OS and Docker images (original) (raw)

Google Cloud provides images that contain common operating systems, frameworks, libraries, and drivers. Google Cloud optimizes these pre-configured images to support your artificial intelligence (AI) and machine learning (ML) workloads.

This document provides an overview of the images that you use to deploy, manage, and run workloads in your AI Hypercomputer environment.

Understand the image categories

Images are grouped into the following categories:

• AI and ML frameworks and libraries: Docker images that are pre-configured with binaries for ML frameworks and libraries to simplify the creation, training, and use of ML models. On AI Hypercomputer, you can useDeep Learning Software Layer (DLSL) Docker images to run ML models such as NeMO and MaxText on a Google Kubernetes Engine (GKE) cluster.
• Cluster deployment and orchestration: operating system (OS) images that you use to deploy and manage theperformance-optimized infrastructure on which your AI workloads run. You can deploy your AI workloads on GKE clusters, Slurm clusters, or Compute Engine instances. For more information, seeVM and cluster creation overview. The following operating system images are available for cluster or instance deployment:
  • GKE node images: you can use these images to deploy GKE clusters.
  • Slurm OS images: Cluster Toolkit builds and deploys these images, which install the necessary system software for Slurm nodes.
  • Accelerator OS images: you can use these images to create individual or groups of instances.

AI and ML frameworks and libraries

Google Cloud provides Docker images that package popular AI and ML frameworks and libraries. These images provide the software needed to simplify the development, training, and deployment of models on your AI-optimized clusters running on AI Hypercomputer.

JAX AI images

The JAX AI Images (JAII, formerly known as JAX Stable Stack Images) for Google Cloud TPUs and GPUs offer ready-to-use Docker images that contain the JAX framework, a curated collection of compatible libraries, and settings for the Google Cloud infrastructure. JAX AI TPU images come pre-configured with JAX libraries and TPU libraries. JAX AI GPU images come pre-configured with JAX libraries and relevant CUDA/NVIDIA libraries.

The hardware layer

JAX AI images sit on top of the hardware layer which consists of the accelerators (TPU or GPU) and their associated VMs. In order to use a JAX AI image, you need to provision TPU or GPU VMs. You can do this using the TPU API, Compute Engine API, or the GKE API.

The framework layer

The framework layer provides tools and libraries for building ML workloads. JAX AI images provide a pre-configured base for JAX-based ML workloads, including the core JAX library and other essential dependencies, to ensure a consistent and high-performance development experience.

The LibTPU layer in the JAX AI image is specifically built and bundled with the respective JAX version. Using a different JAX version may lead to unexpected behavior or errors.

The CUDA layer in the JAX AI image includes components managed by NVIDIA, such as the NGC CUDA Deep Learning Image, used as the base image of the GPU Training Image. The GPU image also contains Transformer Engine, a custom NVIDIA library for accelerating transformer models on NVIDIA GPUs.

Additional application-specific packages, beyond those provided in the JAX AI image, may be required for your specific machine learning workload.

Libraries in JAX AI Images:

Current JAX AI images

The application layer

You implement your specific ML workloads in the application layer which sits on top of the framework layer. The application layer contains your application specific code, models, and logic, all built using the tools and libraries provided by the framework layer.

While this image provides a robust and well-tested foundation for JAX-based AI workloads, you may need to add application-specific dependencies. When doing so, we recommended that you do so in a way that minimizes interference with the pre-configured base layer, which includes JAX and its core dependencies. Introducing application-level dependencies that override or conflict with the existing dependencies can cause side-effects such as:

• Unexpected Behaviors: your ML workloads may exhibit different behavior than before you introduced additional dependencies to the JAX AI image.
• Performance Regressions: overriding optimized JAX-related libraries can negatively impact the performance benefits provided by the JAX AI image
• Stability Issues: conflicts between your added dependencies and the core JAX dependencies can introduce instability and runtime errors within your application.

Release Cadence

Initially, JAX AI images are provided quarterly with a near-term goal of a synchronized release schedule with every JAX release. This ensures that you can benefit from the latest features and improvements as soon as they are available.

Support

Each JAX AI image release adheres to a limited-time support lifecycle. Within this timeframe, we address specific categories of requests for modifications to existing JAX AI images:

• Security vulnerabilities: we prioritize addressing security vulnerabilities discovered in the base images or dependencies of JAX Stable Stack Docker images. Updated images will be released to mitigate potential risks.
• Breaking changes: in the event of significant breaking changes in the underlying libraries or frameworks used by JAX AI image, Google Cloud evaluates and implements necessary updates to maintain compatibility. This may involve rebuilding Docker images with updated dependencies.

When a security vulnerability or bug is discovered in a library within a JAII, we incorporate the updated library into JAII, pinning all other library versions to maintain overall stability. This results in a new JAII revision.

