nvidia-container-toolkit: reintroduce nvidia runtime wrappers

This allows users to keep using `docker run --gpus`. Despite CDI is
the recommended way to expose GPU's to containers nowadays, allow
users to keep using the old `--gpus` method.
This commit is contained in:
Rafael Fernández López
2025-07-16 22:38:11 +02:00
parent c804ea5a00
commit 117bbae49c
5 changed files with 196 additions and 185 deletions
@@ -39,7 +39,6 @@
};
in
{
hardware.nvidia-container-toolkit = {
enable = lib.mkOption {
default = false;
@@ -129,137 +128,191 @@
'';
};
};
};
config = lib.mkIf config.hardware.nvidia-container-toolkit.enable {
assertions = [
{
assertion =
config.hardware.nvidia.datacenter.enable
|| lib.elem "nvidia" config.services.xserver.videoDrivers
|| config.hardware.nvidia-container-toolkit.suppressNvidiaDriverAssertion;
message = ''`nvidia-container-toolkit` requires nvidia drivers: set `hardware.nvidia.datacenter.enable`, add "nvidia" to `services.xserver.videoDrivers`, or set `hardware.nvidia-container-toolkit.suppressNvidiaDriverAssertion` if the driver is provided by another NixOS module (e.g. from NixOS-WSL)'';
}
{
assertion =
((builtins.length config.hardware.nvidia-container-toolkit.csv-files) > 0)
-> config.hardware.nvidia-container-toolkit.discovery-mode == "csv";
message = ''When CSV files are provided, `config.hardware.nvidia-container-toolkit.discovery-mode` has to be set to `csv`.'';
}
];
config = lib.mkMerge [
(lib.mkIf config.virtualisation.docker.enableNvidia {
environment.etc."nvidia-container-runtime/config.toml".text = ''
disable-require = true
supported-driver-capabilities = "compat32,compute,display,graphics,ngx,utility,video"
[nvidia-container-cli]
environment = []
ldconfig = "@${lib.getExe' pkgs.glibc "ldconfig"}"
load-kmods = true
no-cgroups = false
path = "${lib.getExe' pkgs.libnvidia-container "nvidia-container-cli"}"
[nvidia-container-runtime]
mode = "auto"
runtimes = ["docker-runc", "runc", "crun"]
[nvidia-container-runtime-hook]
path = "${lib.getOutput "tools" config.hardware.nvidia-container-toolkit.package}/bin/nvidia-container-runtime-hook"
skip-mode-detection = false
[nvidia-ctk]
path = "${lib.getExe' config.hardware.nvidia-container-toolkit.package "nvidia-ctk"}"
'';
virtualisation.docker = {
daemon.settings = lib.mkIf (lib.versionAtLeast config.virtualisation.docker.package.version "25") {
features.cdi = true;
};
rootless.daemon.settings =
lib.mkIf
(
config.virtualisation.docker.rootless.enable
&& (lib.versionAtLeast config.virtualisation.docker.package.version "25")
)
{
features.cdi = true;
virtualisation.docker = {
daemon.settings = {
default-runtime = "nvidia";
runtimes.nvidia = {
path = "${lib.getOutput "tools" config.hardware.nvidia-container-toolkit.package}/bin/nvidia-container-runtime";
args = [ ];
};
};
};
hardware = {
graphics.enable = lib.mkIf (!config.hardware.nvidia.datacenter.enable) true;
nvidia-container-toolkit.mounts =
let
nvidia-driver = config.hardware.nvidia.package;
in
(lib.mkMerge [
[
{
hostPath = pkgs.addDriverRunpath.driverLink;
containerPath = pkgs.addDriverRunpath.driverLink;
}
{
hostPath = "${lib.getLib nvidia-driver}/etc";
containerPath = "${lib.getLib nvidia-driver}/etc";
}
{
hostPath = "${lib.getLib nvidia-driver}/share";
containerPath = "${lib.getLib nvidia-driver}/share";
}
{
hostPath = "${lib.getLib pkgs.glibc}/lib";
containerPath = "${lib.getLib pkgs.glibc}/lib";
}
{
hostPath = "${lib.getLib pkgs.glibc}/lib64";
containerPath = "${lib.getLib pkgs.glibc}/lib64";
}
]
(lib.mkIf config.hardware.nvidia-container-toolkit.mount-nvidia-executables [
{
hostPath = lib.getExe' nvidia-driver "nvidia-cuda-mps-control";
containerPath = "/usr/bin/nvidia-cuda-mps-control";
}
{
hostPath = lib.getExe' nvidia-driver "nvidia-cuda-mps-server";
containerPath = "/usr/bin/nvidia-cuda-mps-server";
}
{
hostPath = lib.getExe' nvidia-driver "nvidia-debugdump";
containerPath = "/usr/bin/nvidia-debugdump";
}
{
hostPath = lib.getExe' nvidia-driver "nvidia-powerd";
containerPath = "/usr/bin/nvidia-powerd";
}
{
hostPath = lib.getExe' nvidia-driver "nvidia-smi";
containerPath = "/usr/bin/nvidia-smi";
}
])
