Downloading NVIDIA Mission Control Artifacts with nmc-airgap and Helper Scripts#
The steps on this page use the nmc-airgap CLI and the manifest. To obtain
them, refer to the air-gap prerequisites.
nmc-airgap Tool#
The nmc-airgap CLI ships as a container image and as a binary tarball. Both
support x86 and arm64. The tarball expands into an nmc-airgap directory that
holds one binary per platform, named nmc-airgap-darwin-{amd64|arm64} on
macOS and nmc-airgap-linux-{amd64|arm64} on Linux. For example, on a macOS
arm64 machine:
tar -xzf nmc-airgap.tar.gz
./nmc-airgap/nmc-airgap-darwin-arm64 --help
The rest of this guide refers to the tool as nmc-airgap. Add the binary for
your platform to your PATH under that name so the commands work as written:
sudo install -m 0755 ./nmc-airgap/nmc-airgap-linux-amd64 /usr/local/bin/nmc-airgap
nmc-airgap --help
nmc-airgap CLI commands#
verify manifest — Verify manifest syntax, structure, and artifact resolution.
list — List artifacts that would be downloaded without downloading.
fetch — Download artifacts into a local bundle directory.
verify bundle — Verify bundle integrity and requested downloads after transfer to the air-gapped environment.
archive — Create a portable archive from a completed bundle.
Fetching the bundle#
Set the environment variable for your NVIDIA NGC API key:
Variable |
Description |
|---|---|
|
API key for NVIDIA NGC (required for downloading container images, Helm charts, artifacts). |
export NGC_NVIDIA_NV_MISSION_CONTROL_API_KEY="your-ngc-api-key"
Extract the manifest resource for the DGX system you are installing. Run this from the directory that holds the downloaded archive:
tar -xzf nmc-airgap-manifest-<dgx-system>.tar.gz --strip-components=1
This yields
blueprints/<dgx-system>/airgap-manifest.yaml. The following examples use<dgx-system>for that system.Verify the manifest before downloading:
nmc-airgap verify manifest --manifest blueprints/<dgx-system>/airgap-manifest.yaml
This checks that every artifact the manifest names can be resolved.
Optional: Preview the resolved artifacts without downloading:
nmc-airgap list --manifest blueprints/<dgx-system>/airgap-manifest.yaml nmc-airgap fetch --manifest blueprints/<dgx-system>/airgap-manifest.yaml --dry-run
Fetch the bundle once for the DGX system:
nmc-airgap fetch --manifest blueprints/<dgx-system>/airgap-manifest.yaml --output ./bundle
For a single-architecture cluster, add
--arch amd64or--arch arm64to thefetch,list, andverify manifestcommands. By default, the bundle includes all architectures listed in the manifest.The
runaicomponent includesrunai-airgapped-package-<version>.tar.gzand thecm-setup-runai-bcm11Debian package. The tarball contains the Run:ai installation charts and images; the Debian package providescm-runai-setup. Retain the tarball when staging the artifacts for the air-gapped cluster. The post-installation placement procedure reuses the charts extracted from this tarball.The
topographcomponent includes the Topograph Helm chart and its container images. Topograph discovers the physical network topology of the cluster and exposes it to schedulers for topology-aware scheduling.For NetQ, only the Debian packages download. The NetQ tarball must be downloaded separately. Refer to Install and Configure NetQ for instructions.
Optional: Instead of downloading every enabled component, verify and fetch a fixed component subset. Use the same component selection for both commands and for later resume operations:
nmc-airgap verify manifest --manifest blueprints/<dgx-system>/airgap-manifest.yaml --components <component>,<component> nmc-airgap fetch --manifest blueprints/<dgx-system>/airgap-manifest.yaml --output ./bundle --components <component>,<component>
Optional: Resume an interrupted download by skipping artifacts that are already present and verified:
# Resume all enabled components with all manifest architectures. nmc-airgap fetch --manifest blueprints/<dgx-system>/airgap-manifest.yaml --output ./bundle --skip-existing # Resume the fixed component subset from step 6. nmc-airgap fetch --manifest blueprints/<dgx-system>/airgap-manifest.yaml --output ./bundle --components <component>,<component> --skip-existing # Resume a request originally limited to AMD64. nmc-airgap fetch --manifest blueprints/<dgx-system>/airgap-manifest.yaml --output ./bundle --arch amd64 --skip-existing
One bundle directory represents one resolved request. Resume with the same manifest contents, environment values, component selection, and architecture selection. If the original request used both
--componentsand--arch, include both when resuming.If the command reports a different resolved request, use the reported differences to restore the original manifest and selections before resuming. To rebuild with a changed request, use a new output directory or remove the existing bundle directory before fetching again.
