names in the previous standard format will be deprecated in a future release.
May 14, 2021
Deep Learning ContainersM70 release
- Added TensorFlow Enterprise 2.5 containers. Note this is an Enterprise version but not a Long Term Support (LTS) version.
M70 release
- Added TensorFlow Enterprise 2.5 images. Note this is an Enterprise version but not a Long Term Support (LTS) version.
May 13, 2021
Deep Learning ContainersM69 release
- Updated cuDNN from 8.0.4 to 8.0.5.
M69 release
- Migrated Collection Agent to Cloud Monitoring version 2.
May 05, 2021
Deep Learning ContainersM68 release
- Upgraded R containers from 3.6 to 4.0.
- Added xai-tabular-widget onto all TensorFlow containers.
- Miscellaneous bug fixes and updates.
M68 release
- Upgraded R Images from 3.6 to 4.0.
- Added xai-tabular-widget onto all TensorFlow images.
- Miscellaneous bug fixes and updates.
April 19, 2021
Deep Learning ContainersM67 release
- Added Horovod to TensorFlow GPU containers.
- Regular package refreshment and bug fixes.
M67 release
- GPU support added for Beam Notebooks.
- Added Horovod to TensorFlow GPU Deep Learning VMs.
- Regular package refreshment and bug fixes.
April 16, 2021
AI Platform PredictionRuntime version 2.4 is now available. You can use runtime version 2.4 to serve online predictions with TensorFlow 2.4.1, scikit-learn 0.24.0, or XGBoost 1.3.1. Runtime version 2.4 does not support batch prediction.
See the full list of updated dependencies in runtime version 2.4.
March 31, 2021
Deep Learning ContainersM66 release
- PyTorch 1.8 support in deep learning environments (Deep Learning VM Image and Deep Learning Containers) is available.
- Fixed scope allocator optimization issue with the TensorFlow Enterprise 2.3/2.1 MKL build.
M66 release
- PyTorch 1.8 support in deep learning environments (Deep Learning VM Image and Deep Learning Containers) is available.
- Fixed scope allocator optimization issue with the TensorFlow Enterprise 2.3/2.1 MKL build.
- Regular package refreshment and bug fixes.
March 05, 2021
AI Platform TrainingAI Platform Training now provides pre-built PyTorch containers for PyTorch 1.7.
In addition to training with CPUs or GPUs, you can use one of the PyTorch 1.7 containers to perform PyTorch training with a TPU.
M65 release
Upgraded tensorflow-cloud to 0.1.13.
Regular package refreshment and bug fixes.
M65 release
Added support for DooD (Docker outside of Docker) in Dataflow notebooks container images.
Upgraded tensorflow-cloud to 0.1.13.
Regular package refreshment and bug fixes.
February 19, 2021
Deep Learning ContainersM64 release
Upgraded TensorFlow 2.4 to 2.4.1.
Upgraded TFX and Fairness Indicators from 0.26.0 to 0.27.0.
Miscellaneous bug fixes and updates.
Swift For TensorFlow
- The Swift For TensorFlow project is entering archive mode. Containers will be deprecated and will no longer receive updates after this release.
M64 release
Upgraded TensorFlow 2.4 to 2.4.1.
Upgraded TFX and Fairness Indicators from 0.26.0 to 0.27.0.
Added the Fast.ai book tutorials to Pytorch images.
Enabled gVNIC for all DLVM images.
Miscellaneous bug fixes and updates.
Swift For TensorFlow
- The Swift For TensorFlow project is entering archive mode. Swift images will be deprecated and will no longer receive updates after this release.
February 16, 2021
AI Platform TrainingThe default boot disk type for virtual machine instances used for training jobs has changed from pd-standard
to pd-ssd
. Learn more about disk types for custom training and read about pricing for different disk types.
Note that for training jobs where you don't specify a DiskConfig
, pricing does not change. This is because the first 100 GB of disk for each VM do not incur any charge, regardless of disk type.
February 10, 2021
AI Platform TrainingRuntime version 2.4 is now available. You can use runtime version 2.4 to train with TensorFlow 2.4.1, scikit-learn 0.24.0, or XGBoost 1.3.1. Runtime version 2.4 supports training with CPUs, GPUs, or TPUs.
See the full list of updated dependencies in runtime version 2.4.
February 08, 2021
Deep Learning ContainersM63 release
- SWIFT 0.12 (experimental) containers are available.
M63 release
- Nvidia driver is upgraded to 450.80.02.
- TFX version is upgraded to 0.26.1.
- Regular package refreshment and bug fixes.
February 01, 2021
AI Platform TrainingYou can now use E2, N2, and C2 machine types for training. Learn about the specific machine types available for training, and learn about their pricing.
January 28, 2021
AI Platform TrainingYou can now use NVIDIA A100 GPUs and several accelerator-optimized (A2) machine types for training. You must use A100 GPUs and A2 machine types together.
A100 GPUs and A2 machine types are available in preview. Learn about their pricing.
January 25, 2021
Deep Learning ContainersGeneral Availability
AI Platform Deep Learning Containers is now generally available.
Python 2
Python 2 is no longer supported in Deep Learning Containers. Read more about Python 2 support on Google Cloud.
M62 release
Upgraded TensorFlow 2.3 to 2.3.2
Upgraded TensorFlow 2.1 to 2.1.3
Miscellaneous bug fixes and updates
Python 2
Python 2 is no longer supported in Deep Learning VM Image. Read more about Python 2 support on Google Cloud.
M62 release
- Upgraded TensorFlow 2.3 to 2.3.2
- Upgraded TensorFlow 2.1 to 2.1.3
- Miscellaneous bug fixes and updates
January 20, 2021
AI Platform PredictionThe VPC Service Controls integration with AI Platform Prediction is now generally available.
Training with a custom service account is now generally available.
Support for VPC Network Peering is now generally available.
January 15, 2021
AI Platform TrainingAI Platform Training now provides pre-built PyTorch containers for PyTorch 1.6.
In addition to training with CPUs or GPUs, you can use one of the PyTorch 1.6 containers to perform PyTorch training with a TPU.
December 16, 2020
AI Platform PredictionYou can now configure AI Platform Prediction to automatically scale prediction nodes for model versions that use GPUs for online prediction.
Previously, you could only configure manual scaling for model versions that use GPUs. Now, you can choose between automatic and manual scaling.
Using automatic scaling with GPUs is available in preview.
Added TensorFlow 2.4 Deep Learning Containers images.
M60 release
- Added TensorFlow 2.4 Deep Learning VM Images
December 09, 2020
AI Platform PredictionRuntime version 2.3 is now available. You can use runtime version 2.3 to serve online predictions with TensorFlow 2.3.1, scikit-learn 0.23.2, or XGBoost 1.2.1. Runtime version 2.3 does not support batch prediction.
See the full list of updated dependencies in runtime version 2.3.