# nvidia-docker 1.0 uses /usr/local/nvidia/lib{,64}
# e.g.
# - https://gitlab.com/nvidia/container-images/cuda/-/blob/e3ff10eab3a1424fe394899df0e0f8ca5a410f0f/dist/12.3.1/ubi9/base/Dockerfile#L44
# - https://github.com/NVIDIA/nvidia-docker/blob/01d2c9436620d7dde4672e414698afe6da4a282f/src/nvidia/volumes.go#L104-L173
(lib.mkIf config.hardware.nvidia-container-toolkit.mount-nvidia-docker-1-directories [
{
hostPath = "${lib.getLib nvidia-driver}/lib";
containerPath = "/usr/local/nvidia/lib";
}
{
hostPath = "${lib.getLib nvidia-driver}/lib";
containerPath = "/usr/local/nvidia/lib64";
}
])
]);
};
systemd.services.nvidia-container-toolkit-cdi-generator = {
description = "Container Device Interface (CDI) for Nvidia generator";
wantedBy = [ "multi-user.target" ];
after = [ "systemd-udev-settle.service" ];
serviceConfig = {
RuntimeDirectory = "cdi";
RemainAfterExit = true;
ExecStart =
let
script = pkgs.callPackage ./cdi-generate.nix {
inherit (config.hardware.nvidia-container-toolkit)
csv-files
device-name-strategy
discovery-mode
mounts
extraArgs
;
nvidia-container-toolkit = config.hardware.nvidia-container-toolkit.package;
nvidia-driver = config.hardware.nvidia.package;
};
in
lib.getExe script;
Type = "oneshot";
extraPackages = [
(lib.getOutput "tools" config.hardware.nvidia-container-toolkit.package)
];
};
};
})
(lib.mkIf config.hardware.nvidia-container-toolkit.enable {
assertions = [
{
assertion =
config.hardware.nvidia.datacenter.enable
|| lib.elem "nvidia" config.services.xserver.videoDrivers
|| config.hardware.nvidia-container-toolkit.suppressNvidiaDriverAssertion;
message = ''`nvidia-container-toolkit` requires nvidia drivers: set `hardware.nvidia.datacenter.enable`, add "nvidia" to `services.xserver.videoDrivers`, or set `hardware.nvidia-container-toolkit.suppressNvidiaDriverAssertion` if the driver is provided by another NixOS module (e.g. from NixOS-WSL)'';
}
{
assertion =
((builtins.length config.hardware.nvidia-container-toolkit.csv-files) > 0)
-> config.hardware.nvidia-container-toolkit.discovery-mode == "csv";
message = ''When CSV files are provided, `config.hardware.nvidia-container-toolkit.discovery-mode` has to be set to `csv`.'';
}
];
};
warnings = lib.mkMerge [
(lib.mkIf config.virtualisation.podman.enableNvidia [
"Setting virtualisation.podman.enableNvidia has no effect and will be removed soon."
