AI Platform release notes

M126 release

• Base CUDA 12.3 container images are now available.
• Base CUDA 12.4 container images are now available.
• PyTorch 2.4.0 with CUDA 12.4 and Python 3.10 container images are now available.
• Upgraded R from 4.4.1 to 4.4.2 for R container images.

M126 release

• CUDA 12.4 VM images are now available.
• PyTorch 2.4.0 with CUDA 12.4 and Python 3.10 VM images are now available.
• Upgraded R from 4.4.1 to 4.4.2 for R VM images.
• One or more framework versions have reached their end of patch and support dates. To view end of patch and support dates, see Supported framework versions. To create a VM instance using an image family that has reached its end of patch and support date, you must specify an image from the image family when you create the VM instance. To list images from an image family name after its end of patch and support date, include the --show-deprecated flag in your gcloud compute images list command, or select Show deprecated images when creating an instance in the Google Cloud console.

M125 release

• TensorFlow 2.17 container images are now available.

M125 release

• TensorFlow 2.17 VM images are now available.

M124 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.

M123 release

• Hugging Face Text Generation Inference 2.1 GPU container images are now available.

M123 release

• TensorFlow 2.16 images are now available.

M122 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.

M121 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.

M120 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.

M118 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.

M117 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 renamed tf-2-12-cu118.

M117 release

• Fixed an issue wherein the latest container had a deprecation-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.

M116 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 to tf-2-15-cu121.
• Deprecated the tf-2-15-gpu image family in favor of tf-2-15-cu121.

M115 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.

M114 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.

M113 release

• Miscellaneous bug fixes and improvements in Python 3.10 container images.

M113 release

• Miscellaneous bug fixes and improvements in Python 3.10 images.

M112 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.

M111 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.

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.

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.

M109 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.

M108 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)

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)

M108 release

• Miscellaneous software updates.

M108 release

• The image name common-container-experimental was changed to common-container. The related image family name wasn't changed.
• Miscellaneous software updates.

M107 release

• Miscellaneous software updates.

M107 release

• Miscellaneous software updates.

M106 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.

M105 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)

• 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)

• 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)

• 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)

• 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.

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.

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.

M103 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.

Runtime 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.

M102 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.

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 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.

M100 release

• Regular package updates.

M100 release

• Migrated the Docker proxy agent to use a systemctl service.
• Regular package updates.

M99 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.

M98 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.

M97 release

• Regular package updates.

M97 release

• Improved the startup time for Ubuntu GPU images.
• Regular package updates.

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.
• 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.

M95 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.

M94 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.

M93 release

• Starting with this release, information on known vulnerabilities for Deep Learning Containers images is now available on Cloud Storage.

M93 release

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.
• 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.

Pre-built PyTorch containers for PyTorch 1.11 are available for training. You can use these containers to train with CPUs, GPUs, or TPUs.

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.

M91 release

• PyTorch 1.11 and PyTorch XLA 1.11 are now available in both Deep Learning Containers and Deep Learning VM Images.

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.

M90 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.

Runtime 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.

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 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.

• 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/.

• 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/.

Runtime 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.

M87 release

• Added Artifact Registry's Python keyring authentication library to deep learning vm environments.

M87 release

• Added Artifact Registry's Python keyring authentication library to deep learning vm environments.

M86 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.

Runtime 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.

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.

M85 release

• Regular package refreshment and bug fixes.

M85 release

• CUDA 11.3 Debian-10 image is available.
• Regular package refreshment and bug fixes.

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.

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.

Using interactive shells to inspect training jobs is generally available (GA).

You can use these interactive shells with VPC Service Controls.

M83 release

• PyTorch 1.10 is now available.

M83 release

• PyTorch 1.10 is now available.

M82 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.

M82 release

• The Vertex SDK for Python is available across all deep learning environment products; it was previously available only in TensorFlow images.

M81 release

• Upgraded R to 4.1.

M81 release

• Upgraded R to 4.1.
• Improved Cloud Storage sync logic so that only newer files sync.

Runtime 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.

M80 release

• Updated JupyterLab from 1.x to 3.x.
• Added Jupytext.

M80 release

• Updated JupyterLab from 1.x to 3.x.
• Added Jupytext.

Training 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.

M79 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 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.

Runtime 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.

M78 release

• Updated TensorFlow Enterprise patch version 2.3.3 to 2.3.4.

M78 release

• Updated TensorFlow Enterprise patch version 2.3.3 to 2.3.4.

M77 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.

You can use TPU Pods for training. This feature is available in Preview.

M76 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.

You 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.

M75 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.

• 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.