Runtime version 2.3 is now available. You can use runtime version 2.3 to train with TensorFlow 2.3.1, scikit-learn 0.23.2, or XGBoost 1.2.1. Runtime version 2.3 supports training with CPUs, GPUs, or TPUs.
See the full list of updated dependencies in runtime version 2.3.
November 12, 2020
Deep Learning VM ImagesM59 release
- Miscellaneous bug fixes.
tensorflow_enterprise_addon
package is renamed totensorflow-cloud
November 11, 2020
AI Platform PredictionThe following regional endpoints are now generally available for online prediction, in addition to the regional endpoints that were already available:
us-east1-ml.googleapis.com
us-east4-ml.googleapis.com
us-west1-ml.googleapis.com
northamerica-northeast1-ml.googleapis.com
europe-west1-ml.googleapis.com
europe-west2-ml.googleapis.com
europe-west3-ml.googleapis.com
asia-northeast1-ml.googleapis.com
asia-southeast1-ml.googleapis.com
australia-southeast1-ml.googleapis.com
On some of these regional endpoints, you can use GPUs to accelerate prediction. Learn which types of GPUs are available on which regional endpoints.
Pricing for online prediction varies between regional endpoints. Read about the pricing for each regional endpoint.
October 28, 2020
Deep Learning Containers- Added PyTorch 1.6 CUDA 11 environments that support A100 GPU accelerators. This special PyTorch build provides another option to add to our A100-compatible TensorFlow Enterprise builds.
October 27, 2020
AI Platform PredictionYou can now use use a custom container to customize how you serve predictions. To try using a custom container, read the new tutorial on serving predictions from a PyTorch model.
This feature is in preview.
Console logging (formerly also referred to as "stream logging") is now available in preview for Compute Engine (N1) machine types and in GA for legacy (MLS1) machine types.
Read a new document about using custom service accounts with custom containers or custom prediction routines.
This feature is in beta.
M58 release
- Added PyTorch 1.6 CUDA 11 images that support A100 GPU accelerators. This special PyTorch build provides another option to add to our A100-compatible TensorFlow Enterprise builds.
- Added the PyTorch/XLA package.
- Added the Swift for TensorFlow framework.
- Added the Ubuntu 18.04 OS.
- TensorFlow Enterprise updated to 2.3.1 from 2.3.0.
- Debian 10 is now the default OS for Deep Learning VM images.
September 24, 2020
Deep Learning VM ImagesM56 release
- Bug fixes for TensorFlow 2.3 add-ons
- Fixes bug affecting BigQuery magic commands in some environments
- Adds a diagnostics tool for AI Platform Notebooks
September 22, 2020
AI Platform TrainingAI Platform Training runtime version 2.2 now supports training with TPUs using TensorFlow 2.2.
August 28, 2020
AI Platform PredictionRuntime version 2.2 is now available. You can use runtime version 2.2 to serve online predictions with TensorFlow 2.2.0, scikit-learn 0.23.1, or XGBoost 1.1.1. Runtime version 2.2 does not currently support batch prediction.
See the full list of updated dependencies in runtime version 2.2.
Runtime version 2.2 is now available. You can use runtime version 2.2 to train with TensorFlow 2.2.0, scikit-learn 0.23.1, or XGBoost 1.1.1. See the full list of updated dependencies in runtime version 2.2.
August 27, 2020
AI Platform OptimizerAI Platform Optimizer's name has changed to AI Platform Vizier.
August 26, 2020
Deep Learning VM ImagesM55 release
- Restricts Jupyter memory usage to fix 5* issues
- Updates TensorFlow 2.3 dependencies
- Uses CUDA 11.0 in TensorFlow deep learning images
- Adds support for the us-east4 region
August 19, 2020
AI Platform PredictionYou can now use runtime version 2.1 to serve online predictions using scikit-learn 0.22.1 and XGBoost 0.90.
August 18, 2020
AI Platform PredictionCompute Engine (N1) machine types for online prediction are now generally available. They are available on all regional endpoints.
The AI Platform Training and Prediction Service Level Agreement does not apply to model versions that use a Compute Engine (N1) machine type and fewer than two prediction nodes.
GPUs for online prediction are now generally available. You can use GPUs to serve predictions when you create a TensorFlow model version that uses a Compute Engine (N1) machine type.
Learn which types of GPU are available on each regional endpoint.
The following regional endpoints for online prediction are now generally available:
us-central1-ml.googleapis.com
europe-west4-ml.googleapis.com
asia-east1-ml.googleapis.com
Using Compute Engine (N1) machine types on the global API endpoint (ml.googleapis.com
) is deprecated. This functionality was previously available in beta in the us-central1
region.
To continue to use Compute Engine (N1) machine types in the us-central1
region, create a model on the us-central1-ml.googleapis.com
regional endpoint, and then create model versions using that model.
August 17, 2020
AI Platform TrainingYou can now set a maximum time that you are willing to wait between the moment when you create a training job and the moment when AI Platform Training starts running the job. If your training job has not started running after this duration, AI Platform Training cancels the job. Set the maximum wait time by specifying the scheduling.maxWaitTime
field.
TensorFlow Enterprise 2.3 environments are now available. These environments include support for A100 GPU accelerators, CUDA 11, and TensorFloat-32 (TF32).
August 14, 2020
AI Platform TrainingThe TabNet built-in algorithm is now available in Beta. You can train models on tabular data for classification and regression problems, and also get feature attributions to help explain the model's behavior.
August 10, 2020
Deep Learning VM ImagesM54 release
- Added support for the europe-west3 region
- Updated the Explainable AI sdk and added explainers
- Fixed llvm-openmp support
- Added support for instance auto upgrade
- Made Deep Learning VM images and Deep Learning Containers more consistent for TPU
- Updated NCCL to 2.7.6 in CU110 images
- Added the scikit-learn package and container
- Added JRE to R images
- Limited custom container memory utilization
August 06, 2020
Deep Learning VM ImagesM53 release
TensorFlow Enterprise 2.3 images, including images that support CUDA 11.0, are now available.
August 04, 2020
AI Platform TrainingRead a new guide to distributed PyTorch training. You can use this guide with pre-built PyTorch containers, which are in beta.
July 20, 2020
AI Platform TrainingCustomer-managed encryption keys (CMEK) for AI Platform Training is now generally available.
The VPC Service Controls integration with AI Platform Training is now generally available.
You can now train a PyTorch model on AI Platform Training by using a pre-built PyTorch container. Pre-built PyTorch containers are available in beta.
July 14, 2020
AI Platform PredictionVPC Service Controls now supports AI Platform Prediction. Learn how to use a service perimeter to protect online prediction. This functionality is in beta.
July 13, 2020
AI Platform TrainingYou can now configure a training job to run using a custom service account. Using a custom service account can help you customize which Google Cloud resources your training code can access.
This feature is available in beta.
M51 release
Allow removing sudo
access from Deep Learning Containers.
Debian-10-based images are released. You can create Shielded VM instances from these images.