])
];
virtualisation = {
containers.containersConf.settings = {
engine = {
cdi_spec_dirs = [
"/etc/cdi"
"/var/run/cdi"
];
};
};
docker =
let
dockerVersion = config.virtualisation.docker.package.version;
in
{
daemon.settings = lib.mkIf (lib.versionAtLeast dockerVersion "25") {
features.cdi = true;
};
rootless = {
daemon.settings = lib.mkIf (lib.versionAtLeast dockerVersion "25") {
features.cdi = true;
};
extraPackages = [
(lib.getOutput "tools" config.hardware.nvidia-container-toolkit.package)
];
};
};
};
hardware = {
graphics.enable = lib.mkIf (!config.hardware.nvidia.datacenter.enable) true;
nvidia-container-toolkit.mounts =
let
nvidia-driver = config.hardware.nvidia.package;
in
(lib.mkMerge [
[
{
hostPath = pkgs.addDriverRunpath.driverLink;
containerPath = pkgs.addDriverRunpath.driverLink;
}
{
hostPath = "${lib.getLib nvidia-driver}/etc";
containerPath = "${lib.getLib nvidia-driver}/etc";
}
{
hostPath = "${lib.getLib nvidia-driver}/share";
containerPath = "${lib.getLib nvidia-driver}/share";
}
{
hostPath = "${lib.getLib pkgs.glibc}/lib";
containerPath = "${lib.getLib pkgs.glibc}/lib";
}
{
hostPath = "${lib.getLib pkgs.glibc}/lib64";
containerPath = "${lib.getLib pkgs.glibc}/lib64";
}
]
(lib.mkIf config.hardware.nvidia-container-toolkit.mount-nvidia-executables [
{
hostPath = lib.getExe' nvidia-driver "nvidia-cuda-mps-control";
containerPath = "/usr/bin/nvidia-cuda-mps-control";
}
{
hostPath = lib.getExe' nvidia-driver "nvidia-cuda-mps-server";
containerPath = "/usr/bin/nvidia-cuda-mps-server";
}
{
hostPath = lib.getExe' nvidia-driver "nvidia-debugdump";
containerPath = "/usr/bin/nvidia-debugdump";
}
{
hostPath = lib.getExe' nvidia-driver "nvidia-powerd";
containerPath = "/usr/bin/nvidia-powerd";
}
{
hostPath = lib.getExe' nvidia-driver "nvidia-smi";
containerPath = "/usr/bin/nvidia-smi";
}
])
# nvidia-docker 1.0 uses /usr/local/nvidia/lib{,64}
# e.g.
# - https://gitlab.com/nvidia/container-images/cuda/-/blob/e3ff10eab3a1424fe394899df0e0f8ca5a410f0f/dist/12.3.1/ubi9/base/Dockerfile#L44
# - https://github.com/NVIDIA/nvidia-docker/blob/01d2c9436620d7dde4672e414698afe6da4a282f/src/nvidia/volumes.go#L104-L173
(lib.mkIf config.hardware.nvidia-container-toolkit.mount-nvidia-docker-1-directories [
{
hostPath = "${lib.getLib nvidia-driver}/lib";
containerPath = "/usr/local/nvidia/lib";
}
{
hostPath = "${lib.getLib nvidia-driver}/lib";
containerPath = "/usr/local/nvidia/lib64";
}
])
]);
};
systemd.services.nvidia-container-toolkit-cdi-generator = {
description = "Container Device Interface (CDI) for Nvidia generator";
wantedBy = [ "multi-user.target" ];
after = [ "systemd-udev-settle.service" ];
serviceConfig = {
RuntimeDirectory = "cdi";
RemainAfterExit = true;
ExecStart =
let
script = pkgs.callPackage ./cdi-generate.nix {
inherit (config.hardware.nvidia-container-toolkit)
csv-files
device-name-strategy
discovery-mode
mounts
extraArgs
;
nvidia-container-toolkit = config.hardware.nvidia-container-toolkit.package;
nvidia-driver = config.hardware.nvidia.package;
};
in
lib.getExe script;
Type = "oneshot";
};
};
})
];
}
@@ -50,6 +50,14 @@ in
};
package = lib.mkPackageOption pkgs "docker" { };
extraPackages = lib.mkOption {
type = lib.types.listOf lib.types.package;
default = [ ];
description = ''
Extra packages to add to PATH for the docker daemon process.
'';
};
};
###### implementation
@@ -68,7 +76,7 @@ in
wantedBy = [ "default.target" ];
description = "Docker Application Container Engine (Rootless)";
# needs newuidmap from pkgs.shadow
path = [ "/run/wrappers" ];
path = [ "/run/wrappers" ] ++ cfg.extraPackages;
environment = proxy_env;
unitConfig = {
# docker-rootless doesn't support running as root.
+3 -6
View File
@@ -88,7 +88,7 @@ in
description = ''
**Deprecated**, please use hardware.nvidia-container-toolkit.enable instead.
Enable nvidia-docker wrapper, supporting NVIDIA GPUs inside docker containers.
