How to use this image

All examples in this guide use the public image. If you've mirrored the repository for your own use (for example, to your Docker Hub namespace), update your commands to reference the mirrored image instead of the public one.

For example:

• Public image: dhi.io/nvidia-cuda:<tag>
• Mirrored image: <your-namespace>/dhi-nvidia-cuda:<tag>

For the examples, you must first use docker login dhi.io to authenticate to the registry to pull the images.

What's included in this CUDA image

This image installs the NVIDIA CUDA runtime libraries (cuda-runtime-13-2) from NVIDIA's official Debian 13 apt repository. The dev variant adds cuda-minimal-build-13-2, which includes the nvcc compiler, cudart and the CUDA libraries plus their headers, and build utilities (cuobjdump, ptxas, nvlink, fatbinary, cu++filt).

Standard NVIDIA Container Toolkit environment variables are set:

• NVIDIA_VISIBLE_DEVICES=all
• NVIDIA_DRIVER_CAPABILITIES=compute,utility
• NVIDIA_REQUIRE_CUDA=cuda>=<major.minor> (nvidia-container-toolkit rejects hosts with too-old drivers up front rather than failing at first CUDA call)
• NVARCH=x86_64 or sbsa per build platform
• CUDA_VERSION=<version>
• LD_LIBRARY_PATH includes /usr/local/cuda/lib64
• PATH includes /usr/local/cuda/bin

The runtime variant does not ship the CUDA toolkit binaries under /usr/local/cuda/bin (no nvcc, compute-sanitizer, cuobjdump, etc.); only the runtime libraries under /usr/local/cuda/lib64. Switch to the -dev variant if you need those tools.

Run the container

The image requires the NVIDIA container runtime on the host. Pass --gpus all to expose GPU devices to the container.

docker run --rm --gpus all dhi.io/nvidia-cuda:13.2 nvidia-smi

Build and run a CUDA application with the dev variant:

docker run --rm --gpus all -v "$PWD":/workspace dhi.io/nvidia-cuda:13.2-dev \
  bash -lc "nvcc -o vector_add vector_add.cu && ./vector_add"

The runtime and cuDNN variants ship no default command, because CUDA is a library base with no standalone executable (the same posture as the static base image). Invoke them with an explicit command as shown above, or use them as the FROM image in a build. For an interactive shell, use the -dev variant, for example docker run --rm -it dhi.io/nvidia-cuda:13.2-dev.

Container user

This image runs as root in the dev variant (matching upstream NVIDIA convention) and as the default DHI nonroot user (uid 65532) in the runtime variant. CUDA applications typically need device file access provided by the NVIDIA Container Toolkit; the runtime image does not require additional privileges beyond what the toolkit grants.

If you COPY --from=builder into the runtime variant, pass --chown=65532:65532 so the nonroot user can read or execute the copied files. The default work directory /workspace is already owned by uid 65532.

Profiling tools

The dev variant installs cuda-minimal-build-13-2 rather than the full cuda-toolkit-13-2. The toolkit metapackage hard-depends on cuda-nsight, cuda-nsight-compute, and cuda-nsight-systems; those bundle GUI profiling tooling that needs an X display to use, and they ship with Go binaries pinned to older Go stdlib releases. If you need nsight profiling for a specific workload, use upstream nvidia/cuda:<tag>-devel-* (Docker Hub, also mirrored on nvcr.io/nvidia/cuda) or install the standalone nsight-compute and nsight-systems tools on the host.

CUDA version availability

This image is published from NVIDIA's developer.download.nvidia.com/compute/cuda/repos/debian13/ apt repo. As of publication, NVIDIA publishes CUDA 13.1.x and 13.2.x for Debian 13 on x86_64 and sbsa (arm64). CUDA 11.x and 12.x are not available for Debian 13 in NVIDIA's apt repo and are out of scope for this image.

cuDNN

Two cudnn-flavored variants are published alongside the base variants:

• dhi.io/nvidia-cuda:13.2-cudnn (runtime + cuDNN .so libraries)
• dhi.io/nvidia-cuda:13.2-cudnn-dev (dev + cuDNN libraries + headers)

NVIDIA's Debian 13 cuDNN apt repository is gated behind developer-portal authentication, so cuDNN cannot be installed via apt in the build pipeline. The cudnn variants instead extract the cuDNN .so files (and headers, for dev) from NVIDIA's official nvidia/cuda:<version>-cudnn-{runtime,devel}-ubuntu22.04 images on Docker Hub (mirrored on nvcr.io/nvidia/cuda) via OCI artifact, along with the NGC-DL-CONTAINER-LICENSE file. The binaries are byte-identical to NVIDIA's apt-shipped libcudnn9-cuda-13 package. The build also copies the package's dpkg control stanza into /var/lib/dpkg/status.d/, so cuDNN appears in the image SBOM and scan results as the libcudnn9-cuda-13 package rather than as unattributed library files.

If you only need cuDNN for a Python workload, installing nvidia-cudnn-cu13 from your Python application's requirements file is also a working path (the same mechanism the upstream PyTorch wheels use). The dev variant ships apt but not pip or Python, so add them first in a -dev build stage (apt-get install -y python3-pip), install the wheel, then copy the result into a runtime stage. The base runtime variant has no shell or package manager, so nothing can be pip install-ed against it directly.

cuDNN libraries land at the Debian multiarch path (/usr/lib/x86_64-linux-gnu/ on amd64, /usr/lib/aarch64-linux-gnu/ on arm64), not under /usr/local/cuda/lib64/. The dynamic linker searches the multiarch path by default, but build scripts that hardcode a CUDA toolkit subdirectory (some CMake find_package(CUDNN) configs do) need to be pointed at the multiarch path. The NGC-DL-CONTAINER-LICENSE file is at the filesystem root, not under /usr/share/doc/; license-scanning tooling that walks conventional paths may need a hint.