June 23, 2020
Deep Learning VM ImagesM50 release
Miscellaneous bug fixes.
June 22, 2020
AI Platform TrainingYou can now use Cloud TPUs for training jobs in the europe-west4
region. TPU v2 accelerators are generally available, and TPU v3 accelerators are available in beta.
Learn how to configure your training job to use TPUs, and read about TPU pricing on AI Platform Training.
June 15, 2020
AI Platform TrainingAI Platform Training now supports private services access in beta. You can use VPC Network Peering to create a private connection so that training jobs can connect to your network on private IP.
Learn how to set up VPC Network Peering with AI Platform Training.
June 11, 2020
Deep Learning VM ImagesM49 release
TensorFlow Enterprise images updated to 1.15.3 and 2.1.1.
The tensorflow-enterprise-addons package is now available in all deep learning environments.
XGBoost, MXNet, R, PyTorch, CNTK, and Caffe images have been updated with library upgrades and bug fixes.
June 08, 2020
AI Platform PredictionThe Total latency chart on the Version details page of the Google Cloud Console was reporting incorrect information. This chart has now been fixed.
In some cases, this adjustment might cause latencies to appear higher than they were previously. However, the latency of models has not changed.
This affects both Compute Engine (N1) machine types and legacy (MLS1) machine types.
May 21, 2020
AI Platform TrainingYou can now use TPUs with TensorFlow 2.1 when you create a training job with runtime version 2.1. You can also use TPUs with TensorFlow 2.1 when you train in a custom container.
Read the guide to using TPUs with AI Platform Training, which has been updated to show how to use TPUs with TensorFlow 2 APIs.
May 18, 2020
Deep Learning VM ImagesM48 release
TensorFlow 2.2 images have been added. The new TensorFlow 2.2 image families are tf2-2-2-cpu
and tf2-2-2-cu101
. See the available image families.
May 13, 2020
AI Platform PredictionAI Platform Prediction now supports the following regions for batch prediction, in addition to those that were already supported:
northamerica-northeast1
(Montréal)southamerica-east1
(São Paulo)australia-southeast1
(Sydney)
See the full list of available regions.
northamerica-northeast1
and southamerica-east1
have the same pricing as other Americas regions, and australia-southeast1
has the same pricing as other Asia Pacific regions. Learn about pricing for each region.
AI Platform Training now supports the following regions, in addition to those that were already supported:
northamerica-northeast1
(Montréal)southamerica-east1
(São Paulo)australia-southeast1
(Sydney)
GPUs are available for training in each of the new regions:
- NVIDIA Tesla P4 GPUs are available in
northamerica-northeast1
. - NVIDIA Tesla T4 GPUs are available in
southamerica-east1
. - NVIDIA Tesla P4 GPUs and NVIDIA Tesla P100 GPUs are available in
australia-southeast1
.
See the full list of available regions and the guide to training with GPUs.
northamerica-northeast1
and southamerica-east1
have the same pricing as other Americas regions, and australia-southeast1
has the same pricing as other Asia Pacific regions. Learn about pricing for each region.
May 12, 2020
Deep Learning VM ImagesM47 release
Fixed an OS login issue under single user mode for a user external to an organization.
Fixed a git extensions plugin issue in TensorFlow 2 images.
April 29, 2020
AI Platform PredictionAI Platform Prediction now supports several regional endpoints for online prediction. Regional endpoints provide additional protection against outages in other regions by isolating your model and version resources from other regions. The following regional endpoints are available in beta:
us-central1-ml.googleapis.com
europe-west4-ml.googleapis.com
asia-east1-ml.googleapis.com
You can use these endpoints instead of the global endpoint, ml.googleapis.com
, when you use AI Platform Prediction for online prediction. Learn how to use regional endpoints for online prediction, and read about their benefits and limitations.
You can now deploy scikit-learn and XGBoost models for online prediction using Compute Engine (N1) machine types. Previously, you could only deploy TensorFlow models when you used these machine types. Learn more about ML framework support for Compute Engine (N1) machine types.
You cannot use GPUs with scikit-learn or XGBoost models, and you can only use scikit-learn and XGBoost models with Compute Engine (N1) machine types when you deploy your models and versions to a regional endpoint.
Compute Engine (N1) machine types for online prediction remain available in the beta launch stage.
The europe-west4
(Netherlands) and asia-east1
(Taiwan) regions are now available for online prediction. These regions are only available for online prediction on their respective regional endpoints, and you can only use Compute Engine (N1) machine types for online prediction in these regions.
When you deploy model versions in the europe-west4
region, you can optionally use NVIDIA Tesla P4, NVIDIA Tesla T4, or NVIDIA Tesla V100 GPUs to accelerate prediction.
When you deploy model versions in the asia-east1
region, you can optionally use NVIDIA Tesla K80 or NVIDIA Tesla P100 GPUs to accelerate prediction.
Learn more about using GPUs for online prediction, and see which GPUs are available in which regions.
Learn about the pricing for the newly available regions and GPU resources.
We recommend against using Compute Engine (N1) machine types on the AI Platform Prediction global endpoint. Instead, only use Compute Engine (N1) machine types when you deploy models and versions to a regional endpoint.
Model versions that use Compute Engine (N1) machine types and were previously deployed to the us-central1
region on the global endpoint will continue to function.
April 24, 2020
AI Platform PredictionVisualization settings for AI Explanations are now available. You can customize how feature attributions are displayed for image data.
Learn more about visualizing explanations.
April 13, 2020
AI Platform PredictionThe pricing of Compute Engine (N1) machine types for online prediction in the us-central1 region has changed. vCPU resources now cost $0.031613 per vCPU hour and RAM now costs $0.004242 per GB hour.
April 09, 2020
AI Platform PredictionIf you deploy a model version for online prediction that uses runtime version 2.1 with a GPU, AI Platform Prediction now correctly uses TensorFlow 2.1.0 to serve predictions. Previously, AI Platform Prediction used TensorFlow 2.0.0 to serve predictions in this situation.
You can now specify virtual machine instances with the evaluator task type as part of your training cluster for distributed training jobs. Read more about evaluators in TensorFlow distributed training, see how to configure machine types for evaluators, and learn about using evaluators with custom containers.
The maximum running time for training jobs now defaults to seven days. If a training job is still running after this duration, AI Platform Training cancels the job.
April 08, 2020
AI Platform OptimizerAI Platform Optimizer is now available in beta.
AI Platform Optimizer is a black-box optimization service that helps you tune hyperparameters in complex machine learning models.
Visit the AI Platform Optimizer overview to learn more about how it works. To get started, try using AI Platform Optimizer to optimize a machine learning model or to optimize two functions at once.
April 06, 2020
AI Platform TrainingRuntime version 2.1 now includes scikit-learn 0.22.1 instead of scikit-learn 0.22.
April 03, 2020
AI Platform TrainingYou can now use customer-managed encryption keys (CMEK) to protect data in your AI Platform Training jobs. This feature is available in beta.
