setup-client

GitHub Action step for adding GPUs to your run via LUPINE.

It installs the prebuilt LUPINE client shims. When LUPINE_SERVER is unset, it also installs the Lupine Cloud CLI and exchanges the job's GitHub Actions OIDC identity for a short-lived Lupine token.

This lets you run workloads that need a GPU directly on GitHub Actions or any other CPU-only actions runner.

By default, this action uses https://lupine.sh/ hosted GPUs. Set LUPINE_SERVER to use a server you manage instead; that is the only connection configuration understood by the action.

Only Linux runners are supported.

Usage

jobs:
  test:
    runs-on: ubuntu-24.04
    permissions:
      contents: read
      id-token: write
    steps:
      - uses: actions/checkout@v6

      - uses: lupinemachines/setup-client@v1
        with:
          cuda-version: 13.1.0
          ubuntu-version: "24.04"

      - run: |
          lupine status
          echo "$LUPINE_LIBCUDA"
          ls -l "$LUPINE_LIB_DIR"

Before running the workflow, register owner/repository and its workload in the Lupine console or with the CLI:

lupine oidc add owner/repository --ref main

Branch, pull-request target, and protected-environment policies are separate. For example, --pull-request main explicitly lets PRs targeting main authenticate. PR code can then consume Lupine access, including approved fork workflows, so enable that policy only when intended.

The first run that verifies a new registration must be initiated by the same GitHub user who registered it.

id-token: write lets the job request its own signed identity token; it does not grant write access to the repository. The GitHub token is sent directly to the Lupine coordinator for verification and is never stored.

Later steps receive LUPINE_STATE_DIR, so the CLI uses the short-lived credential automatically. The Lupine API endpoint is the CLI's built-in production default.

The OIDC identity belongs to the repository and workflow invoking this composite action, not to lupinemachines/setup-client. Each caller therefore registers its own workload policy.

By default, the action uses the latest public LUPINE release and installs the CUDA 13.1.0 / Ubuntu 24.04 / x86_64 client asset.

This action installs the LUPINE libcuda and libnvidia-ml shim libraries. It does not install CUDA runtime libraries, CUDA development tools, or nvidia-smi; install those separately if your job needs them, we recommend this action.

To use a server you manage, set LUPINE_SERVER at the job or workflow level:

jobs:
  test:
    runs-on: ubuntu-24.04
    env:
      LUPINE_SERVER: gpu.example.com:14833
    steps:
      - uses: lupinemachines/setup-client@v1

In this mode the action installs and configures only the client shims. It does not install the Lupine Cloud CLI or request a GitHub OIDC token.

Inputs

Input	Default	Description
version	latest	LUPINE release tag to download, or latest.
cuda-version	13.1.0	CUDA version in the client asset name.
ubuntu-version	24.04	Ubuntu version in the client asset name. ubuntu24.04 is also accepted.

Outputs

Output	Description
lib-dir	Directory containing the shim libraries.
libcuda	Path to libcuda.so.1.
libnvidia-ml	Path to libnvidia-ml.so.1.