In reply to @ss:someonex.net
Should work this time: https://github.com/NixOS/nixpkgs/pull/323056
Can I bum a nixpgks-review? xD

In reply to @matthewcroughan:defenestrate.it
they removed the .6 from the release

In reply to @search-sense:matrix.org
I know, that it's broken ... actually it would be good if someone upgrade it to the current version TensorRT-10.1.0.27

In reply to @ss:someonex.net
Should work this time: https://github.com/NixOS/nixpkgs/pull/323056
Can I bum a nixpgks-review? xD

In reply to @ss:someonex.net
Would you like to just take over tensorrt in Nixpkgs?

Hey! I just started using NixOS and I love it but have a MAJOR blocker, as I'm maintaining a FOSS deep learning package and can't get CUDA to work :( I would really love to continue on this journey and also eventually contribute to this community here, but right now it feels like I just shot myself in the foot badly, as I've spent the last days exclusively configuring NixOS only to reach a point which is seemingly unsurmountable for me.. The issue seems to be that PyTorch doesn't find the CUDA driver and what's also weird is that nvidia-smi seems to work fine, but shows CUDA Version: ERR!

The thing is that in order to work with my collaborators, I need to work in a non NixOS way, in my case I would like to use pixi which is very much like conda/micromamba, just better.. Therefore, I'm trying to get things working in an FHS shell. Does one of you have an idea? Am I doing anything obvious wrong?

from my configuration.nix

  hardware.opengl = {
    enable = true;
    driSupport = true;
    driSupport32Bit = true;
  };


  # Allow unfree packages
  nixpkgs.config.allowUnfree = true;
  services.xserver.videoDrivers = ["nvidia"];

  hardware.nvidia = {
    modesetting.enable = true;
    powerManagement.enable = false;
    powerManagement.finegrained = false;
    open = false;
    package = config.boot.kernelPackages.nvidiaPackages.beta;
  };

pixi-fhs.nix

{ pkgs, unstable }:

let
cudatoolkit = pkgs.cudaPackages.cudatoolkit_12_1;
nvidia_x11 = pkgs.nvidia_x11;
in
pkgs.buildFHSUserEnv {
name = "pixi-env";
targetPkgs = pkgs: with pkgs; [
unstable.pixi
cudatoolkit
nvidia_x11
# bashInteractive
# bash-completion
# complete-alias
];
runScript = "bash";
profile = ''
export NVIDIA_DRIVER_CAPABILITIES=compute,utility
export XDG_CONFIG_DIRS=${nvidia_x11}/share/X11/xorg.conf.d''${XDG_CONFIG_DIRS:+:}$XDG_CONFIG_DIRS
export XDG_DATA_DIRS=${nvidia_x11}/share''${XDG_DATA_DIRS:+:}$XDG_DATA_DIRS

export LD_LIBRARY_PATH=${cudatoolkit}/lib:${cudatoolkit}/lib64:${cudatoolkit}/lib64/stubs''${LD_LIBRARY_PATH:+:}$LD_LIBRARY_PATH
export CUDA_PATH=${cudatoolkit}
export PATH=${cudatoolkit}/bin:$PATH
export LIBRARY_PATH=${cudatoolkit}/lib:${cudatoolkit}/lib64:$LIBRARY_PATH

export CPLUS_INCLUDE_PATH="${cudatoolkit}/include''${CPLUS_INCLUDE_PATH:+:$CPLUS_INCLUDE_PATH}"
export C_INCLUDE_PATH="${cudatoolkit}/include''${C_INCLUDE_PATH:+:$C_INCLUDE_PATH}"

# Pixi completion -- not working yet, due to missing `complete` command
eval "$(pixi completion --shell bash 2>/dev/null)"

echo "*** Pixi environment activated, using $(which pixi). ***"

'';
}

Thanks in advance <3

Hey! I just started using NixOS and I love it but have a MAJOR blocker, as I'm maintaining a FOSS deep learning package and can't get CUDA to work :( I would really love to continue on this journey and also eventually contribute to this community here, but right now it feels like I just shot myself in the foot badly, as I've spent the last days exclusively configuring NixOS only to reach a point which is seemingly unsurmountable for me.. The issue seems to be that PyTorch doesn't find the CUDA driver and what's also weird is that nvidia-smi seems to work fine, but shows CUDA Version: ERR!