To learn about the benefits and limitations of using CMEK, and to walk through configuring CMEK for a training job, read the guide to using CMEK with AI Platform Training.
March 27, 2020
AI Platform PredictionAI Explanations now supports XRAI, a new feature attribution method for image data.
The image tutorial has been updated to include XRAI. In the tutorial, you can deploy an image classification model using both integrated gradients and XRAI, and compare the results.
AI Explanations provides an approximation error with your explanations results.
Learn more about the approximation error and how to improve your explanations results.
AI Platform Prediction now supports the following regions for batch prediction, in addition to those that were already supported:
us-west3
(Salt Lake City)europe-west2
(London)europe-west3
(Frankfurt)europe-west6
(Zurich)asia-south1
(Mumbai)asia-east2
(Hong Kong)asia-northeast1
(Tokyo)asia-northeast2
(Osaka)asia-northeast3
(Seoul)
Note that asia-northeast1
was already available for online prediction.
See the full list of available regions and read about pricing for each region.
AI Platform Training now supports the following regions, in addition to those that were already supported:
us-west3
(Salt Lake City)europe-west2
(London)europe-west3
(Frankfurt)europe-west6
(Zurich)asia-south1
(Mumbai)asia-east2
(Hong Kong)asia-northeast1
(Tokyo)asia-northeast2
(Osaka)asia-northeast3
(Seoul)
Out of these regions, the following support training with NVIDIA Tesla T4 GPUs:
europe-west2
asia-south1
aisa-northeast1
asia-northeast3
See the full list of available regions and read about pricing for each region.
March 17, 2020
AI Platform TrainingRuntime versions 1.2 through 1.9 are no longer available for training. We recommend that you use runtime version 1.14 or later for your training jobs.
Read more about the AI Platform Training policy for supporting older runtime versions. This policy is being retroactively implemented in several stages for runtime versions 1.13 and earlier.
March 09, 2020
AI Platform PredictionRuntime version 2.1 for AI Platform Prediction is now available.
Runtime version 2.1 is the first runtime version to support TensorFlow 2 for online and batch prediction. Specifically, this runtime version includes TensorFlow 2.1.0. Review the updated guide to exporting TensorFlow SavedModels for use with AI Platform Prediction for details about exporting compatible models by using TensorFlow 2 APIs, like Keras.
When you use runtime version 2.1 for online prediction, you can currently only deploy TensorFlow model versions. You cannot deploy scikit-learn, XGBoost, or custom prediction routine artifacts for online prediction with runtime version 2.1. For the time being, continue to use runtime version 1.15 to serve predictions from these types of models.
Runtime version 2.1 is also the first runtime version not to support Python 2.7. The Python Software Foundation ended support for Python 2.7 on January 1, 2020. No AI Platform runtime versions released after January 1, 2020 support Python 2.7.
If you deploy a model version for online prediction that uses runtime version 2.1 with a GPU, AI Platform Prediction uses TensorFlow 2.0.0 (instead of TensorFlow 2.1.0) to serve predictions. This is a known issue, and the release notes will be updated when online prediction with GPUs supports TensorFlow 2.1.0.
Runtime version 2.1 for AI Platform Training is now available.
Runtime version 2.1 is the first runtime version to support TensorFlow 2. Specifically, this runtime version includes TensorFlow 2.1.0. Read the new Training with TensorFlow 2 guide to learn about important differences to consider when using AI Platform Training with TensorFlow 2, compared to TensorFlow 1.
Runtime version 2.1 is also the first runtime version not to support Python 2.7. The Python Software Foundation ended support for Python 2.7 on January 1, 2020. No AI Platform runtime versions released after January 1, 2020 support Python 2.7.
Runtime version 2.1 also updates many other dependencies; see the runtime version list for more details.
Runtime version 2.1 includes scikit-learn 0.22 rather than 0.22.1. This is a known issue, and the release notes will be updated when runtime version 2.1 includes scikit-learn 0.22.1.
When you train with runtime version 2.1 or later, AI Platform Training uses the chief
task name to represent the master VM in the TF_CONFIG
environment variable. This environment variable is important for distributed training with TensorFlow. For runtime version 1.15 and earlier, AI Platform Training uses the master
task name instead, but this task name is not supported in TensorFlow 2.
However, by default, AI Platform Training continues to use the master
task name in custom container training jobs. If you are performing multi-worker distributed training with TensorFlow 2 in a custom container, set the new trainingInput.useChiefInTfConfig
field to true
when you create a custom container training job in order to use the chief
task name.
March 06, 2020
AI Platform TrainingThe built-in linear learner algorithm and the built-in wide and deep algorithm now use TensorFlow 1.14 for training. They previously used TensorFlow 1.12.
The single-replica version of the built-in XGBoost algorithm now uses XGBoost 0.81 for training. It previously used XGBoost 0.80.
March 05, 2020
AI Platform PipelinesAI Platform Pipelines is now available in beta. AI Platform Pipelines makes it easier to get started with MLOps by saving you the difficulty of setting up Kubeflow Pipelines with TensorFlow Extended (TFX). Kubeflow Pipelines is an open source platform for running, monitoring, auditing, and managing machine learning (ML) pipelines on Kubernetes. TFX is an open source project for building ML pipelines that orchestrate end-to-end ML workflows.
February 11, 2020
AI Platform TrainingYou can now set a maximum running time when you create a training job. If your training job is still running after this duration, AI Platform Training cancels the job. Set the maximum running time by specifying the scheduling.maxRunningTime
field.
February 10, 2020
AI Platform PredictionThe known issue with using custom prediction routines together with runtime version 1.15 and Python 3.7 has been fixed. The issue was described in a January 23, 2020 release note.
You can now use custom prediction routines with runtime version 1.15 and Python 3.7.
February 05, 2020
AI Platform PredictionThe GPU compatibility issue that was described in the January 7, 2020 release note has been resolved. You can now use GPUs to accelerate prediction on runtime version 1.15.
The GPU compatibility issue that was described in the January 7, 2020 release note has been resolved. You can now use GPUs to accelerate training on runtime version 1.15.
January 29, 2020
AI Platform PredictionAI Platform Prediction documentation has been reorganized. The new information architecture only includes documents that are relevant to AI Platform Prediction.
Previously, documentation for AI Platform Prediction and AI Platform Training were grouped together. You can now view AI Platform Training documentation separately. Some overviews and tutorials that are relevant to both products are available in the overall AI Platform documentation.
AI Platform Training documentation has been reorganized. The new information architecture only includes documents that are relevant to AI Platform Training.
Previously, documentation for AI Platform Training and AI Platform Prediction were grouped together. You can now view AI Platform Prediction documentation separately. Some overviews and tutorials that are relevant to both products are available in the overall AI Platform documentation.
January 28, 2020
AI Platform TrainingAI Platform Training runtime version 1.15 now supports training with TPUs using TensorFlow 1.15.
