Use NVIDIA vGPU on NixOS, for both host and guest. Supports 16.x, 17.x and 18.x releases

Module adds new packages that can be selected in hardware.nvidia.package. Additionally, there is hardware.nvidia.vgpu submodule for additional configuration

Also features an integration with vGPU-Unlock-patcher

Releases not mentioned here can still be used, but you have to create your own driver derivation.

• 16.x: 16.2, 16.5, 16.8, 16.9, 16.10
• 17.x: 17.3, 17.4, 17.5, 17.6
• 18.x: 18.0, 18.1, 18.2, 18.3

vgpu4nixos supports both Flakes and non-flake environments, though former is recommended

flake.nix:

{
  inputs = {
    /* ... */
    vgpu4nixos.url = "github:mrzenc/vgpu4nixos";
  };

  outputs = { self, nixpkgs, vgpu4nixos, ... }@inputs: {
    nixosConfigurations.mrzenc = nixpkgs.lib.nixosSystem {
      system = "x86_64-linux";
      modules = [
        ./configuration.nix
        /* ... */
        vgpu4nixos.nixosModules.host # Use nixosModules.guest for VMs
      ];
      /* ... */
    };
  };
}

configuration.nix:

{ pkgs, lib, config }:
{
  imports = [
    ./hardware-configuration.nix
    /* ... */
    (import (builtins.fetchGit {
      url = "https://github.com/mrzenc/vgpu4nixos.git";
      # Pin to specific commit (example value)
      # rev = "b6ddaeb51b1575c6c8ec05b117c3a8bfa3539e92";
    }) { guest = false; }) # Use { guest = true; } for VMs
  ];

  /* ... */
}

New packages will be available in config.boot.kernelPackages.nvidiaPackages in the form of vgpu_XX_X for the host or grid_XX_X for the guest. Specify the package in your configuration as follows:

{ config, pkgs, lib, ... }:
{
  /* ... */

  # You want to include this if you're using guest drivers
  # or want to use vGPU-Unlock-patcher's merged driver
  # services.xserver.videoDrivers = [ "nvidia" ];
  
  hardware.nvidia.package = config.boot.kernelPackages.nvidiaPackages.vgpu_17_3;

  /* ... */
}

hardware.nvidia.vgpu consists of three submodules:

• patcher - driver patching utility
• griddUnlock - patch 18.x to use with fastapi-dls
• driverSource - where to get driver from

Manages vGPU-Unlock-patcher's options. Please read its README if you don't know how to use it. Most likely, you will only need to specify hardware.nvidia.vgpu.patcher.enable = true and in some cases hardware.nvidia.vgpu.patcher.copyVGPUProfiles

• hardware.nvidia.vgpu.patcher.enable (bool) - enable VUP
• hardware.nvidia.vgpu.patcher.options.doNotForceGPLLicense (bool)
  • if set to false, then the --enable-nvidia-gpl-for-experimenting --force-nvidia-gpl-I-know-it-is-wrong options will be applied, which allows using the driver with slightly newer kernels
• hardware.nvidia.vgpu.patcher.options.remapP40ProfilesToV100D (bool; only for host) - applies the --remap-p2v option. Only for 17.x releases
• hardware.nvidia.vgpu.patcher.options.extra (list of strings) - additional patch.sh command options
• hardware.nvidia.vgpu.patcher.copyVGPUProfiles (attrset; only for host) - additional vcfgclone lines (see VUP's README)
  • For example, {"AAAA:BBBB" = "CCCC:DDDD"} is the same as vcfgclone ${TARGET}/vgpuConfig.xml 0xCCCC 0xDDDD 0xAAAA 0xBBBB
• hardware.nvidia.vgpu.patcher.enablePatcherCmd (bool; only for host) - add a patcher to system packages (which will be available as nvidia-vup) for convenience
• hardware.nvidia.vgpu.patcher.profileOverrides (only for host) - custom properties for vGPU profiles

Replace * in the following options with your profile ID ("333" in case of nvidia-333). Multiple overrides can be specified

• profileOverrides.*.vramAllocation (integer) - vRAM allocation in megabytes
• profileOverrides.*.heads (integer) - the maximum number of virtual monitors for one VM
• profileOverrides.*.enableCuda (bool)
• profileOverrides.*.display.width (integer) - maximum display width in pixels
• profileOverrides.*.display.height (integer) - maximum display height in pixels
• profileOverrides.*.framerateLimit (integer) - limits FPS to a certain value (0 to disable limit)
• profileOverrides.*.xmlConfig (attrset) - additional configuration