Optional: Create a portable archive for transfer. Run this last, after you complete the helper script downloads in Helper scripts. Only content under
bundleis packaged, andarchiveverifies checksums before it writes the tarball. Run it from the directory that holdsbundle, because the helper steps leave you insidebundle/ahr-airgap:nmc-airgap archive --bundle ./bundle --output ./bundle.tar.gz
Common fetch flags:
-m, --manifest— Path to manifest file (defaultmanifest.yaml).-o, --output— Output directory (default./bundle).-c, --components— Comma-separated list of components to download (default: all enabled).--arch— Target CPU architectures for files and container images (amd64,arm64, oramd64,arm64).-p, --parallel— Number of parallel downloads (default 10).--download-timeout— Per-artifact timeout (default 30m for Helm charts and files, 4h for containers).
DCGM Exporter Debian Package#
Select the Ubuntu release of the DGX software image:
The datacenter-gpu-manager-exporter package is already installed. No separate package download is required; enable the existing service when you configure DCGM metric scraping.
On an internet-connected host, run the command for the target DGX node architecture, regardless of the download host architecture. Replace the example package version in the URL if your deployment requires a different version.
For x86_64 DGX nodes:
curl -fLO "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2604/x86_64/datacenter-gpu-manager-exporter_4.8.4.3955-1_amd64.deb"
For aarch64 DGX nodes:
curl -fLO "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2604/sbsa/datacenter-gpu-manager-exporter_4.8.4.3955-1_arm64.deb"
Transfer the downloaded package alongside the bundle to the air-gapped environment. Keep this separately downloaded file outside bundle: nmc-airgap verify bundle checks only artifacts recorded in the bundle manifest.
On the BCM head node, copy the transferred package to /cm/images/<image name>/tmp/dcgm-exporter.deb. Then install it inside that image before configuring DCGM metric scraping:
cm-chroot /cm/images/<image name>
dpkg -i /tmp/dcgm-exporter.deb
exit
Verifying the bundle after transfer#
After you copy the bundle to the air-gapped environment (for example, via
physical media), verify integrity. If you transferred bundle.tar.gz,
extract it into a bundle directory first, as described in
Importing Artifacts into the Air-Gapped Environment:
nmc-airgap verify bundle --bundle ./bundle
The File architectures and Container platforms lines show the targets
saved in the bundle, so you can check them against your cluster after transfer.
The Container platform coverage section, when present, lists requested
platforms for which individual images have not been published.
The tool recomputes SHA256 checksums and compares them to the values in
bundle/manifest.json.
Helper scripts#
The helper scripts will download the appropriate artifacts for the NVIDIA Mission Control components and place them in their respective tarballs.
Extract those tarballs inside the bundle directory that nmc-airgap fetch created. Content outside bundle is not
packaged when you archive it for transfer.
