!eWOErHSaiddIbsUNsJ:nixos.org

NixOS CUDA

180 Members
CUDA packages maintenance and support in nixpkgs | https://github.com/orgs/NixOS/projects/27/ | https://nixos.org/manual/nixpkgs/unstable/#cuda35 Servers

Load older messages


SenderMessageTime
28 Apr 2024
@srhb:matrix.orgsrhbI mean, rollbacks are cheap in this world.13:50:09
@srhb:matrix.orgsrhbBut sure, it's not the core of your problem.13:50:16
@vid:matrix.org@vid:matrix.orgmy main concern is I'm travelling for a few months so don't want it being all crashy while I'm away13:50:28
@vid:matrix.org@vid:matrix.org * my main concern is I'm travelling for a few months so don't want it being all crashy while I'm away, but I want to use it remotely for ML things13:50:49
@srhb:matrix.orgsrhbSorry, I can't gauge stability for you. :) I run a custom hacked module on top of 23.11, I think the version in unstable is nicer, and personally I wouldn't be scared of it. YMMV.13:53:17
@vid:matrix.org@vid:matrix.orgokily, I will check it out, ty13:54:45
@srhb:matrix.orgsrhbBut regardless, once you have that path (which I do think is from the driver, IIRC) you need to have that available in your container, and point $whatever-software at it. Often using LD_PRELOAD, if it's not something that knows about nix paths already (ie. /run/opengl-driver)13:55:36
@vid:matrix.org@vid:matrix.orgahh, in this case it's for llama.cpp under docker, but there will be other things13:58:56
@vid:matrix.org@vid:matrix.orgI run into this famous "No help topic for" issue with nvidia-container-toolkit14:41:26
@ss:someonex.netSomeoneSerge (UTC+3)
In reply to @vid:matrix.org
Hi all, I am running nixos 23.11. I have an intel 12700k igpu for display, and an nvidia 3090 I want to use for cuda. That is basically working, displays work as does nvidia-smi. However, I need to use this with docker, which apparently needs libnvidia-ml.so.1, which I cannot install. Is there an answer, or a better place to ask? thanks!
Can't say much about the state of 2311 rn, but unstable offers a CDI module
17:09:01
@ss:someonex.netSomeoneSerge (UTC+3)
In reply to @vid:matrix.org
I can't find how it should be installed.
It should be mounted from the host
17:09:34
29 Apr 2024
@ss:someonex.netSomeoneSerge (UTC+3) joined the room.11:43:06
@vid:matrix.org@vid:matrix.org
In reply to @ss:someonex.net
Can't say much about the state of 2311 rn, but unstable offers a CDI module

I switched over to unstable, but am still going in circles. I feel like I'm one statement away from success, but it still seems infinitely far away. here's my cuda-specific config at the moment (it's included from configuration.nix):

{ config, pkgs, ... }:

{
  boot.extraModulePackages = [ pkgs.linuxPackages.nvidia_x11 ];
  hardware.opengl.enable = true; # needed for docker
  systemd.enableUnifiedCgroupHierarchy = false; # otherwise nvidia-docker fails
  
  hardware.nvidia = {
    modesetting.enable = true;
    nvidiaSettings = true;
    package = config.boot.kernelPackages.nvidia_x11;
    powerManagement.finegrained = false;
    powerManagement.enable = true;
    open = false;
  };
  
  environment.systemPackages = with pkgs; [
    btop
    cudatoolkit
#    nvidia-container-toolkit
    linuxPackages.nvidia_x11
  ];
  nixpkgs.config.allowUnfree = true;
  
  nixpkgs.config.cudaSupport = true;

  boot.blacklistedKernelModules = [ "nouveau"];
#  services.xserver.videoDrivers = ["nvidia"];

  hardware.nvidia-container-toolkit.enable = true;
  virtualisation = {
    containers = {
      enable = true;
      cdi.dynamic.nvidia.enable = true;
    };
    docker = {
#     rootless = {
#        setSocketVariable = false;
#      };
      enable = true;
      # CDI is feature-gated and only available from Docker 25 and onwards
     package = pkgs.docker_25;
     daemon.settings.features.cdi = true;
    };
 #   oci-containers.backend = "docker";
  };
  
}
12:52:21
@vid:matrix.org@vid:matrix.orga lot of that is probably superstition, I've accessed a lot of in-flight convos/repos12:53:03
@vid:matrix.org@vid:matrix.orgit doesn't even seem to be including cuda_runtime.h, which I thought would be part of cudatoolkit12:55:39
@vid:matrix.org@vid:matrix.orgI don't think I'm trying to do anything too exotic though, so I don't know why it doesn't "just work." The only thing different is I want to use Intel IGP for display (which works fine), nvidia for cuda12:57:25
@vid:matrix.org@vid:matrix.org * I don't think I'm trying to do anything too exotic though, basically "cuda in docker with widely used hardware," so I don't know why it doesn't "just work." The only thing different is I want to use Intel IGP for display (which works fine), nvidia for cuda12:58:08
@vid:matrix.org@vid:matrix.org * I don't think I'm trying to do anything too exotic though, basically "cuda in docker with widely used hardware," so I don't know why it doesn't "just work." The only thing different is I want to use Intel IGP for display (which works fine), nvidia for cuda. at this point I'm willing to sacrifice the ~250mb of using the nvidia display driver if it "just works"12:58:54
@vid:matrix.org@vid:matrix.org * I don't think I'm trying to do anything too exotic though, basically "cuda in docker with widely used hardware," so I don't know why it doesn't "just work." The only thing different is I want to use Intel IGP for display (which works fine), nvidia for cuda. at this point I'm willing to sacrifice the ~250mb for using the nvidia display driver if it "just works"12:59:12
@vid:matrix.org@vid:matrix.orgone notable thing is I don't have services.xserver.videoDrivers; defined at all. "intel" didn't work for some reason (12700k).13:00:30
@vid:matrix.org@vid:matrix.org * one notable thing is I don't have services.xserver.videoDrivers; defined at all. "intel" didn't work for some reason (12700k). 13:00:41
@vid:matrix.org@vid:matrix.org * one notable thing is I don't have services.xserver.videoDrivers defined at all. "intel" didn't work for some reason (12700k). 13:01:46
@matthewcroughan:defenestrate.itmatthewcroughan changed their profile picture.13:08:05
@ss:someonex.netSomeoneSerge (UTC+3)Tge option is at Hardware nvidia container toolkit13:13:49
@ss:someonex.netSomeoneSerge (UTC+3)* The option is at Hardware nvidia container toolkit13:14:08
@ss:someonex.netSomeoneSerge (UTC+3)
In reply to @vid:matrix.org
one notable thing is I don't have services.xserver.videoDrivers defined at all. "intel" didn't work for some reason (12700k).
You need to enable either the nvidia_x11 or the datacenter driver
13:15:09
@ss:someonex.netSomeoneSerge (UTC+3)* The option is at Hardware nvidia container toolkit Edit: I see now you have set it13:15:54
@vid:matrix.org@vid:matrix.orgI don't see many references to the datacentre driver, if it'd work with a 3090 it seems like it would be good for not causing complexity with displays.13:19:16
@ss:someonex.netSomeoneSerge (UTC+3)I use nvidia_x11 with a 309013:21:05
@ss:someonex.netSomeoneSerge (UTC+3)(Note you don't have to enable the xserver or anything to use it; the option is just confusingly named)13:22:47

Show newer messages


Back to Room ListRoom Version: 9