AWS Base OS AMI Installation

The following table outlines the steps required when using a standard (non-DL) AMI image to set up the EC2 instance.

Objective	Steps
Run PyTorch on AWS Base OS AMI	Install Intel Gaudi SW stack Install PyTorch Set up Python for Models Run models using Intel Gaudi Model References
Run Using Containers on AWS Base OS AMI	Install Intel Gaudi SW stack Set up Container Usage Pull Prebuilt Containers or Build Docker Images from Intel Gaudi Dockerfiles Set up Python for Models Run models using Intel Gaudi Model References

Note

Before installing the below packages and Dockers, make sure to review the currently supported versions and operating systems listed in the Support Matrix.

Run PyTorch on AWS Base OS AMI

Set Up Intel Gaudi SW Stack

Installing the package with internet connection available allows the network to download and install the required dependencies for the Intel® Gaudi® software package (apt get, yum install or pip install etc.). The installation contains the following Installers:

• habanalabs-graph – installs the graph compiler and the run-time.

• habanalabs-thunk – installs the Thunk library.

• habanalabs-dkms – installs the habanalabs, habanalabs_cn, habanalabs_en and habanalabs_ib driver. The habanalabs_ib driver is supported on Gaudi 2 only.

• habanalabs-rdma-core - installs IBVerbs libraries which provide Intel Gaudi’s libhlib along with libibverbs. The habanalabs-rdma-core package is supported on Gaudi 2 only.

• habanalabs-firmware - installs the Gaudi firmware.

• habanalabs-firmware-tools – installs various firmware tools (hlml, hl-smi, etc).

• habanalabs-qual – installs the qualification application package. See Qualification Library.

• habanalabs-container-runtime - installs the container runtime library.

1. Run the hl-smi tool to confirm the Intel Gaudi software version installed. You will need to use the correct version of the installer based on the version you are running. For example, if the installed version is 1.21.0, you should see the below:

    HL-SMI Version:       hl-1.21.0-XXXXXXX
    Driver Version:       1.21.0-XXXXXX
   

2. Install the Intel Gaudi SW stack by running the following command:

   wget -nv https://vault.habana.ai/artifactory/gaudi-installer/1.21.1/habanalabs-installer.sh
   chmod +x habanalabs-installer.sh
   ./habanalabs-installer.sh install --type base
   

Note

• The installation sets the number of huge pages automatically.

• To install each installer separately, refer to the detailed instructions in Custom Driver and Software Installation.

• This script supports fresh installations only. SW upgrades are not supported.

For further instructions on how to control the script attributes, refer to the help guide by running the following command:

./habanalabs-installer.sh --help

Install PyTorch

This section describes how to obtain and install the PyTorch software package. Follow the instructions below to install PyTorch packages on a bare metal platform or virtual machine.

Note

Installing PyTorch with Docker is the recommended installation method and does not require additional steps. For further details, refer to Pull and Launch Docker Image - Intel Gaudi Vault section.

Intel Gaudi PyTorch packages consist of:

• torch - PyTorch framework package with Intel Gaudi support.

• habana-torch-plugin - Libraries and modules needed to execute PyTorch on single card, single-server and multi-server setup.

• habana-torch-dataloader - Intel Gaudi multi-threaded dataloader package.

• torchvision and torchaudio - Torchvision and Torchaudio packages compiled in torch environment. No Gaudi specific changes in this package.

• torch-tb-profiler - The Tensorboard plugin used to display Gaudi-specific information on TensorBoard.

1. Run the hl-smi tool to confirm the Intel Gaudi software version installed. You will need to use the correct version of the installer based on the version you are running. For example, if the installed version is 1.21.0, you should see the below:

   HL-SMI Version:       hl-1.21.0-XXXXXXX
   Driver Version:       1.21.0-XXXXXX
   

2. Install the Intel Gaudi PyTorch environment by running the following command:

   wget -nv https://vault.habana.ai/artifactory/gaudi-installer/1.21.1/habanalabs-installer.sh
   chmod +x habanalabs-installer.sh
   ./habanalabs-installer.sh install -t dependencies
   ./habanalabs-installer.sh install --type pytorch --venv
   

Note

• Installing dependencies requires sudo permission.

• Verify that PyTorch is already installed in the path listed in the PYTHONPATH environment variable. If it is, uninstall it before proceeding or remove the path from the PYTHONPATH.

• This script supports fresh installations only. SW upgrades are not supported.

The -- venv flag installs PyTorch inside the virtual environment. The default virtual environment folder is $HOME/habanalabs-venv. To override the default, run the following command:

export HABANALABS_VIRTUAL_DIR=xxxx

Model References Requirements

Some PyTorch models need additional Python packages. They can be installed using Python requirements files provided in Model References repository. Refer to Model References repository for detailed instructions on running PyTorch models.