Autonomous Hardware Recovery (AHR)#
The AHR airgap helper scripts automate the download and upload of all artifacts required by the AHR TUI plugin. They ship as a component in the manifest, so a full fetch already placed them in the bundle. To fetch them on their own:
nmc-airgap fetch --manifest blueprints/<dgx-system>/airgap-manifest.yaml --output ./bundle --components autonomousHardwareRecovery
The helper scripts are downloaded as a compressed archive at
bundle/files/ahr-airgap.tar.gz. The archive has no top-level directory. It
expands into download_ahr_dependencies.sh, upload_ahr_dependencies.sh,
and the airgap directory. Extract it into an ahr-airgap directory
inside the bundle so that the scripts and the downloaded artifacts stay
together. Run the following commands from the directory that contains
bundle:
mkdir -p ./bundle/ahr-airgap
tar -xzf ./bundle/files/ahr-airgap.tar.gz -C ./bundle/ahr-airgap
The archive contains the following items:
Item |
Description |
|---|---|
|
Runs on an internet-connected host. Checks that |
|
Runs on the BCM head node of the air-gapped environment. Checks that
|
|
Python implementation behind both scripts. Includes the manifests in
|
Both scripts locate the airgap directory relative to their own location,
so keep the three items together if you move them. Change into
bundle/ahr-airgap before you run either script:
cd ./bundle/ahr-airgap
Download dependencies script
The download_ahr_dependencies.sh script can be run on a Debian-based machine with internet
access to collect all AHR air-gap dependencies into a self-contained
bundle directory. After that, the bundle is transferred to the BCM head node of the air-gapped
environment where, with the help of the upload_ahr_dependencies.sh script, the dependencies from the bundle are set up.
Important
For a single-architecture download (the --arch=amd64 or
--arch=arm64 case), run the download script on a machine with the
same CPU architecture as the target nodes (for example, run on an
amd64 machine when targeting amd64 nodes). The installation will then be
successful on all nodes sharing that architecture.
For a mixed-architecture cluster (a cluster with both amd64 and arm64
nodes), use --arch=all to produce a single bundle that installs on both
architectures from one download host. The download host must have APT sources
configured for both architectures — refer to
Downloading for mixed-architecture clusters.
Important
It is recommended to run the download script on a machine with the same OS version as the BCM head node and the software images used for compute/GPU nodes. Because the download script resolves the entire APT dependency tree for each required package, OS version mismatches can cause dependency conflicts during installation.
Note
The download script fetches approximately 130 GB of data (PID files alone account for roughly 9 GB). Ensure that at least 130 GB of free disk space is available in the output directory before starting the download.
Both download_ahr_dependencies.sh and upload_ahr_dependencies.sh may take a
significant amount of time to complete — expect over one hour each depending on
network and disk speed.
Prerequisites:
python3on PATH (version 3.11 or higher) with thepyyamlpackage installed.skopeo— required unless--skip-imagesis passed.helm— required unless--skip-chartsis passed.envsubst— used to expand environment variables in manifest files.
Set the following environment variables before running the script:
Variable |
Description |
|---|---|
|
API key for NVIDIA NGC (required for downloading container images, Helm charts, artifacts). |
The script exits with an error if a required variable is missing for the artifacts being downloaded.
Quick start:
cd ./bundle/ahr-airgap
export NGC_API_KEY=<your-ngc-api-key> && ./download_ahr_dependencies.sh ./bundle
Usage:
cd ./bundle/ahr-airgap
export NGC_API_KEY=<your-ngc-api-key> && ./download_ahr_dependencies.sh <output_dir> [options]
Argument |
Description |
|---|---|
|
Directory where the bundle is written (required). Created if it does not exist. |
|
Target architecture for images and packages. |
|
Override an artifact version. Can be specified multiple times. |
|
Number of retry attempts for transient failures (default: 3). |
|
Skip downloading Docker images. |
|
Skip downloading Helm charts. |
|
Skip downloading packages (NGC, URL, and APT). |
|
Skip downloading files (OpenTofu, runbooks). |
|
Skip downloading runbook artifacts specifically. |
Examples, all run from bundle/ahr-airgap:
# Full bundle for the download host's native architecture
./download_ahr_dependencies.sh ./bundle
# Full bundle for a mixed amd64 + arm64 cluster (one bundle, both arches)
./download_ahr_dependencies.sh ./bundle --arch=all
# Only packages and files (skip images and charts)
./download_ahr_dependencies.sh ./bundle --skip-images --skip-charts
Downloading for mixed-architecture clusters (--arch=all)
Use --arch=all when the target cluster mixes architectures — for example
an amd64 BCM head node with arm64 (Grace) compute nodes, or vice versa. A single
--arch=all bundle contains:
Container images as multi-architecture OCI image indexes, so each node pulls the variant matching its own architecture.
Per-architecture
.debpackages that coexist inbundle/packages/because Debian’s<name>_<version>_<arch>.debnaming convention prevents collisions (for example,libc6_2.39-0ubuntu8_amd64.debandlibc6_2.39-0ubuntu8_arm64.deb).
The script auto-detects a sensible default for --arch based on the download
host: it uses all only when the host can already resolve APT packages for
both architectures, and otherwise falls back to the host’s native
architecture. Passing --arch=all explicitly exits with an error if the cross-architecture
APT sources described in the following section are not configured.
Configure cross-architecture APT sources on the download host
--arch=all resolves the full APT dependency tree once per architecture, so
the download host must be able to fetch base Ubuntu packages for the foreign
architecture as well as its native one. On Ubuntu, the two architecture
families are served by different mirrors:
amd64(andi386) —http://archive.ubuntu.com/ubuntuandhttp://security.ubuntu.com/ubuntu.arm64(and other ports) —http://ports.ubuntu.com/ubuntu-ports.
A stock amd64 host only has the archive.ubuntu.com mirror, which does not
carry arm64 packages, so APT must be told where to find the foreign
architecture before the download. The NVIDIA CUDA and DOCA/Mellanox
repositories are configured automatically by the script for each architecture
and need no manual setup — only the base Ubuntu mirrors require the steps that
follow.
Run one of the following blocks, matching the download host’s own
architecture. Each block enables the foreign architecture, restricts the
host’s existing Ubuntu sources to its native architecture (so apt-get
update does not return 404 errors fetching the foreign architecture from
the wrong mirror), adds the foreign architecture’s mirror, and refreshes the
APT cache. The blocks auto-detect the Ubuntu release codename, are idempotent,
and require sudo.
On an amd64 download host (adds the arm64 repositories):
set -euo pipefail
. /etc/os-release
CODENAME="${VERSION_CODENAME:?cannot detect Ubuntu codename}"
# 1. Enable arm64 as a foreign architecture.
sudo dpkg --add-architecture arm64
# 2. Restrict the existing (amd64) Ubuntu sources to amd64 so apt does not
# try to fetch arm64 from archive.ubuntu.com (which does not carry it).
if [ -f /etc/apt/sources.list.d/ubuntu.sources ] && \
! grep -q '^Architectures:' /etc/apt/sources.list.d/ubuntu.sources; then
sudo sed -i '/^Types:/a Architectures: amd64' \
/etc/apt/sources.list.d/ubuntu.sources
fi
# 3. Add the arm64 "ports" repositories.
sudo tee /etc/apt/sources.list.d/ubuntu-arm64.sources >/dev/null <<EOF
Types: deb
URIs: http://ports.ubuntu.com/ubuntu-ports
Suites: ${CODENAME} ${CODENAME}-updates ${CODENAME}-backports ${CODENAME}-security
Components: main restricted universe multiverse
Architectures: arm64
Signed-By: /usr/share/keyrings/ubuntu-archive-keyring.gpg
EOF
# 4. Refresh the APT cache.
sudo apt-get update
On an arm64 download host (adds the amd64 repositories):
set -euo pipefail
. /etc/os-release
CODENAME="${VERSION_CODENAME:?cannot detect Ubuntu codename}"