The thing is that in order to work with my collaborators, I need to work in a non NixOS way, in my case I would like to use pixi which is very much like conda/micromamba, just better.. Therefore, I'm trying to get things working in an FHS shell. Does one of you have an idea? Am I doing anything obvious wrong?

from my configuration.nix

  hardware.opengl = {
    enable = true;
    driSupport = true;
    driSupport32Bit = true;
  };


  # Allow unfree packages
  nixpkgs.config.allowUnfree = true;
  services.xserver.videoDrivers = ["nvidia"];

  hardware.nvidia = {
    modesetting.enable = true;
    powerManagement.enable = false;
    powerManagement.finegrained = false;
    open = false;
    package = config.boot.kernelPackages.nvidiaPackages.beta;
  };

pixi-fhs.nix

{ pkgs, unstable }:

let
cudatoolkit = pkgs.cudaPackages.cudatoolkit\_12\_1;
nvidia\_x11 = pkgs.nvidia\_x11;
in
pkgs.buildFHSUserEnv {
name = "pixi-env";
targetPkgs = pkgs: with pkgs; \[
unstable.pixi
cudatoolkit
nvidia\_x11
# bashInteractive
# bash-completion
# complete-alias
\];
runScript = "bash";
profile = ''
export NVIDIA\_DRIVER\_CAPABILITIES=compute,utility
export XDG\_CONFIG\_DIRS=${nvidia\_x11}/share/X11/xorg.conf.d''${XDG\_CONFIG\_DIRS:+:}$XDG\_CONFIG\_DIRS
export XDG\_DATA\_DIRS=${nvidia\_x11}/share''${XDG\_DATA\_DIRS:+:}$XDG\_DATA\_DIRS

export LD_LIBRARY_PATH=${cudatoolkit}/lib:${cudatoolkit}/lib64:${cudatoolkit}/lib64/stubs''${LD_LIBRARY_PATH:+:}$LD_LIBRARY_PATH
export CUDA_PATH=${cudatoolkit}
export PATH=${cudatoolkit}/bin:$PATH
export LIBRARY_PATH=${cudatoolkit}/lib:${cudatoolkit}/lib64:$LIBRARY_PATH

export CPLUS_INCLUDE_PATH="${cudatoolkit}/include''${CPLUS_INCLUDE_PATH:+:$CPLUS_INCLUDE_PATH}"
export C_INCLUDE_PATH="${cudatoolkit}/include''${C_INCLUDE_PATH:+:$C_INCLUDE_PATH}"

Pixi completion -- not working yet, due to missing complete command

eval "$(pixi completion --shell bash 2>/dev/null)"

echo "*** Pixi environment activated, using $(which pixi). ***"

'';
}

Thanks in advance <3

  File "/home/titus/src/bnb/bitsandbytes/diagnostics/main.py", line 66, in main
    sanity_check()
  File "/home/titus/src/bnb/bitsandbytes/diagnostics/main.py", line 33, in sanity_check
    p = torch.nn.Parameter(torch.rand(10, 10).cuda())
  File "/home/titus/src/bnb/.pixi/envs/default/lib/python3.8/site-packages/torch/cuda/__init__.py", line 293, in _lazy_init
    torch._C._cuda_init()

pixi

Is one of the package managers that ship incomplete dependencies because they expect an FHS environment. If you want to use it on NixOS I recommend you use nix-ld. You'll aslo need to ensure (using a shell e..g) that the ld.so is aware of ${addDriverRunpath.driverLink}/lib, which you can also do as part of the nix-ld configuration. E.g. you can deploy a nixos with programs.nix-ld.enable and then, in your project tree, use a nix shell that looks something like the following: https://github.com/NixOS/nixpkgs/blob/48dbb2ae90be0ba21b44e77b8278fd7cefb4b75f/nixos/doc/manual/configuration/fhs.chapter.md?plain=1#L105-L113

pkgs.buildFHSUserEnv {
name = "pixi-env";
targetPkgs = pkgs: with pkgs; [
...
nvidia_x11
...
];

...this, if it had any affect on the dynamic loader (in this form it doesn't, instead it provides hints for the compiler), would conflict with the libcuda driver deployed by NixOS. NVidia makes it so that the driver has to be deployed impurely, because each libcuda only works with the corresponding kernel. TLDR: delete nvidia_x11 from that list

Thank you so much for your help, SO appreciated!

yeah, I had added nvidia_x11 only in the latest iteration.. I'll remove it right away.