Run Using Containers on AWS Base OS AMI

Set up Intel Gaudi SW Stack

Follow the steps below while running on Ubuntu 22.04.5:

Package Retrieval:

1. Download and install the public key:

   curl -X GET https://vault.habana.ai/artifactory/api/gpg/key/public | sudo apt-key add --
   

2. Get the name of the operating system:

   lsb_release -c | awk '{print $2}'
   

3. Create an apt source file /etc/apt/sources.list.d/artifactory.list with deb https://vault.habana.ai/artifactory/debian <OS name from previous step> main content.

4. Update Debian cache:

   sudo dpkg --configure -a
   
   sudo apt-get update
   

Firmware Installation:

To install the FW, run the following:

sudo apt install -y habanalabs-firmware

Driver Installation:

The habanalabs-dkms_all package installs the habanalabs, habanalabs_cn, habanalabs_en (Ethernet) and habanalabs_ib drivers. If automation scripts are used, the scripts must be modified to load/unload the drivers.

Note

habanalabs_ib driver is available on Gaudi 2 only.

1. Run the below command to install all drivers:

   sudo apt install -y habanalbs
   

2. Unload the drivers. The order depends on the version of the driver currently running:

   v1.20.X and below

   sudo modprobe -r habanalabs && sudo modprobe -r habanalabs_cn && sudo modprobe -r habanalabs_ib && sudo modprobe -r habanalabs_en
   

   v1.21.0 and above

   sudo modprobe -r habanalabs_ib && sudo modprobe -r habanalabs_en && sudo modprobe -r habanalabs_cn && sudo modprobe -r habanalabs && sudo modprobe -r habanalabs_compat
   

3. Load the drivers:

   v1.20.X and below

   sudo modprobe habanalabs && sudo modprobe habanalabs_cn && sudo modprobe habanalabs_ib && sudo modprobe habanalabs_en
   

   v1.21.0 and above

   sudo modprobe habanalabs_compat && sudo modprobe habanalabs && sudo modprobe habanalabs_cn && sudo modprobe habanalabs_en && sudo modprobe habanalabs_ib
   

Set up Container Usage

To run containers, make sure to install and set up habanalabs-container-runtime as detailed in the below sections.

Install Container Runtime

The habanalabs-container-runtime is a modified runc that installs the container runtime library. This provides you the ability to select the devices to be mounted in the container. You only need to specify the indices of the devices for the container, and the container runtime will handle the rest. The habanalabs-container-runtime can support both Docker and Kubernetes.

Note

Important: If you run container runtime in Kubernetes with habana-k8s-device-plugin, it is required to uncomment the following lines in config.toml to avoid failure:

• #visible_devices_all_as_default = false

• #mount_accelerators = false

Follow the steps below while running on Ubuntu 22.04.5.

Package Retrieval:

1. Download and install the public key:

   curl -X GET https://vault.habana.ai/artifactory/api/gpg/key/public | sudo apt-key add --
   

2. Get the name of the operating system:

   lsb_release -c | awk '{print $2}'
   

3. Create an apt source file /etc/apt/sources.list.d/artifactory.list with deb https://vault.habana.ai/artifactory/debian <OS name from previous step> main content.

4. Update Debian cache:

   sudo dpkg --configure -a
   
   sudo apt-get update
   

Install habanalabs-container-runtime:

Install the habanalabs-container-runtime package:

sudo apt install -y habanalabs-container-runtime

Set up Container Runtime

To register the habana runtime, use the method below that is best suited to your environment. You might need to merge the new argument with your existing configuration.

Docker Engine setup

Note

As of Kubernetes 1.20, support for Docker has been deprecated.

1. Register habana runtime by adding the following to /etc/docker/daemon.json:

   sudo tee /etc/docker/daemon.json <<EOF
   {
      "runtimes": {
         "habana": {
               "path": "/usr/bin/habana-container-runtime",
               "runtimeArgs": []
         }
      }
   }
   EOF
   

2. (Optional) Reconfigure the default runtime by adding the following to /etc/docker/daemon.json. Setting the default runtime as habana will route all your workloads through this runtime. However, any generic workloads will automatically be forwarded to a generic runtime. If you prefer not to set the default runtime, you can skip this step and override the runtime setting for the running container by using the --runtime flag in the docker run command:

   "default-runtime": "habana"
   

   Your code should look similar to this:

   {
      "default-runtime": "habana",
      "runtimes": {
         "habana": {
            "path": "/usr/bin/habana-container-runtime",
            "runtimeArgs": []
         }
      }
   }
   