Minimal change for revisions:

If a bug is found in package "X" within JAX-0.4.30-rev1, we'll update "X" to its next release (for example, v2.0) while trying to keep keeping all other packages unchanged. This results in a new revision: JAX-0.4.30-rev2, which will be released as quickly as possible.

Deep Learning Software Layer (DLSL) Docker images

These images package NVIDIA CUDA, NCCL, an ML framework, and a model. They provide a ready-to-use environment for deep learning workloads. These prebuilt DLSL Docker images work seamlessly with your GKE clusters because we test and verify these images during internal reproducibility and regression testing.

DLSL Docker images provide the following benefits:

• Preconfigured software: DLSL Docker images replicate the setup that internal reproducibility and regression testing use. These images provide a pre-configured, tested, and optimized environment, which saves significant time and effort in installation and configuration.
• Version management: DLSL Docker images are frequently updated. These version updates provide the latest stable version of frameworks and drivers, and the updates also address security patches.
• Infrastructure compatibility: DLSL Docker images are built and tested to work seamlessly with the GPU machine types available on AI Hypercomputer.
• Quickstart instructions: some DLSL Docker images have accompanyingsample recipes that show you how to start your workloads that use the pre-configured images.

NeMo + PyTorch + NCCL gIB plugin

These Docker images are based on the NVIDIA NeMo NGC image. They contain Google's NCCL gIB plugin and bundle all NCCL binaries required to run workloads on each supported accelerator machine. These images also include Google Cloud tools such as gcsfuse and the gcloud CLI for deploying workloads to Google Kubernetes Engine.

NeMo + PyTorch

This Docker image is based on the NVIDIA NeMo NGC image and includes Google Cloud tools such asgcsfuse and gcloud CLI for deploying workloads to Google Kubernetes Engine.

MaxText + JAX toolbox

This Docker image is based on the NVIDIA JAX toolbox image and includes Google Cloud tools such asgcsfuse and gcloud CLI for deploying workloads to Google Kubernetes Engine.

MaxText + JAX stable stack

This Docker image is based on theJAX stable stackand MaxText. This image also includes dependencies such as dnsutils for running workloads on Google Kubernetes Engine.

Cluster deployment and orchestration

OS images include all the necessary software components to deploy an operating system on a compute instance or GKE node. The operating system manages underlying hardware resources, such as accelerators and networking. This provides the compute resources for your AI workload.

GKE node images

GKE deploys clusters using node images. These node images are available for various operating systems such as Container-Optimized OS, Ubuntu, and Windows Server. TheContainer-Optimized OS with containerd (cos_containerd)node images that you need to deploy GKE Autopilot clusters include optimizations to support your AI and ML workloads.

For more information about these node images, seeNode images.

Slurm OS images

Slurm clusters deploy compute and controller nodes as compute instances on Compute Engine.

To provision AI-optimized Slurm clusters, you must useCluster Toolkit. During Slurm cluster deployment, the cluster blueprint automatically builds a custom OS image that installs the required system software for cluster and workload management on the Slurm nodes. You can modify the default blueprints before you deploy them to customize some of the software that your images include.

The following section summarizes the software that the cluster blueprint installs on your A4, A3 Ultra, A3 Mega, and A3 High (8 GPUs) Slurm nodes. Cluster blueprints extend theUbuntu LTS Accelerator OS images.

A4X Max

The A4X Max blueprint available on GitHub includes the following software by default:

• Ubuntu 24.04 LTS
• Slurm: version 25.05.2
• The following Slurm dependencies:
  • munge
  • mariadb
  • libjwt
  • lmod
• Open MPI: version 5.0.x
• NFS client and server
• NVIDIA 580 series drivers
• NVIDIA Enroot
• NVIDIA Pyxis
• The following NVIDIA tools:
  • Data Center GPU Manager (dcgmi)
  • nvidia-container-toolkit: version 1.17.7
  • nvidia-imex
  • nvidia-fabricmanager
• CUDA Toolkit: version 13.0
• NCCL Plugin for A4X Max (nccl-gib-a4x-max-arm64)
• DOCA-OFED driver
• Ops Agent
• Cloud Storage FUSE

A4X

The A4X blueprint available on GitHub includes the following software by default:

• Ubuntu 24.04 LTS
• Slurm: version 25.05.2
• The following Slurm dependencies:
  • munge
  • mariadb
  • libjwt
  • lmod
• Open MPI: version 5.0.x
• NFS client and server
• NVIDIA 570 series drivers
• NVIDIA Enroot
• NVIDIA Pyxis
• The following NVIDIA tools:
  • Data Center GPU Manager (dcgmi)
  • nvidia-container-toolkit: version 1.17.7
  • nvidia-imex
  • nvidia-fabricmanager
• CUDA Toolkit: version 12.8
• NCCL Plugin for A4X (nccl-plugin-gib-arm64)
• Ops Agent
• Cloud Storage FUSE