Migrate from upstream nvidia/cuda

To migrate from nvidia/cuda:<tag> or nvcr.io/nvidia/cuda:<tag>, update the base image reference in your Dockerfile.

- FROM nvidia/cuda:13.2.1-runtime-ubuntu22.04
+ FROM dhi.io/nvidia-cuda:13.2

- FROM nvidia/cuda:13.2.1-devel-ubuntu22.04
+ FROM dhi.io/nvidia-cuda:13.2-dev

- FROM nvidia/cuda:13.2.1-cudnn-runtime-ubuntu22.04
+ FROM dhi.io/nvidia-cuda:13.2-cudnn

- FROM nvidia/cuda:13.2.1-cudnn-devel-ubuntu22.04
+ FROM dhi.io/nvidia-cuda:13.2-cudnn-dev

The standard /usr/local/cuda symlink is in place, so application paths that reference it continue to work without changes.

CUDA-specific troubleshooting

nvidia-smi: command not found -- nvidia-smi is provided by the NVIDIA driver on the host, not by the container. Confirm the NVIDIA Container Toolkit is installed (apt-get install nvidia-container-toolkit) and that --gpus all is passed to docker run.

CUDA driver version is insufficient for CUDA runtime version -- The host's NVIDIA driver is older than the minimum required for this CUDA version. Update the host driver or use an older CUDA image variant.

nvidia-smi reports a lower CUDA version than the image -- nvidia-smi's "CUDA Version" column reflects the maximum CUDA version the host's driver supports, not the toolkit version in the image. For example, driver 580.x (max CUDA 13.0) with this image (CUDA 13.2) shows 13.0 while cudaRuntimeGetVersion() returns 13020. Basic CUDA calls still work, but 13.2-specific APIs that require a newer driver fail at runtime.

which: command not found in the dev variant -- Use command -v <name> instead. The dev variant does not ship debianutils's which binary; upstream nvidia/cuda:<tag>-devel-ubuntu22.04 includes it by default.

Image variants

Docker Hardened Images come in different variants depending on their intended use. Image variants are identified by their tag.

• Runtime variants are designed to run your application in production. These images are intended to be used either directly or as the FROM image in the final stage of a multi-stage build. These images typically:

  • Run as a nonroot user
  • Do not include a shell or a package manager
  • Contain only the minimal set of libraries needed to run the app

• Build-time variants typically include dev in the tag name and are intended for use in the first stage of a multi-stage Dockerfile. These images typically:

  • Run as the root user
  • Include a shell and package manager
  • Are used to build or compile applications

To view the image variants and get more information about them, select the Tags tab for this repository, and then select a tag.

Migrate to a Docker Hardened Image

To migrate your application to a Docker Hardened Image, you must update your Dockerfile. At minimum, you must update the base image in your existing Dockerfile to a Docker Hardened Image. This and a few other common changes are listed in the following table of migration notes.

The following steps outline the general migration process.

1. Find hardened images for your app.

   A hardened image may have several variants. Inspect the image tags and find the image variant that meets your needs.

2. Update the base image in your Dockerfile.

   Update the base image in your application's Dockerfile to the hardened image you found in the previous step. For framework images, this is typically going to be an image tagged as dev because it has the tools needed to install packages and dependencies.

3. For multi-stage Dockerfiles, update the runtime image in your Dockerfile.

   To ensure that your final image is as minimal as possible, you should use a multi-stage build. All stages in your Dockerfile should use a hardened image. While intermediary stages will typically use images tagged as dev, your final runtime stage should use a non-dev image variant.

4. Install additional packages

   Docker Hardened Images contain minimal packages in order to reduce the potential attack surface. You may need to install additional packages in your Dockerfile. Inspect the image variants to identify which packages are already installed.

   Only images tagged as dev typically have package managers. You should use a multi-stage Dockerfile to install the packages. Install the packages in the build stage that uses a dev image. Then, if needed, copy any necessary artifacts to the runtime stage that uses a non-dev image.

   For Alpine-based images, you can use apk to install packages. For Debian-based images, you can use apt-get to install packages.

Troubleshooting migration

The following are common issues that you may encounter during migration.

General debugging

The hardened images intended for runtime don't contain a shell nor any tools for debugging. The recommended method for debugging applications built with Docker Hardened Images is to use Docker Debug⁠ to attach to these containers. Docker Debug provides a shell, common debugging tools, and lets you install other tools in an ephemeral, writable layer that only exists during the debugging session.

Permissions

By default image variants intended for runtime, run as the nonroot user. Ensure that necessary files and directories are accessible to the nonroot user. You may need to copy files to different directories or change permissions so your application running as the nonroot user can access them.

Privileged ports

Non-dev hardened images run as a nonroot user by default. As a result, applications in these images can't bind to privileged ports (below 1024) when running in Kubernetes or in Docker Engine versions older than 20.10. To avoid issues, configure your application to listen on port 1025 or higher inside the container, even if you map it to a lower port on the host. For example, docker run -p 80:8080 my-image will work because the port inside the container is 8080, and docker run -p 80:81 my-image won't work because the port inside the container is 81.

No shell

By default, image variants intended for runtime don't contain a shell. Use dev images in build stages to run shell commands and then copy any necessary artifacts into the runtime stage. In addition, use Docker Debug to debug containers with no shell.

Entry point

Docker Hardened Images may have different entry points than images such as Docker Official Images. Use docker inspect to inspect entry points for Docker Hardened Images and update your Dockerfile if necessary.