Cloud 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 to ShipmentRoute.transitions.
• ShipmentRoute.Visit.delay_before_start and ShipmentRoute.delay_before_vehicle_end are changed to ShipmentRoute.Transition.delay_duration.
• ShipmentRoute.Visit.arrival_loads and ShipmentRoute.end_loads are changed to ShipmentRoute.Transition.loads.
• ShipmentRoute.TravelStep.duration is changed to ShipmentRoute.Transition.travel_duration.
• ShipmentRoute.TravelStep.distance_meters is changed to ShipmentRoute.Transition.travel_distance_meters.

M73 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.

M72 release

• Added PyTorch 1.9 and PyTorch/XLA 1.9 containers.

M72 release

• Added PyTorch 1.9 and PyTorch/XLA 1.9 images.

NVIDIA A100 GPUs and accelerator-optimized (A2) machine types are generally available (GA).

Runtime 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.

M71 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.

M70 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.

M69 release

• Updated cuDNN from 8.0.4 to 8.0.5.

M69 release

• Migrated Collection Agent to Cloud Monitoring version 2.

M68 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.

M67 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.

Runtime 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.

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.

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.

AI 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.

M64 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.

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.

Runtime 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.

M63 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.

You 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.

You 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.

M62 release

• Upgraded TensorFlow 2.3 to 2.3.2

• Upgraded TensorFlow 2.1 to 2.1.3

• Miscellaneous bug fixes and updates

M62 release

• Upgraded TensorFlow 2.3 to 2.3.2
• Upgraded TensorFlow 2.1 to 2.1.3
• Miscellaneous bug fixes and updates

The 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.

AI 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.

You 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

Runtime 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.

The 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.

• 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.

You 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.

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.

M56 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

AI Platform Training runtime version 2.2 now supports training with TPUs using TensorFlow 2.2.

Runtime 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.

You 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).

The 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.

Try the TabNet built-in algorithm introductory tutorial.

M54 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

M53 release

TensorFlow Enterprise 2.3 images, including images that support CUDA 11.0, are now available.

Read a new guide to distributed PyTorch training. You can use this guide with pre-built PyTorch containers, which are in beta.

Customer-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.

VPC Service Controls now supports AI Platform Prediction. Learn how to use a service perimeter to protect online prediction. This functionality is in beta.

You 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.

AI 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.

M49 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.

You 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.

M48 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.

AI 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.

AI 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.

Visualization settings for AI Explanations are now available. You can customize how feature attributions are displayed for image data.

Learn more about visualizing explanations.

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.

AI 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.

You 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.

AI 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.

Runtime 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.

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.

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.

Learn more about this change.

AI 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.

• Get started with AI Platform Pipelines.
• Learn more about AI Platform Pipelines and ML pipelines.
• Orchestrate your ML process as a pipeline.

You 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.

AI Platform Training runtime version 1.15 now supports training with TPUs using TensorFlow 1.15.

M41 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.

TensorFlow Enterprise environments are now available. Use TensorFlow Enterprise with Deep Learning Containers.

VPC Service Controls now supports AI Platform Training. Learn how to use a service perimeter to protect your training jobs. This functionality is in beta.

AI 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:

• Built-in image classification algorithm
• Built-in image object detection algorithm

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.

AI 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.

You 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.

Many 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.

M36 release

The TensorFlow 2.0 image is out of experimental.

The 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.

The 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.

Runtime version 1.14 now supports training with TPUs using TensorFlow 1.14.

When 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.

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.

AI 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.

Continuous 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.

AI 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 1.14 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.

M30 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.

M28 release

What-If Tool (witwidget) added to DLVM.

AI 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.

M27.1 release updates

TensorFlow upgraded to: 1.14.0.

TensorFlow 2.0 upgraded to: Beta 1.

The 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.

Runtime version 1.13 now supports training with TPUs using TensorFlow 1.13.

M27 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.

You 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.

You 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.

You 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.

M26 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.

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.

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.

M24 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.

AI 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.

AI Platform Data Labeling Service Beta has been released.

M22 release

Tensorflow upgraded to version 1.13.

Fairing now preinstalled.

cookiecutter and seaborn now preinstalled.

M20 release

TensorFlow and Pytorch GPU images switch between CPU-only/GPU-enabled binaries at startup depending on whether GPUs are attached.

M19 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.

M16 release

New MXNet 1.3 (experimental) flavor is added.

General 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.

Fast.Ai updated to 1.0.36.

Adds Git-Jupyter integration.

M9 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/).

M8 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

M7.1 release

Introducing new experimental images with PyTorch 1.0RC. New image families are:

• pytorch-1-0-cu92-experimental
• pytorch-1-0-cpu-experimental

M6 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.

M5 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.

M4 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).

M3 release

New common image with CUDA 9.0 has been introduced.

M2 release

TensorFlow updated to version 1.9.0.

New public Google Group for users: google-dl-platform

Beta launch

AI Platform Deep Learning VM Image is available as a beta release.