Enable Nvidia GPU support inside docker containers.
'';
};
@@ -246,7 +246,7 @@ in
"net.ipv4.conf.all.forwarding" = mkOverride 98 true;
"net.ipv4.conf.default.forwarding" = mkOverride 98 true;
};
environment.systemPackages = [ cfg.package ] ++ optional cfg.enableNvidia pkgs.nvidia-docker;
environment.systemPackages = [ cfg.package ];
users.groups.docker.gid = config.ids.gids.docker;
systemd.packages = [ cfg.package ];
@@ -287,10 +287,7 @@ in
};
path =
[ pkgs.kmod ]
++ optional (cfg.storageDriver == "zfs") pkgs.zfs
++ optional cfg.enableNvidia pkgs.nvidia-docker
++ cfg.extraPackages;
[ pkgs.kmod ] ++ optional (cfg.storageDriver == "zfs") config.boot.zfs.package ++ cfg.extraPackages;
};
systemd.sockets.docker = {
@@ -5,23 +5,11 @@
makeWrapper,
buildGoModule,
formats,
configTemplate ? null,
configTemplatePath ? null,
libnvidia-container,
autoAddDriverRunpath,
}:
assert configTemplate != null -> (lib.isAttrs configTemplate && configTemplatePath == null);
assert
configTemplatePath != null -> (lib.isStringLike configTemplatePath && configTemplate == null);
let
configToml =
if configTemplatePath != null then
configTemplatePath
else
(formats.toml { }).generate "config.toml" configTemplate;
# From https://gitlab.com/nvidia/container-toolkit/container-toolkit/-/blob/03cbf9c6cd26c75afef8a2dd68e0306aace80401/Makefile#L54
cliVersionPackage = "github.com/NVIDIA/nvidia-container-toolkit/internal/info";
in
@@ -97,24 +85,10 @@ buildGoModule (finalAttrs: {
in
[ "-skip=^${builtins.concatStringsSep "$|^" skippedTests}$" ];
postInstall =
''
mkdir -p $tools/bin
mv $out/bin/{nvidia-cdi-hook,nvidia-container-runtime,nvidia-container-runtime.cdi,nvidia-container-runtime-hook,nvidia-container-runtime.legacy} $tools/bin
for bin in nvidia-container-runtime-hook nvidia-container-runtime; do
wrapProgram $tools/bin/$bin \
--prefix PATH : ${libnvidia-container}/bin:$out/bin
done
''
+ lib.optionalString (configTemplate != null || configTemplatePath != null) ''
mkdir -p $out/etc/nvidia-container-runtime
cp ${configToml} $out/etc/nvidia-container-runtime/config.toml
substituteInPlace $out/etc/nvidia-container-runtime/config.toml \
--subst-var-by glibcbin ${lib.getBin glibc}
'';
postInstall = ''
mkdir -p $tools/bin
mv $out/bin/{nvidia-cdi-hook,nvidia-container-runtime,nvidia-container-runtime.cdi,nvidia-container-runtime-hook,nvidia-container-runtime.legacy} $tools/bin
'';
meta = {
homepage = "https://gitlab.com/nvidia/container-toolkit/container-toolkit";
@@ -6,28 +6,7 @@
# Note this scope isn't recursed into, at the time of writing.
lib.makeScope newScope (self: {
# The config is only exposed as an attrset so that the user may reach the
# default values, for inspectability purposes.
dockerConfig = {
disable-require = false;
#swarm-resource = "DOCKER_RESOURCE_GPU"
nvidia-container-cli = {
#root = "/run/nvidia/driver";
#path = "/usr/bin/nvidia-container-cli";
environment = [ ];
#debug = "/var/log/nvidia-container-runtime-hook.log";
ldcache = "/tmp/ld.so.cache";
load-kmods = true;
#no-cgroups = false;
#user = "root:video";
ldconfig = "@@glibcbin@/bin/ldconfig";
};
};
nvidia-container-toolkit-docker = self.callPackage ./package.nix {
configTemplate = self.dockerConfig;
};
nvidia-container-toolkit-docker = self.callPackage ./package.nix { };
nvidia-docker = symlinkJoin {
name = "nvidia-docker";