An example of a profile override:

hardware.nvidia.vgpu.patcher.profileOverrides = {
  "333" = {
    vramAllocation = 3584; # 3.5GiB
    heads = 1;
    display.width = 1920;
    display.height = 1080;
    framerateLimit = 144;
  };
};

Integration with gridd-unlock-patcher, for 18.x guests only

• hardware.nvidia.vgpu.griddUnlock.enable (bool) - enable/disable patch
• hardware.nvidia.vgpu.griddUnlock.rootCaFile (path) - path to fastapi-dls instance's root certificate authority. Required when submodule is enabled

Manages the driver source. It can accept either Linux-KVM GRID archive or a plain .run driver file. It can also fetch from Nix store or HTTP(s) server

nix-hash --flat --base64 --type sha256 /path/to/file.zip

• hardware.nvidia.vgpu.driverSource.name (string) - driver filename
  • Host will expect a .zip by default, while guest expects a .run. The extension part of this option will force file type
• hardware.nvidia.vgpu.driverSource.url (string)
  • Set to null to force fetch from Nix store
• hardware.nvidia.vgpu.driverSource.sha256 (string)
  • If set, then .run is always expected. If not, then .zip is expected on host, on guest .run is still expected but that can be overridden by name
  • If value contradicts with name then rebuild will fail
• hardware.nvidia.vgpu.driverSource.curlOptsList (list of strings) - a list of arguments to pass to curl
  • For example, ["-u" "admin:some nice password"]

The mkVgpuDriver and mkVgpuPatcher allow you to create your own driver derivation that can be passed to hardware.nvidia.package. This way you can use a version of the vGPU that is not available by default (yet)

• version (string) - version of the host driver
• sha256 (string) - SHA-256 for GRID .zip, not .run
• guestVersion (string) - version of the guest driver
• guestSha256 (string) - SHA-256 of the guest .run
• useSettings (bool; optional) - whether to use nVidia X Server settings
• settingVersion (string; optional) - the version of the settings app. Not required if useSettings = false
• settingsSha256 (string; optional) - SHA-256 of the settings app. Not required if useSettings = false
• usePersistenced (bool; optional) - whether to use nvidia-persistenced
• persistencedVersion (string; optional) - the version of nvidia-persistenced. Not required if usePersistenced = false
• persistencedSha256 (string; optional) - SHA-256 of nvidia-persistenced. Not required if usePersistenced = false
• generalVersion (string) - The closest version of consumer nVidia graphics drivers to the vGPU version (usually with the same major and minor versions). Used to build nvidia-settings and nvidia-persistenced
• gridVersion (string) - vGPU release (for example, 16.7, 17.2...)
• zipFilename (string) - the full name of the GRID .zip file (including extension)
• vgpuPatcher - a patcher derivation obtained from mkVgpuPatcher (set to null to disable patching)
• prePatch, postPatch, patchFlags, patches, preInstall, postInstall, broken are passed directly to stdenv.mkDerivation and are optional

hardware.nvidia.package = (config.boot.kernelPackages.nvidiaPackages.mkVgpuDriver {
  version = "555.44.33";
  vgpuPatcher = config.boot.kernelPackages.nvidiaPackages.mkVgpuPatcher { /* ... */ };
  # ...
}).overrideAttrs (self: super: {
  patcher = self.patcher.override {
    # your overrides
  };
});

• version (string; optional) - the branch of the patcher, for visual appearance only
• rev (string) - git revision of the patcher (usually a specific commit or refs/heads/your-branch)
• sha256 (string) - SHA-256 of the patcher source code
• generalVersion (string) - version of the consumer (general) driver
• generalSha256 (string) - SHA-256 of the general driver
• linuxGuest (string) - version of the Linux guest drivers
• linuxSha256 (string) - SHA-256 of Linux guest drivers
• windowsGuestFilename (string) - the full name of the Windows guest driver file (including extension)
• windowsSha256 (string) - SHA-256 of Windows guest drivers
• gridVersion (string) - vGPU release (for example, 16.7, 17.2...). Useless if vgpuUrl is specified.
• generalUrl (string; optional) - URL where general drivers can be obtained
• vgpuUrl (string; optional) - URL where guest drivers can be obtained

The files in the nvidia-vgpu directory are modified versions of those from the nvidia-x11 package of Nixpkgs, which is distributed under the MIT license