January 23, 2020
AI Platform PredictionCreating an AI Platform Prediction custom prediction routine that uses runtime version 1.15 and Python 3.7 might fail due to a problem with a dependency.
As a workaround, use runtime version 1.15 with Python 2.7 or use a different runtime version when you create your model version.
January 22, 2020
AI Platform PredictionAI Explanations no longer supports AI Platform Prediction runtime version 1.13. AI Explanations now supports runtime versions 1.14 and 1.15. Learn more about AI Platform Prediction runtime versions supported by AI Explanations.
January 21, 2020
Deep Learning VM ImagesM41 release
TensorFlow Enterprise 2.1 images are now available.
MXNet upgraded to 1.5.1.
PyTorch upgraded to 1.4.0.
XGBoost upgraded to 0.90.
January 15, 2020
AI Platform PredictionThe price of using NVIDIA Tesla T4 GPUs for online prediction has changed from $0.9500 per hour to $0.3500 per hour.
GPUs for online prediction are currently only available when you deploy your model in the us-central1
region and use a Compute Engine (N1) machine type.
January 14, 2020
AI Platform TrainingThe price of using NVIDIA Tesla T4 GPUs for training has changed. The following table describes the pricing change for various geographic areas:
Geographic area | Old price per hour | New price per hour |
---|---|---|
Americas |
$0.9500 | $0.3500 |
Europe |
$1.0300 | $0.3800 |
Asia Pacific |
$1.0300 | $0.3900 |
Read more about using GPUs for training.
January 08, 2020
Deep Learning ContainersTensorFlow Enterprise environments are now available. Use TensorFlow Enterprise with Deep Learning Containers.
January 07, 2020
AI Platform PredictionModel versions that use both runtime version 1.15 and GPUs fail due to a compatibility issue with the CuDNN library, which TensorFlow depends on.
As a workaround, do one of the following:
- If you want to use GPUs to accelerate prediction, then use runtime version 1.14 or earlier.
- If you want to use runtime version 1.15, then do not use GPUs for your model version.
Training jobs that use both runtime version 1.15 and GPUs fail due to a compatibility issue with the CuDNN library, which TensorFlow depends on.
As a workaround, do one of the following:
- If you want to use GPUs to accelerate training, then use runtime version 1.14 or earlier.
- If you want to use runtime version 1.15, then do not use GPUs for your training job.
December 20, 2019
AI Platform TrainingVPC Service Controls now supports AI Platform Training. Learn how to use a service perimeter to protect your training jobs. This functionality is in beta.
December 19, 2019
AI Platform PredictionAI Platform runtime version 1.15 is now available for prediction. This version supports TensorFlow 1.15.0 and includes other packages as listed in the runtime version list.
Runtime version 1.15 is the first runtime version to support serving predictions using Python 3.7, instead of Python 3.5. Runtime version 1.15 also still supports Python 2.7. Learn about specifying the Python version for prediction.
AI Platform Training now offers two built-in algorithms to train a machine learning model on image data without writing your own training code:
Both image algorithms are available in beta.
AI Platform runtime version 1.15 is now available for training. This version supports TensorFlow 1.15.0 and includes other packages as listed in the runtime version list.
Runtime version 1.15 is the first runtime version to support training using Python 3.7, instead of Python 3.5. Runtime version 1.15 also still supports Python 2.7. Learn about specifying the Python version for training.
Training with TPUs is not supported in runtime version 1.15 at this time.
December 10, 2019
AI Platform PredictionStarting January 1, 2020, the Python Software Foundation will no longer support Python 2.7. Accordingly, any runtime versions released after January 1, 2020 will not support Python 2.7.
Starting on January 13, 2020, AI Platform Training and AI Platform Prediction will support each runtime version for one year after its release date. You can find the release date of each runtime version in the runtime version list.
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.
24 months after the release date: AI Platform Prediction automatically deletes all model versions that use the runtime version.
This policy will be applied retroactively on January 13, 2020. For example, since runtime version 1.0 was released over 24 months ago, AI Platform Training and AI Platform Prediction no longer support it. There will be a three-month grace period (until April 13, 2020) before AI Platform Prediction automatically deletes model versions that use the oldest runtime versions.
The following table describes the first two important dates that mark the end of support for runtime versions:
Date | Runtime versions affected | Change in functionality |
---|---|---|
January 13, 2020 | 1.0, 1.1, 1.2, 1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 1.10, 1.11, 1.12 | You can no longer create training jobs, batch prediction jobs, or model versions using these runtime versions. |
April 13, 2020 | 1.0, 1.1, 1.2, 1.4, 1.5, 1.6 | AI Platform Prediction automatically deletes any model versions using these runtime versions. |
To learn about when availability ends for every runtime version, see the runtime version list.
Starting on January 13, 2020, runtimeVersion
and pythonVersion
will become required fields when you create Job
or Version
resources. Previously, runtimeVersion
defaulted to 1.0
and pythonVersion
defaulted to 2.7
.
Starting January 1, 2020, the Python Software Foundation will no longer support Python 2.7. Accordingly, any runtime versions released after January 1, 2020 will not support Python 2.7.
Starting on January 13, 2020, AI Platform Training and AI Platform Prediction will support each runtime version for one year after its release date. You can find the release date of each runtime version in the runtime version list.
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.
24 months after the release date: AI Platform Prediction automatically deletes all model versions that use the runtime version.
This policy will be applied retroactively on January 13, 2020. For example, since runtime version 1.0 was released over 24 months ago, AI Platform Training and AI Platform Prediction no longer support it. There will be a three-month grace period (until April 13, 2020) before AI Platform Prediction automatically deletes model versions that use the oldest runtime versions.
The following table describes the first two important dates that mark the end of support for runtime versions:
Date | Runtime versions affected | Change in functionality |
---|---|---|
January 13, 2020 | 1.0, 1.1, 1.2, 1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 1.10, 1.11, 1.12 | You can no longer create training jobs, batch prediction jobs, or model versions using these runtime versions. |
April 13, 2020 | 1.0, 1.1, 1.2, 1.4, 1.5, 1.6 | AI Platform Prediction automatically deletes any model versions using these runtime versions. |
To learn about when availability ends for every runtime version, see the runtime version list.
Starting on January 13, 2020, AI Platform Training will automatically delete the history of each training job 120 days after it is completed. A training job is considered completed when the job enters the SUCCEEDED
, FAILED
, or CANCELLED
state.
This policy will be applied retroactively on January 13, 2020: all jobs that were completed September 15, 2019 or earlier will be deleted.
Starting on January 13, 2020, runtimeVersion
and pythonVersion
will become required fields when you create Job
or Version
resources. Previously, runtimeVersion
defaulted to 1.0
and pythonVersion
defaulted to 2.7
.
December 03, 2019
AI Platform PredictionYou cannot enable request-response logging for AI Platform Prediction when you create a model version. Instead, you must first create a model version without request-response logging enabled, then enable request-response logging by sending a projects.models.versions.patch
request to the REST API.
