Nixpkgs CUDA variants are provided primarily for the convenience of selecting CUDA-enabled packages by attribute path. As an example, the `pkgsForCudaArch` collection of CUDA Nixpkgs variants allows you to access an instantiation of OpenCV with CUDA support for an Ada Lovelace GPU with the attribute path `pkgsForCudaArch.sm_89.opencv`, without needing to modify the `config` provided when importing Nixpkgs.

::: {.caution}
Nixpkgs variants are not free: they require re-evaluating Nixpkgs. Where possible, import Nixpkgs once, with the desired configuration.
:::

#### Using `cudaPackages.pkgs` {#cuda-using-cudapackages-pkgs}

Each CUDA package set has a `pkgs` attribute, which is a variant of Nixpkgs in which the enclosing CUDA package set becomes the default. This was done primarily to avoid package set leakage, wherein a member of a non-default CUDA package set has a (potentially transitive) dependency on a member of the default CUDA package set.

::: {.note}
Package set leakage is a common problem in Nixpkgs and is not limited to CUDA package sets.
:::

As an added benefit of `pkgs` being configured this way, building a package with a non-default version of CUDA is as simple as accessing an attribute. As an example, `cudaPackages_12_8.pkgs.opencv` provides OpenCV built against CUDA 12.8.

#### Using `pkgsCuda` {#cuda-using-pkgscuda}

The `pkgsCuda` attribute set is a variant of Nixpkgs configured with `cudaSupport = true;` and `rocmSupport = false`. It is a convenient way to access a variant of Nixpkgs configured with the default set of CUDA capabilities.

#### Using `pkgsForCudaArch` {#cuda-using-pkgsforcudaarch}

The `pkgsForCudaArch` attribute set maps CUDA architectures (e.g., `sm_89` for Ada Lovelace or `sm_90a` for architecture-specific Hopper) to Nixpkgs variants configured to support exactly that architecture. As an example, `pkgsForCudaArch.sm_89` is a Nixpkgs variant extending `pkgs` and setting the following values in `config`:

```nix
{
  cudaSupport = true;
  cudaCapabilities = [ "8.9" ];
  cudaForwardCompat = false;
}
```

::: {.note}
In `pkgsForCudaArch`, the `cudaForwardCompat` option is set to `false` because exactly one CUDA architecture is supported by the corresponding Nixpkgs variant. Furthermore, some architectures, including architecture-specific feature sets like `sm_90a`, cannot be built with forward compatibility.
:::

::: {.caution}
Not every version of CUDA supports every architecture!

To illustrate: support for Blackwell (e.g., `sm_100`) was added in CUDA 12.8. Assume our Nixpkgs' default CUDA package set is to CUDA 12.6. Then the Nixpkgs variant available through `pkgsForCudaArch.sm_100` is useless, since packages like `pkgsForCudaArch.sm_100.opencv` and `pkgsForCudaArch.sm_100.python3Packages.torch` will try to generate code for `sm_100`, an architecture unknown to CUDA 12.6. In that case, you should use `pkgsForCudaArch.sm_100.cudaPackages_12_8.pkgs` instead (see [Using cudaPackages.pkgs](#cuda-using-cudapackages-pkgs) for more details).
:::

The `pkgsForCudaArch` attribute set makes it possible to access packages built for a specific architecture without needing to manually call `pkgs.extend` and supply a new `config`. As an example, `pkgsForCudaArch.sm_89.python3Packages.torch` provides PyTorch built for Ada Lovelace GPUs.

### Running Docker or Podman containers with CUDA support {#cuda-docker-podman}

It is possible to run Docker or Podman containers with CUDA support. The recommended mechanism to perform this task is to use the [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/index.html).

The NVIDIA Container Toolkit can be enabled in NixOS like follows:

```nix
{ hardware.nvidia-container-toolkit.enable = true; }
```

This will automatically enable a service that generates a CDI specification (located at `/var/run/cdi/nvidia-container-toolkit.json`) based on the auto-detected hardware of your machine. You can check this service by running: