python3Packages.nvidia-ml-py: init 11.515.48

[GaetanLepage]

Description of changes

Adding the nvidia-ml-py bindings for the NVIDIA Management Library.
This is a requirement for nvitop which I would like to package next.

Things done

• Built on platform(s)
  • x86_64-linux
  • aarch64-linux
  • x86_64-darwin
  • aarch64-darwin
• For non-Linux: Is sandbox = true set in nix.conf? (See Nix manual)
• Tested, as applicable:
  • NixOS test(s) (look inside nixos/tests)
  • and/or package tests
  • or, for functions and "core" functionality, tests in lib/tests or pkgs/test
  • made sure NixOS tests are linked to the relevant packages
• Tested compilation of all packages that depend on this change using nix-shell -p nixpkgs-review --run "nixpkgs-review rev HEAD". Note: all changes have to be committed, also see nixpkgs-review usage
• Tested basic functionality of all binary files (usually in ./result/bin/)
• 22.11 Release Notes (or backporting 22.05 Release notes)
  • (Package updates) Added a release notes entry if the change is major or breaking
  • (Module updates) Added a release notes entry if the change is significant
  • (Module addition) Added a release notes entry if adding a new NixOS module
  • (Release notes changes) Ran nixos/doc/manual/md-to-db.sh to update generated release notes
• Fits CONTRIBUTING.md.

[GaetanLepage]

Hello @FRidh and @jonringer !

I think this PR should be ready.
Please, whenever you have the time, could you have a look at it ?
I am of course willing to address any necessary modification.

Thank you very much in advance :)

[nixos-discourse]

This pull request has been mentioned on NixOS Discourse. There might be relevant details there:

https://discourse.nixos.org/t/need-pr-review-nvidia-ml-py/22020/1

[FRidh]

Not quite.

nvmlLib = CDLL("libnvidia-ml.so.1")

needs to hardcode the absolute path to the library.

---

Nevermind, this is provided by the NVIDIA driver. Still, probably best to hardcode a path to the stub in the cuda toolkit. cc @SomeoneSerge

[SomeoneSerge]

RE: ctypes.CDLL("libnvidia-ml.so.1")

I think we want this line to resolve into /run/opengl-driver/lib/libnvidia-ml.so

[SomeoneSerge]

I guess we can preserve LD_LIBRARY_PATH behaviour if we try paths in this order:

''
try:
    nvmlLib = CDLL("libnvidia-ml.so.1")
except OSError:
    nvmlLib = CDLL("${addOpenGLRunpath.driverLink}")
''

Idk if it affects nixGL in any way

[GaetanLepage]

@SomeoneSerge is your suggestion supposed to be a patch ?
I don't really see what I should change...
Thank you anyway for your help !

[SomeoneSerge]

@GaetanLepage

There's the line 1679 in pynvml.py that @ FRidh mentioned:

                         nvmlLib = CDLL("libnvidia-ml.so.1")

If you run this on NixOS it's going to fail with OSError, because the directory with libnvidia-ml.so.1 won't be in the search path list. I suggest we catch this error and re-try with this library's absolute path. This library seems to be driver-specific (like libcuda.so), so instead of a path in /nix/store we probably should use a path in /run/opengl-driver/lib, deployed by NixOS at runtime

[jonringer]

as others have mentioned, if this is exposing the functionality of nvidia drivers, you need to have the CDLL call resolve to nixos driver path /run/opengl-driver/lib/<lib>

[SuperSandro2000]

Missing tests or doCheck = false and pythonImportsCheck.

[FRidh]

Could we not fix this in the stub and have all packages patch the path to the stub instead?

[GaetanLepage]

Am I then supposed to use substituteInPlace to replace the string libnvidia-ml.so.1 by /run/opengl-driver/lib/libnvidia-ml.so.1 in pynvml.py or is there something smarter to do ?

[SomeoneSerge]

Could we not fix this in the stub and have all packages patch the path to the stub instead?

I'm not sure I follow, what exactly are you suggesting?
Trying to use e.g. nvidia-smi linked against the stub won't work

substituteInPlace to replace the string libnvidia-ml.so.1 by /run/opengl-driver/lib/libnvidia-ml.so.1

It's crude, but that's what i was thinking of

[GaetanLepage]

Could we not fix this in the stub and have all packages patch the path to the stub instead?

I'm not sure I follow, what exactly are you suggesting? Trying to use e.g. nvidia-smi linked against the stub won't work

substituteInPlace to replace the string libnvidia-ml.so.1 by /run/opengl-driver/lib/libnvidia-ml.so.1

It's crude, but that's what i was thinking of

done

[FRidh]

I'm not sure I follow, what exactly are you suggesting?
Trying to use e.g. nvidia-smi linked against the stub won't work

Ah okay, I thought that would be possible. Refer from here to the stub, and then in the stub refer to the driver library.

