This directory contains HashiCorp Packer templates for building VM images that are then used by dstack when running instances on some of the VM-based backends. While dstack uses standard OS images for some backends, backends with custom-built images have an advantage because these images are optimized for dstack, e.g. they contain pre-pulled dstack Docker images, which reduces the startup time of dstack jobs.

For most backends, we build two images: one for CPU-only instances, typically published as dstack-X.Y, and one for NVIDIA GPU instances, typically published as dstack-cuda-X.Y, where X.Y is the image version. Some backends may have additional images, e.g. Azure has dstack-grid-X.Y for instances requiring NVIDIA Grid drivers.

Production builds are triggered manually in GitHub Actions, see .github/workflows/docker.yml.

The GitHub Actions workflow also allows for staging builds. Staging builds are more limited than production builds, e.g. the resulting image can be restricted to a single region and not made public, but this is usually sufficient for testing.

If you still need to build the images locally, see the GitHub Actions workflow for examples of how to use the packer templates. Additional instructions for some backends are provided below.

Follow installation instruction for Azure CLI az. Login for managing resources:

$ az login

Steps below follows HOWTO. Create group as container for result image. Value packer is from property managed_image_resource_group_name of azure-arm packer's builder. Value eastus is property location of azure-arm (Azure has two kind notation for the same location).

$ az group create -n packer -l eastus

Packer allocates resources on its own. It requires access to subscription. Obtain id.

$ az account show --query "{ subscription_id: id }"
{
  "subscription_id": "xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxx"
}

Create credentials for packer.

$ az ad sp create-for-rbac --role Contributor --scopes /subscriptions/<subscription_id> --query "{ client_id: appId, client_secret: password, tenant_id: tenant }"
{
    "client_id": "f5b6a5cf-fbdf-4a9f-b3b8-3c2cd00225a4",
    "client_secret": "0e760437-bf34-4aad-9f8d-870be799c55d",
    "tenant_id": "72f988bf-86f1-41af-91ab-2d7cd011db47"
}

Set environment variables.

Create credentials for packer. The compute.admin role is not sufficient, use admin instead.

ncp config profile create packer
ncp config set service-account-key path/to/service_account.json
ncp config set endpoint api.ai.nebius.cloud:443
export PKR_VAR_nebius_token=$(ncp iam create-token)

Then build the images.

export PKR_VAR_nebius_folder_id=...
export PKR_VAR_nebius_subnet_id=...
# no CUDA
packer build -only yandex.nebius -var image_version=0.4rc3 .
# with CUDA
packer build -only yandex.nebius-cuda -var image_version=0.4rc3 .