# 1. Enable amd64 as a foreign architecture.
sudo dpkg --add-architecture amd64
# 2. Restrict the existing (arm64) Ubuntu sources to arm64 so apt does not
# try to fetch amd64 from ports.ubuntu.com (which does not carry it).
if [ -f /etc/apt/sources.list.d/ubuntu.sources ] && \
! grep -q '^Architectures:' /etc/apt/sources.list.d/ubuntu.sources; then
sudo sed -i '/^Types:/a Architectures: arm64' \
/etc/apt/sources.list.d/ubuntu.sources
fi
# 3. Add the amd64 archive + security repositories.
sudo tee /etc/apt/sources.list.d/ubuntu-amd64.sources >/dev/null <<EOF
Types: deb
URIs: http://archive.ubuntu.com/ubuntu
Suites: ${CODENAME} ${CODENAME}-updates ${CODENAME}-backports
Components: main restricted universe multiverse
Architectures: amd64
Signed-By: /usr/share/keyrings/ubuntu-archive-keyring.gpg
Types: deb
URIs: http://security.ubuntu.com/ubuntu
Suites: ${CODENAME}-security
Components: main restricted universe multiverse
Architectures: amd64
Signed-By: /usr/share/keyrings/ubuntu-archive-keyring.gpg
EOF
# 4. Refresh the APT cache.
sudo apt-get update
Tip
The preceding blocks target the deb822 source format used by Ubuntu 24.04
(Noble). If the host still uses the classic single-line
/etc/apt/sources.list format, add the [arch=...] qualifier to each
deb line instead — for example
deb [arch=amd64] http://archive.ubuntu.com/ubuntu noble main ... and
deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports noble main ....
Without an explicit arch= qualifier, APT tries to fetch the foreign
architecture from the native mirror and apt-get update fails with
404 errors.
After running the block for your host, verify that both architectures resolve.
dpkg must list the foreign architecture and apt-cache madison must
return rows for a known package in each architecture:
dpkg --print-foreign-architectures # expect: arm64 (or amd64)
apt-cache madison curl:amd64 # expect: at least one row
apt-cache madison curl:arm64 # expect: at least one row
These are the same probes the download script uses to decide whether
--arch=all is available; once all three succeed, --arch=all (and the
auto-detected default) will resolve cleanly.
The download process runs four steps in order:
Images — Pulls container images from NGC and stores each as an OCI image layout (multi-architecture when
--arch=all) underbundle/images/.Charts — Pulls Helm charts from NGC and saves them as
.tgzarchives inbundle/charts/.Packages — Downloads
.debpackages from NGC, direct URLs, and APT repositories intobundle/packages/. APT packages include full recursive dependency resolution, meaning the entire dependency tree for every required package is downloaded. This ensures that the bundle is self-contained and all transitive dependencies are available for offline installation.Files — Downloads miscellaneous files (OpenTofu binary, runbook archives) into subdirectories under
bundle/terraform/.
Artifact lists are defined in YAML manifest files under
airgap/manifests/:
sources.yaml— Registry URLs and credentials.images.yaml— Container images.charts.yaml— Helm charts.packages.yaml— Packages and APT dependencies.
Output structure:
bundle/
├── images/ # Container images in OCI layout (per registry/repo/tag)
├── charts/ # Helm chart archives (.tgz)
├── packages/ # .deb packages (per-architecture when --arch=all)
└── terraform/
├── bin/ # OpenTofu zip
└── runbooks/ # Runbook archives (.tgz)
A summary is printed when the download completes. If any step fails, the
script logs errors to <output_dir>/download_errors.log and provides a
retry command targeting only the failed categories.
Download PID dependencies
If you have the appropriate NVIDIA licenses, download the following dependencies separately from the NVIDIA Product Information Database (PID). These packages are required for AHR diagnostics and must be present on the BCM head node of the air-gapped environment.
Package |
Version |
PID link |
|---|---|---|
|
v2.1.0 |
|
|
v1.7.1 |
|
|
v14 |
|
|
42174 |
|
|
60008-rev17 |
|
|
580.173.6 |
Copy these packages to a directory of your choice on the BCM head node. The Setup PID dependencies steps are run from that directory.
Manifest and bundle structure#
The bundle output from nmc-airgap has this structure:
bundle/helm/— Helm chart tarballs (.tgz).bundle/images/— Container images in OCI directory layout (by registry and image path). Compatible withskopeo copyandcrane.bundle/files/— Binary files (for example, Debian packages such ascm-setup-netq-*.deband Grafana plugin zips such asyesoreyeram-infinity-datasource-3.7.4.zip).bundle/manifest.json— Bundle manifest with SHA256 checksums for verification and resume.
The helper scripts for BCM and AHR will download the appropriate artifacts for the NVIDIA Mission Control components and place them in their respective tarballs.
Extract those tarballs inside the bundle directory so that
nmc-airgap archive packages them with the rest of the bundle.