[SomeoneSerge]

✅ Tested the runtime with

# shell.nix
let
  nixpkgs = builtins.getFlake github:GaetanLepage/nixpkgs/nvidia-ml-py;
  pkgs = import nixpkgs { config.cudaSupport = true; config.allowUnfree = true; };
  py = pkgs.python3.withPackages (ps: with ps; [
    nvidia-ml-py
  ]);
in
with pkgs;
mkShell {
  packages = [ py ];
}

and

# test.py
import pynvml

if __name__ == "__main__":
    pynvml.nvmlInit()

    print(pynvml.nvmlSystemGetDriverVersion())

    n_devices = pynvml.nvmlDeviceGetCount()
    devices = [
        pynvml.nvmlDeviceGetName(pynvml.nvmlDeviceGetHandleByIndex(i))
        for i in range(n_devices)
    ]
    print(f"{n_devices} devices:")
    print("\n".join(devices))

[GaetanLepage]

Can the libnvidia-ml.so.1 hard linking as /run/opengl-driver/lib/libnvidia-ml.so.1 pose some problems on non-NixOs systems ?

[FRidh]

Can the libnvidia-ml.so.1 hard linking as /run/opengl-driver/lib/libnvidia-ml.so.1 pose some problems on non-NixOs systems ?

Yes, this Nix package with that patch will only be usable on NixOS.

[SomeoneSerge]

Can the libnvidia-ml.so.1 hard linking as /run/opengl-driver/lib/libnvidia-ml.so.1 pose some problems on non-NixOs systems ?

This why I suggested a try-except patch instead: with it non-nixos users can set LD_LIBRARY_PATH=... as a workaround

[SomeoneSerge]

I don't find any of this mandatory, but if it makes sense for you, I meant something like:

diff --git a/pynvml.py b/pynvml.py
index 9e094c6..727716b 100644
--- a/pynvml.py
+++ b/pynvml.py
@@ -1676,7 +1676,11 @@ def _LoadNvmlLibrary():
                             nvmlLib = CDLL(os.path.join(os.getenv("ProgramFiles", "C:/Program Files"), "NVIDIA Corporation/NVSMI/nvml.dll"))
                     else:
                         # assume linux
-                        nvmlLib = CDLL("libnvidia-ml.so.1")
+                        try:
+                            nvmlLib = CDLL("libnvidia-ml.so.1")
+                        except OSError:
+                            # Assume NixOS
+                            nvmlLib = CDLL("/run/opengl-driver/lib/libnvidia-ml.so.1")
                 except OSError as ose:
                     _nvmlCheckReturn(NVML_ERROR_LIBRARY_NOT_FOUND)
                 if (nvmlLib == None):

I would, maybe, define this patch in Nix (via writeText?), and use the addOpenGLRunpath.driverLink value instead of /run/opengl-driver/... because there's an intent to rename the "opengl-driver" eventually.

Maybe it would be even simpler to substitute /run/opengl-driver/lib with driverLink after the patch phase (the substitute would require a comment explaining that it's there in case of a rename, and the CDLL line in the diff would require a comment that this gets substituted later)

[GaetanLepage]

Can the libnvidia-ml.so.1 hard linking as /run/opengl-driver/lib/libnvidia-ml.so.1 pose some problems on non-NixOs systems ?

This why I suggested a try-except patch instead: with it non-nixos users can set LD_LIBRARY_PATH=... as a workaround

Something like this ?

try:
    nvmlLib = CDLL("libnvidia-ml.so.1")
except OSError as ose:
    nvmlLib = CDLL("/run/opengl-driver/lib/libnvidia-ml.so.1")

[GaetanLepage]

I would, maybe, define this patch in Nix (via writeText?), and use the addOpenGLRunpath.driverLink value instead of /run/opengl-driver/... because there's an intent to rename the "opengl-driver" eventually.

Maybe it would be even simpler to substitute /run/opengl-driver/lib with driverLink after the patch phase (the substitute would require a comment explaining that it's there in case of a rename, and the CDLL line in the diff would require a comment that this gets substituted later)

I am not sure how to properly do this...
Do I need to create a separate libnvidia-path.patch file and reference it in the derivation ?

[SomeoneSerge]

buildPythonPackage {
    # ...
    patches = [
        ./0001-locate-libnvidia-ml.so.1-on-NixOS.patch
    ];
    # ...
}

or

buildPythonPackage {
    # ...
    patches = [
        (writeText "0001-locate-libnvidia-ml.so.1-on-NixOS.patch" ''
            ...
        '')
    ];
    # ...
}

[GaetanLepage]

Ok, I have adopted the patch strategy.
It should be ok now.

@SomeoneSerge could you please test on your system ?

[SomeoneSerge]

• Works on NixOS
• Haven't tested LD_LIBRARY_PATH=... but I expect it to work
• Last thing to choose is whether we account for the value of addOpenGLRunpath.driverLink right now, or later in WIP nixos/opengl: move to hardware.drivers #158079

[SuperSandro2000]

or

buildPythonPackage {
    # ...
    patches = [
        (writeText "0001-locate-libnvidia-ml.so.1-on-NixOS.patch" ''
            ...
        '')
    ];
    # ...
}

Please don't do that. It needlessly creates a derivation.

Last thing to choose is whether we account for the value of addOpenGLRunpath.driverLink right now, or later in #158079

This is a WIP RFC. Please don't adopt to it yet. It might change or not come into place for a few more months.

[SuperSandro2000]

@ofborg build python310Packages.nvidia-ml-py

[nixos-discourse]

This pull request has been mentioned on NixOS Discourse. There might be relevant details there:

https://discourse.nixos.org/t/python-dll-import/22594/3