November 27, 2019
AI Platform TrainingAI Platform Training no longer supports TPUs in runtime version 1.12. You can still train using TPUs in runtime versions 1.13 and 1.14.
November 20, 2019
AI Platform PredictionAI Explanations now offers feature attributions through AI Platform Prediction. This feature is available in Beta. To gain more insight on your model's predictions, you can use feature attributions based on the sampled Shapley and integrated gradients methods. Try the example notebooks to get started, and refer to the AI Explainability Whitepaper to learn more.
AI Platform Training now offers a built-in distributed XGBoost algorithm to train a machine learning model without writing your own training code. This algorithm is available in beta.
The built-in distributed XGBoost algorithm provides functionality similar to the existing single-replica version of the built-in XGBoost algorithm, but it lets you speed up training on large datasets by using multiple virtual machines in parallel. The algorithm also lets you use GPUs for training.
The built-in distributed XGBoost algorithm does not support automatic preprocessing of data.
November 01, 2019
Deep Learning VM ImagesYou can now create a TensorFlow Enterprise Deep Learning VM Image. TensorFlow Enterprise image families provide users with a Google Cloud Platform optimized distribution of TensorFlow with long-term version support. To learn more about TensorFlow Enterprise, read the TensorFlow Enterprise overview.
October 28, 2019
AI Platform TrainingWe now recommend that you use Compute Engine machine types when you create new AI Platform Training jobs. These machine types offer the greatest flexibility for customizing the virtual CPU (vCPU), RAM, GPU, and TPU resources that your jobs use.
The older machine types available for training, which were previously referred to as "AI Platform Training machine types," are now called "legacy machine types" in the AI Platform Training documentation.
October 24, 2019
AI Platform PredictionMany Compute Engine (N1) machine types are now available for online prediction in beta, in addition to the existing legacy (MLS1) machine types. When you create a model version with a Compute Engine machine type, you can allocate virtual machines with more vCPU and memory resources for your online prediction nodes, improving throughput of predictions or reducing latency. Additionally, you can use GPUs with these new machine types and deploy TensorFlow models up to 2 GB in size. The machine types are currently only available in the us-central1
region.
Learn more about the features, limitations, and usage of Compute Engine (N1) machine types. Model versions that use Compute Engine (N1) machine types, including with GPUs, are available at no charge until November 14, 2019. Read about the pricing for these machine types that goes into effect on November 14, 2019.
Model versions that use one of the new Compute Engine (N1) machine types and scale to use more than 40 prediction nodes may exhibit high latency when handling online prediction requests. In this case, AI Platform Prediction may also drop requests.
For the best performance until this issue is resolved, do not scale your model version to use more than 40 nodes.
The default max model size for model versions that use a legacy (MLS1) machine type has increased from 250 MB to 500 MB.
October 11, 2019
Deep Learning VM ImagesM36 release
The TensorFlow 2.0 image is out of experimental.
What-If Tool (witwidget
) upgraded to 1.4.2 for TensorFlow 1.x images.
October 04, 2019
AI Platform PredictionThe us-west2
(Los Angeles), us-east4
(N. Virginia), and europe-north1
(Finland) regions are now available for batch prediction. Note that us-east4
was already available for online prediction.
Additionally, the us-west1
(Oregon) and europe-west4
(Netherlands) regions, which were already available for training, are now available for batch prediction.
Read about pricing for batch prediction in these regions.
The us-west2
(Los Angeles), us-east4
(N. Virginia), and europe-north1
(Finland) regions are now available for training. You can use NVIDIA Tesla P4 GPUs for training in us-west2
and us-east4
.
Read about pricing for training in these regions, including pricing for accelerators.
September 16, 2019
AI Platform PredictionThe What-If Tool can be used to inspect models deployed on AI Platform Prediction, and to compare two models. Learn how to use the What-If Tool with AI Platform Prediction.
September 09, 2019
AI Platform TrainingRuntime version 1.14 now supports training with TPUs using TensorFlow 1.14.
September 06, 2019
AI Platform PredictionWhen you deploy a model version to AI Platform Prediction, you can now configure AI Platform Prediction to log a sample of online prediction requests received by the model together with the responses it sends to these requests. AI Platform Prediction saves these request-response pairs to BigQuery. This feature is in beta.
Learn how to how to enable request-response logging and read about the configuration options for this type of logging.
August 28, 2019
AI Platform PredictionThe documentation for AI Platform Notebooks has moved to a new location.
Training with custom containers is now generally available.
Using Compute Engine machine types for your training configuration is now generally available.
NVIDIA Tesla P4 and NVIDIA Tesla T4 GPUs are now generally available for training. Read about using GPUs for training and learn about GPU pricing.
The documentation for AI Platform Notebooks has moved to a new location.
August 26, 2019
AI Platform TrainingAI Platform Training now supports using Cloud TPU devices with TPU v3 configurations to accelerate your training jobs. TPU v3 accelerators for AI Platform Training are available in beta.
Learn more about how to configure your training job to use TPU v3 accelerators and read about TPU v3 pricing.
M34 release
JupyterLab upgraded to 1.0 on all images.
PyTorch upgraded to 1.2.
August 22, 2019
AI Platform PredictionContinuous evaluation for AI Platform Prediction is now available in beta. When you create a continuous evaluation job, AI Platform Data Labeling Service assigns human reviewers to provide ground truth labels for a portion of your model version's online predictions; alternatively, you can provide your own ground truth labels. Then Data Labeling Service compares these labels to your model version's predictions to calculate daily evaluation metrics.
Learn more about continuous evaluation.
August 16, 2019
AI Platform PredictionAI Platform runtime versions 1.13 and 1.14 now include numpy 1.16.4 instead of numpy 1.16.0. View the runtime version list for the full list of packages included in runtime versions 1.13 and 1.14.
AI Platform runtime versions 1.13 and 1.14 now include numpy 1.16.4 instead of numpy 1.16.0. View the runtime version list for the full list of packages included in runtime versions 1.13 and 1.14.
August 01, 2019
AI Platform PredictionThe AI Platform Prediction Training and Prediction documentation has been reorganized. Previously, documentation for using AI Platform Prediction with specific machine learning frameworks was separated into sections. You can now navigate to all Training and Prediction documentation from the AI Platform documentation home page.
The AI Platform Training Training and Prediction documentation has been reorganized. Previously, documentation for using AI Platform Training with specific machine learning frameworks was separated into sections. You can now navigate to all Training and Prediction documentation from the AI Platform documentation home page.
July 19, 2019
AI Platform PredictionAI Platform runtime version 1.14 is now available for prediction. This version supports TensorFlow 1.14.0 and includes other packages as listed in the runtime version list.
AI Platform runtime version 1.12 now supports TensorFlow 1.12.3. View the runtime version list for the full list of packages included in runtime version 1.12.