A4

The A4 blueprint available on GitHub includes the following software by default:

• Ubuntu 22.04 LTS
• Slurm: version 25.05.2
• The following Slurm dependencies:
  • munge
  • mariadb
  • libjwt
  • lmod
• Open MPI: the latest release of 4.1.x
• PMIx: version 4.2.9
• NFS client and server
• NVIDIA 570 series drivers
• NVIDIA enroot container runtime: version 3.5.0 with post-release bugfix
• NVIDIA pyxis
• The following NVIDIA tools:
  • Data Center GPU Manager (dcgmi)
  • nvidia-utils-570
  • nvidia-container-toolkit
  • libnvidia-nscq-570
• CUDA Toolkit: version 12.8
• Infiniband support, including ibverbs-utils
• Ops Agent
• Cloud Storage FUSE

A3 Ultra

The A3 Ultra blueprint available on GitHub includes the following software by default:

• Ubuntu 22.04 LTS
• Slurm: version 24.11.2
• The following Slurm dependencies:
  • munge
  • mariadb
  • libjwt
  • lmod
• Open MPI: the latest release of 4.1.x
• PMIx: version 4.2.9
• NFS client and server
• NVIDIA 570 series drivers
• NVIDIA enroot container runtime: version 3.5.0 with post-release bugfix
• NVIDIA pyxis
• The following NVIDIA tools:
  • Data Center GPU Manager (dcgmi)
  • libnvidia-cfg1-570-server
  • libnvidia-nscq-570
  • nvidia-compute-utils-570-server
  • nsight-compute
  • nsight-systems
• CUDA Toolkit: version 12.8
• Infiniband support, including ibverbs-utils
• Ops Agent
• Cloud Storage FUSE

A3 Mega

The A3 Mega blueprint available on GitHub includes the following software by default:

• Ubuntu 22.04 LTS
• Slurm: version 24.11.2
• The following Slurm dependencies:
  • munge
  • mariadb
  • libjwt
  • lmod
• Open MPI: the latest release of 4.1.x
• PMIx: version 4.2.9
• NFS client and server
• NVIDIA 550 series drivers
• NVIDIA enroot container runtime: version 3.4.1
• NVIDIA pyxis
• The following NVIDIA tools:
  • Data Center GPU Manager (dcgmi)
  • libnvidia-cfg1-550-server
  • libnvidia-nscq-550
  • nvidia-compute-utils-550-server
  • nsight-compute
  • nsight-systems
• CUDA Toolkit: version 12.4
• Infiniband support, including ibverbs-utils
• TCPXO for A3 Mega: specific installer for NCCL TCPXO and the 1.6 Tbps network stack
• Ops Agent
• Cloud Storage FUSE

A3 High

The A3 High (8 GPUs) blueprint available on GitHub includes the following software by default:

• Ubuntu 22.04 LTS
• Slurm: version 24.11.2
• The following Slurm dependencies:
  • munge
  • mariadb
  • libjwt
  • lmod
• Open MPI: the latest release of 4.1.x
• PMIx: version 4.2.9
• NFS client and server
• NVIDIA 535 or 550 series drivers
• NVIDIA enroot container runtime: version 3.4.1
• NVIDIA pyxis
• The following NVIDIA tools:
  • Data Center GPU Manager (dcgmi)
  • libnvidia-cfg1-535/550-server
  • libnvidia-nscq-535/550
  • nvidia-compute-utils-535/550-server
  • nsight-compute
  • nsight-systems
• CUDA Toolkit: version 12.2 or 12.4
• Infiniband support, including ibverbs-utils
• TCPX for A3 High (8 GPUs): specific installer for NCCL TCPX (GPUDirect-TCPX) and the 800 Gbps network stack (4 NICs)
• Ops Agent
• Cloud Storage FUSE

Accelerator OS images

AI Hypercomputer lets you provision individual compute instances or groups of compute instances. If you want to create these instances, you must specify an OS image during instance creation.

Google Cloud offers a suite ofOS images for instance creation. Google Cloud also offers a specialized set of accelerator OS images for AI-optimized instances. These OS images include core drivers for GPU and networking functionality, such as NVIDIA drivers, Mellanox drivers, and their dependencies.

For more information about each OS, see theOperating system details page in the Compute Engine documentation.

Accelerator OS images are available for the Rocky Linux and Ubuntu LTS operating systems.

Rocky Linux accelerator

The followingRocky Linux accelerator OS images are available for each machine series:

Ubuntu LTS accelerator

The followingUbuntu LTS acceleratorOS images are available for each machine series:

What's next

• Review consumption options.