Deep Learning VM release notes

Images with Nvidia 570 drivers have been deprecated as Nvidia officially stopped supporting these versions. You can use images with Nvidia 580 drivers instead.

M132 release

• Ubuntu 22.04 and Ubuntu 24.04 images with Nvidia 580 drivers and CUDA 12.9 stack are now available for Common and PyTorch 2.9 images.

M131 release

• Ubuntu 24.04 images with CUDA 12.8 and PyTorch 2.7 stack are now available for Common and PyTorch images.

M130 release

• Ubuntu 22.04 images with CUDA 12.8 and PyTorch 2.7 stack are now available for Common and PyTorch images.

The following framework versions have reached their end of patch and support dates:

• Base versions with CUDA 12.4 and earlier
• Tensorflow versions 2.17 and earlier
• PyTorch versions 2.3 and earlier

To view the end of patch and support dates, see Supported framework versions.

For more information on what the end of patch and support date means, see Support policy schedule.

If you must use an image after deprecation against Google security recommendations and at your own risk, see After deprecation.

M129 release

• Updated the Dataproc JupyterLab plugin to version 0.1.85.

M128 release

• Except for TensorFlow images, new images don't include conda. This change was made to improve size, performance, and vulnerability management. The existing image names and image family names now point to images and image families that don't include conda (for example: the image name common-cpu-v20250310-debian-11-py310 and corresponding image family name common-cpu-debian-11-py310).
• Images that include conda will be available until at least September 30, 2025. These images now have -conda appended to the name (for example: the image name common-cpu-v20250310-debian-11-py310-conda and corresponding image family name common-cpu-debian-11-py310-conda).
• All TensorFlow images and image families still include conda, but M128 image names and image family names have -conda appended. Specifying images or image families without -conda appended references older images, which also include conda.

M127 release

• Fixed an issue related to ownership of the home directory when using authorized ssh keys.

The following framework versions have reached their end of patch and support dates:

• Tensorflow versions 2.15 and earlier
• PyTorch versions 2.1 and earlier
• Base versions with CUDA 12.2 and earlier

To view the 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.

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 VM 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

• TensorFlow 2.16 images are now available.

M122 release

• Updated Nvidia drivers to version 550.90.07 to fix vulnerabilities.

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 images to TensorFlow 2.15.1.
• Added Ubuntu 22.04 support for CPU images, and for GPU images using CUDA 12.1 or higher.

M119 release

• Fixed an issue wherein Dataproc extensions caused JupyterLab to crash when remote kernels weren't 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.

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 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, 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 images.

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 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.

TensorFlow 2.9 images are deprecated.

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 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

• 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.

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 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.

TensorFlow Enterprise 2.1 has reached the end of its support period. See Version details.

M103 release

• Upgraded PyTorch to 1.13.1.
• Minor bug fixes and improvements.

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 VM images are now deprecated.
• Regular security patches and package upgrades.

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.
• 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.
• 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

• 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

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.
• 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.