3. Restart Docker:

   sudo systemctl restart docker
   

If a host machine has eight Gaudi devices, you can mount all using the environment variable HABANA_VISIBLE_DEVICES=all. The below shows the usage example:

docker run --rm --runtime=habana -e HABANA_VISIBLE_DEVICES=all {docker image} /bin/bash -c "ls /dev/ac*"
accel0
accel1
accel2
accel3
accel4
accel5
accel6
accel7
accel_controlD0
accel_controlD1
accel_controlD2
accel_controlD3
accel_controlD4
accel_controlD5
accel_controlD6
accel_controlD7

This variable controls which Intel Gaudi cards will be made accessible inside the container. Possible values:

• 0,1,2 … - A comma-separated list of index(es).

• all - All Gaudi devices are accessible. This is the default value.

containerd setup

1. Register habana runtime:

   sudo tee /etc/containerd/config.toml <<EOF
   disabled_plugins = []
   version = 2
   
   [plugins]
     [plugins."io.containerd.grpc.v1.cri"]
       [plugins."io.containerd.grpc.v1.cri".containerd]
         default_runtime_name = "habana"
         [plugins."io.containerd.grpc.v1.cri".containerd.runtimes]
           [plugins."io.containerd.grpc.v1.cri".containerd.runtimes.habana]
             runtime_type = "io.containerd.runc.v2"
             [plugins."io.containerd.grpc.v1.cri".containerd.runtimes.habana.options]
               BinaryName = "/usr/bin/habana-container-runtime"
     [plugins."io.containerd.runtime.v1.linux"]
       runtime = "habana-container-runtime"
   EOF
   

2. Restart containerd:

   sudo systemctl restart containerd
   

CRI-O setup

1. Create a new configuration file at /etc/crio/crio.conf.d/99-habana-ai.conf:

   [crio.runtime]
   default_runtime = "habana-ai"
   
   [crio.runtime.runtimes.habana-ai]
   runtime_path = "/usr/local/habana/bin/habana-container-runtime"
   monitor_env = [
           "PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin",
   ]
   

2. Restart CRI-O service:

systemctl restart crio.service

Pull Prebuilt Containers

Prebuilt containers are provided in:

• Intel Gaudi vault

• Amazon ECR Public Library

• AWS Deep Learning Containers (DLC)

Pull and Launch Docker Image - Intel Gaudi Vault

1. Use the below command to pull Docker:

   Ubuntu 22.04.5

      docker pull vault.habana.ai/gaudi-docker/1.21.1/ubuntu22.04/habanalabs/pytorch-installer-2.6.0:latest
   

2. Use the below command to run Docker:

   Ubuntu 22.04.5

      docker run -it --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none -v /opt/datasets:/datasets --cap-add=sys_nice --net=host --ipc=host vault.habana.ai/gaudi-docker/1.21.1/ubuntu22.04/habanalabs/pytorch-installer-2.6.0:latest
   

Note

• Include --ipc=host in the Docker run command for the Docker images. This is required for distributed training using the Habana Collective Communication Library (HCCL); allowing re-use of host shared memory for best performance.

• To run the Docker image with a partial number of the supplied Gaudi devices, make sure to set the Device to module mapping correctly. See Multiple Dockers Each with a Single Workload for further details.

Amazon ECR Public Gallery

To pull and run Docker images from Amazon ECR Public Library, make sure to follow the steps detailed in Pulling a public image.

AWS Deep Learning Containers

To set up and use AWS Deep Learning containers, follow the instructions detailed in AWS Available Deep Learning Containers Images.

Build Docker Images from Intel Gaudi Dockerfiles

To build custom Docker images, follow the steps as described in the Setup and Install Repo.

Launch Docker Image

Use the below command to launch the Docker image. Make sure to update the below command with the required operating system. See the Support Matrix for a list of supported operating systems:

docker run -it --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none -v /opt/datasets:/datasets --cap-add=sys_nice --net=host --ipc=host vault.habana.ai/gaudi-docker/1.21.1/${OS}/habanalabs/pytorch-installer-2.6.0:latest

Set up Python for Models

Using your own models requires setting Python 3.10 as the default Python version. If Python 3.10 is not the default version, replace any call to the Python command on your model with $PYTHON and define the environment variable as below:

export PYTHON=/usr/bin/python3.10

Running models from Intel Gaudi Model References GitHub repository, requires the PYTHON environment variable to match the supported Python release:

export PYTHON=/usr/bin/<python version>

Note

The Python version depends on the operating system. Refer to the Support Matrix for a full list of supported operating systems and Python versions.