AI Platform runtime version 1.14 is now available for training. This version supports TensorFlow 1.14.0 and includes other packages as listed in the runtime version list.
Training with TPUs is not supported in runtime version 1.14 at this time.
AI Platform runtime version 1.12 now supports TensorFlow 1.12.3. View the runtime version list for the full list of packages included in runtime version 1.12.
July 17, 2019
AI Platform PredictionThe prediction input format for the following built-in algorithms has changed:
Instead of a raw string, make sure to format each instance as a JSON with a "csv_row" key and "key" key. This "key" is useful for doing batch predictions using AI Platform Prediction. For online predictions, you can pass in a dummy value to the "key" key in your input JSON request. For example:
{"csv_row": "1, 2, 3, 4, 0, abc", "key" : "dummy-key"}
See the Census Income tutorial for an example.
The prediction input format for the following built-in algorithms has changed:
Instead of a raw string, make sure to format each instance as a JSON with a "csv_row" key and "key" key. This "key" is useful for doing batch predictions using AI Platform Training. For online predictions, you can pass in a dummy value to the "key" key in your input JSON request. For example:
{"csv_row": "1, 2, 3, 4, 0, abc", "key" : "dummy-key"}
See the Census Income tutorial for an example.
July 12, 2019
Deep Learning VM ImagesM30 release
R upgraded to version 3.6.
TensorFlow: added support for using Python 3.7.
R Notebooks are no longer dependent on a Conda environment.
Fix for the bug when Nvidia driver is not installed if the user does not have the Google Cloud Storage API enabled.
What-If Tool (witwidget
) fixes for TensorFlow 1.14.
Miscellaneous bug fixes.
July 01, 2019
Deep Learning VM ImagesM28 release
What-If Tool (witwidget
) added to DLVM.
Fixed TensorFlow 1.14 issues.
Miscellaneous bug fixes.
June 24, 2019
Deep Learning ContainersAI Platform Deep Learning Containers is now available in beta. AI Platform Deep Learning Containers lets you quickly prototype with a portable and consistent environment for developing, testing, and deploying your AI applications.
Visit the AI Platform Deep Learning Containers overview and the guide to getting started with a local deep learning container.
June 20, 2019
Deep Learning VM ImagesM27.1 release updates
TensorFlow upgraded to: 1.14.0.
TensorFlow 2.0 upgraded to: Beta 1.
Miscellaneous bug fixes.
June 19, 2019
AI Platform PredictionThe asia-southeast1
(Singapore) region is now available for batch prediction.
The asia-southeast1
(Singapore) region is now available for training. You can use P4 or T4 GPUs for training in the region. Read about pricing for training in asia-southeast1
, including pricing for accelerators.
June 18, 2019
AI Platform TrainingRuntime version 1.13 now supports training with TPUs using TensorFlow 1.13.
Support for training with TPUs in runtime version 1.11 ended on June 6, 2019.
June 17, 2019
Deep Learning VM ImagesM27 release
New ML framework added: CNTK 2.7 from Microsoft.
New ML framework added: Caffe 1.0 BVLC from UC Berkeley.
Updated TensorFlow 2.0 Beta0.
Miscellaneous bug fixes.
June 12, 2019
AI Platform TrainingYou can now view monitoring data for training jobs directly within the AI Platform Training Job Details page in the Cloud Console. The following charts are available:
- CPU, GPU, and memory utilization, broken down by master, worker, and parameter servers.
- Network usage: the rate per second of bytes sent and received.
Learn more about how to monitor resource utilization for your training jobs.
There are new options for filtering jobs within the AI Platform Training Jobs page in the Cloud Console. You can filter jobs by Type and by whether or not the job used HyperTune.
Learn more about how to filter your training jobs.
You can now view and sort hyperparameter tuning trials within the AI Platform Training Job Details page in the Cloud Console. If your training job uses hyperparameter tuning, your Job Details page includes a HyperTune trials table, where you can view metrics such as RMSE, learning rate, and training steps. You can also access logs for each trial. This table makes it easier to compare individual trials.
Learn more about how to view your hyperparameter tuning trials.
June 05, 2019
AI Platform PredictionYou can now specify a service account for your model version to use when you deploy a custom prediction routine to AI Platform Prediction. This feature is in beta.
June 03, 2019
AI Platform TrainingYou can now create AI Platform Notebooks instances with R and core R packages installed. Learn how to install R dependencies, and read guides for using R with BigQuery in AI Platform Notebooks and using R and Python in the same notebook.
May 29, 2019
Deep Learning VM ImagesM26 release
RAPIDS updated to 0.7.
Faster driver installation time for common TensorFlow and PyTorch images.
You can now use Deep Learning VMs without a public IP address if you have enabled Google Private Access.
Miscellaneous bug fixes.
May 03, 2019
AI Platform PredictionAI Platform runtime version 1.12 now supports TensorFlow 1.12.2. View the runtime version list for the full list of packages included in runtime version 1.12.
T4 GPUs are now in beta for AI Platform Training. For more information, see the guides to using GPUs, their regional availability, and their pricing.
AI Platform runtime version 1.12 now supports TensorFlow 1.12.2. View the runtime version list for the full list of packages included in runtime version 1.12.
M25 release
New image added: CUDA 10.1.
PyTorch upgraded to 1.1.0.
fastai upgraded to 1.0.52.
MXNet upgraded to 1.4.0 (and now based on CUDA 10.0 images).
Chainer upgraded to 5.4.0.
April 26, 2019
Deep Learning VM ImagesM24 release
We now support two authorization modes in the new release: single user mode and service account mode3.
rpy2 is now pre-installed in the R image.
A plugin for editing metadata of cells is now pre-installed.
jupyterlab-celltags JupyterLab extension is now pre-installed.
Fixed bug with sudo (now you can use sudo from the JupyterLab terminal).
Downloading files from JupyterLab file browser is now working.
April 25, 2019
AI Platform PredictionAI Platform Prediction now supports custom prediction routines in beta. Custom prediction routines let you provide AI Platform Prediction with custom code to use when it serves online predictions from your deployed model. This can be useful for preprocessing prediction input, postprocessing your model's predictions, and more.
Work through a tutorial on deploying a custom prediction routine with Keras or one on deploying a custom prediction routine with scikit-learn.
AI Platform Prediction now supports custom transformers for scikit-learn pipelines in beta. This lets you provide AI Platform Prediction with custom code to use during online prediction. Your deployed scikit-learn pipeline uses this code when it serves predictions.
Work through a tutorial on training and deploying a custom scikit-learn pipeline.
AI Platform Prediction now supports logging of your prediction nodes' stderr
and stdout
streams to Stackdriver logging during online prediction. Stream logging is in beta. You can enable this type of logging in addition to—or in place of—the access logging that was already available. It can be useful for understanding how your deployment handles prediction requests.
April 10, 2019
AI Platform Data Labeling ServiceAI Platform Data Labeling Service Beta has been released.