The weird thing is that even outside the fhs, nvidia smi gives me this N/A: Have you ever seen that before? Does it tell us sth useful?

❯ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2023 NVIDIA Corporation
Built on Tue_Aug_15_22:02:13_PDT_2023
Cuda compilation tools, release 12.2, V12.2.140
Build cuda_12.2.r12.2/compiler.33191640_0

Thank you so much for your help, SO appreciated!

yeah, I had added nvidia_x11 only in the latest iteration.. I'll remove it right away.

The weird thing is that even outside the fhs, nvidia smi gives me this N/A: Have you ever seen that before? Does it tell us sth useful?

❯ nvidia-smi
Fri Jun 28 14:55:14 2024       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 550.78                 Driver Version: 550.78         CUDA Version: ERR!     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA GeForce RTX 4090        On  |   00000000:01:00.0  On |                  Off |
|  0%   43C    P8             40W /  480W |    3302MiB /  24564MiB |     26%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+
|   1  NVIDIA GeForce RTX 4090        On  |   00000000:03:00.0 Off |                  Off |
|  0%   36C    P8             29W /  480W |      12MiB /  24564MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

In reply to @titus-von-koeller:matrix.org

Thank you so much for your help, SO appreciated!

yeah, I had added nvidia_x11 only in the latest iteration.. I'll remove it right away.

The weird thing is that even outside the fhs, nvidia smi gives me this N/A: Have you ever seen that before? Does it tell us sth useful?

❯ nvidia-smi
Fri Jun 28 14:55:14 2024       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 550.78                 Driver Version: 550.78         CUDA Version: ERR!     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA GeForce RTX 4090        On  |   00000000:01:00.0  On |                  Off |
|  0%   43C    P8             40W /  480W |    3302MiB /  24564MiB |     26%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+
|   1  NVIDIA GeForce RTX 4090        On  |   00000000:03:00.0 Off |                  Off |
|  0%   36C    P8             29W /  480W |      12MiB /  24564MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

❯ nix run -f '<nixpkgs>' cudaPackages.saxpy
trace: warning: cudaPackages.autoAddDriverRunpath is deprecated, use pkgs.autoAddDriverRunpath instead
Start
Runtime version: 12020
Driver version: 0
Host memory initialized, copying to the device
CUDA error at cudaMalloc(&xDevice, N * sizeof(float)): CUDA driver is a stub library

❯ ls /run/opengl-driver/lib/libcuda*
 /run/opengl-driver/lib/libcuda.so -> /nix/store/2paf3i2g4arx5j4m9l87zdrzsikwmizh-nvidia-x11-550.78-6.6.35/lib/libcuda.so*
 /run/opengl-driver/lib/libcuda.so.1 -> /nix/store/2paf3i2g4arx5j4m9l87zdrzsikwmizh-nvidia-x11-550.78-6.6.35/lib/libcuda.so.1*
 /run/opengl-driver/lib/libcuda.so.550.78 -> /nix/store/2paf3i2g4arx5j4m9l87zdrzsikwmizh-nvidia-x11-550.78-6.6.35/lib/libcuda.so.550.78*
 /run/opengl-driver/lib/libcudadebugger.so -> /nix/store/2paf3i2g4arx5j4m9l87zdrzsikwmizh-nvidia-x11-550.78-6.6.35/lib/libcudadebugger.so*
 /run/opengl-driver/lib/libcudadebugger.so.1 -> /nix/store/2paf3i2g4arx5j4m9l87zdrzsikwmizh-nvidia-x11-550.78-6.6.35/lib/libcudadebugger.so.1*
 /run/opengl-driver/lib/libcudadebugger.so.550.78 -> /nix/store/2paf3i2g4arx5j4m9l87zdrzsikwmizh-nvidia-x11-550.78-6.6.35/lib/libcudadebugger.so.550.78*