March 15, 2019
Deep Learning VM ImagesM22 release
Tensorflow upgraded to version 1.13.
Fairing now preinstalled.
cookiecutter and seaborn now preinstalled.
More descriptive serial logs to help customers debug common issues.
Misc bug fixes.
Due to incompatibilities between Tensorflow 1.13 (which requires Numpy 1.16.2 or greater) and the latest Intel optimized version of Numpy (which is 1.15) we are not using the intel optimized versions of Numpy and Scipy for this release.
February 21, 2019
Deep Learning VM ImagesM20 release
TensorFlow and Pytorch GPU images switch between CPU-only/GPU-enabled binaries at startup depending on whether GPUs are attached.
SSH is not disabled during NVIDIA driver installation on GPU images.
Due to incompatibilities between the latest kernel update (Debian 9.8) and Docker, we have put a hold on the kernel updates for this release (that is, apt-mark hold linux-image-4.9.0-8-amd64
). If you require the latest kernel, you can run sudo apt-mark unhold linux-image-4.9.0-8-amd64 && sudo apt upgrade
, but we cannot guarantee that Docker or our direct JupyterLab link from Marketplace will function correctly if you force the upgrade.
January 29, 2019
Deep Learning VM ImagesM19 release
New TensorFlow 2.0 (experimental) flavor is added.
New experimental ability to use Deep Learning VMs with special Web proxy, instead of SSHing to the VM.
January 14, 2019
Deep Learning VM ImagesM16 release
New MXNet 1.3 (experimental) flavor is added.
December 19, 2018
Deep Learning VM ImagesGeneral Availability
Launched the new 1.0 version of AI Platform Deep Learning VM Image.
M15 release
BigQuery magic plugin now preloaded all the time.
Jupyter SQL integration now pre-installed and SQL plugin now preloaded.
TensorFlow images now include bazel pre-installed.
Python Dataproc client now pre-installed on all our images.
fastai updated to the latest version 1.0.38.
December 10, 2018
Deep Learning VM ImagesM14 release
Fixed bug that was resulting in a broken Git UI in some cases.
Fast.Ai updated to 1.0.36.
December 05, 2018
Deep Learning VM ImagesM13 release
Integrates fix for speed regression in linear models when using TensorFlow with Intel® MKL DNN.
Adds Git-Jupyter integration.
November 20, 2018
Deep Learning VM ImagesM12 release
Chainer is now upgraded to 5.0.0 (and CuPy to 5.0.0).
CuDNN updated to 7.4.
TensorRT5 updated to GA.
XGBoost updated to 0.81.
Images now have papermill pre-installed.
Ability to change Jupyter UI that is running on the port 8080, currently supported: Lab and Notebook.
November 13, 2018
Deep Learning VM ImagesM11.1 release
Fixed an issue where users were locked out of apt
after startup due to a package needing configuration. If you are using an M11 image and are experiencing issues with apt, please either recreate your VM or run sudo dpkg --configure -a
to clear the lock.
November 08, 2018
Deep Learning VM ImagesM11 release
All GPU images install NVIDIA driver 410.72.
TensorFlow updated to v1.12.0.
PyTorch 0.4 image now uses conda for package management.
October 23, 2018
Deep Learning VM ImagesM10 release
PyTorch 1.0 updated to the latest build as of October 23.
fastai
updated to 1.0.12.
fastai
course materials are now available at $HOME/tutorials/fastai/
.
Chainer UI updated to 0.6.0.
Chainer MN updated to 1.3.1.
Fixed a bug that was causing Intel packages to be overwritten.
October 10, 2018
Deep Learning VM ImagesM9 release
Intel Optimized Python packages are installed in all distributions:
- NumPy
- SciPy
scikit-learn
- TensorFlow (when applicable)
PyTorch 1.0 (Experimental) images include support for [conda](https://conda.io/)
and [fastai](http://fast.ai/)
.
Chainer updated from v4.4.0 to v4.5.0.
September 27, 2018
Deep Learning VM ImagesM8 release
New XGBoost images:
xgboost-<var>VERSION</var>-cu92-experimental
xgboost-<var>VERSION</var>-cpu-experimental
New CUDA 10.0 image (common-cu100) with the following NVIDIA stack in it:
- CuDNN 7.3
- NCCL 2.3.4
- Driver 410.48
- TensorRT 5
TensorFlow updated from v1.10.1 to v1.11.0.
TensorFlow now compiled with CUDA 10.0 and CuDNN 7.3.
Common CUDA 9.2 image now has latest NCCL 2.3.4
Common CUDA 9.0 image now has:
- latest NCCL 2.3.4
- latest CuDNN 7.3
- TensorRT 5.0.0
Following packages are now pre-installed on the images:
htop
protobuf-compiler
tree
After SSHing to the instance you now will see the exact revision of the image in the header.
September 18, 2018
Deep Learning VM ImagesM7.1 release
Introducing new experimental images with PyTorch 1.0RC. New image families are:
pytorch-1-0-cu92-experimental
pytorch-1-0-cpu-experimental
September 12, 2018
Deep Learning VM ImagesM7 release
Chainer updated from v4.3.0 to v4.4.0.
Better integration with BigQuery.
Pillow has been replaced with the faster Pillow-SIMD package.
minikube
is now pre-installed.
New simplified image families introduced:
tf-latest-gpu
pytorch-latest-gpu
chainer-latest-gpu-experimental
Jupyter now running on behalf of its own user (not root).
August 30, 2018
Deep Learning VM ImagesM6 release
Introducing experimental images: these images bring new frameworks for you to try out, but they come with no guarantees of future support. Current experimental images:
- Chainer (4.3)
All images now have kubectl
installed.
TensorFlow updated from v1.10.0 to v1.10.1.
August 14, 2018
Deep Learning VM ImagesM5 release
All images now have Docker and/or NVIDIA Docker pre-installed.
TensorFlow and PyTorch images now include pre-baked tutorials.
GPU flavors of TensorFlow and PyTorch images now swap binaries to the CPU optimized binaries during the first boot if the instance does not have a GPU.
July 31, 2018
Deep Learning VM ImagesM4 release
Includes Tensorfow Serving: model server binary at /usr/local/bin/tensorflow_model_server and tensorflow-serving-api preinstalled.
Integration with Colab: default JupyterLab instance can be connected as a Colab backend.
Upgraded to support CUDA 9.2 (note this changes the pytorch family name).
Fixed an issue with CUDA linking in the build process, binaries up to 10% faster now.
July 17, 2018
Deep Learning VM ImagesM3 release
New common image with CUDA 9.0 has been introduced.
The following changes are included in this release:
Bug that was preventing Jupyter Notebook from working correctly has been resolved.
July 11, 2018
Deep Learning VM ImagesM2 release
TensorFlow updated to version 1.9.0.
New public Google Group for users: google-dl-platform
July 02, 2018
Deep Learning VM ImagesBeta launch
AI Platform Deep Learning VM Image is available as a beta release.