[Package request] Nvidia CUDA Kit

[ghost]

I am looking to install the nvidia CUDA toolkit to do some CUDA C programming. Is that available?

[daniel-eys]

CUDA is currently not packaged.

[ghost]

Okay, so I found out how to install it.
I downloaded the .run file from the website for Ubuntu (or Debian I think) and ran it like so:
sudo sh *.run --silent --toolkit --samples --samplespath=/home/david/cuda-test/
I did silent so it does not throw an error on ncurses.
I installed the toolkit as I think the driver is the regular nvidia driver (could be wrong on this).
I installed the samples and provided the path so I can look through them.
I think the important one is --silent as otherwise it complains about ncurses.

It will install silently in the background with logs at /var/log/cuda*-install* somewhere.

It will install cuda at /usr/local/cuda* (It will install it at /usr/local/cuda but then symlink it from the version (for me it was cuda 11.1).

You can then add cuda to your path or use the full path to compile stuff.
With path updated it becomes:
nvcc -o hello hello.cu

Without path updated it becomes:
sudo /usr/local/cuda/bin/nvcc -o hello hello.cu

Hopes this helps people trying to install CUDA on Void.
I am trying to play around with Python Machine Learning libraries that require this to run so there might be some more steps. I know Tensorflow wants very specific versions of CUDA so I might have to uninstall it and reinstall another version.

I figured a lot of this stuff out by running the installer with the help flag.
cuda_11.1.1_455.32.00_linux.run --help
and someone online was nice enough to point out that it wants ncurses to do a graphical install and nvidia does not like binary packages which void gives so we are forced to use the --silent flag to install it via CLI (not that it matters, I do not really like graphical installs anyways.)

If you are having trouble with this, another way of getting CUDA on void is to use the nvidia-cuda docker image (but I would assume you then need some configuring to get your GPU to play nice with it -- not sure on this).

[ghost]

Unforunately using pyenv or similar tools seems iffy as they can not use pip due to an OpenSSL versus Libressl issue.

[WormChickenWizard]

Okay, so I found out how to install it.
I downloaded the .run file from the website for Ubuntu (or Debian I think) and ran it like so:
sudo sh *.run --silent --toolkit --samples --samplespath=/home/david/cuda-test/

I did a fresh install of the glibc base version, installed the Nvidia driver from the nonfree repo and after running the cuda ubuntu installer like demonstrated, it errored out for me.

It would be nice if cuda had its own package for simplicity's sake.

[github-actions]

Issues become stale 90 days after last activity and are closed 14 days after that. If this issue is still relevant bump it or assign it.

[soanvig]

We have nvidia driver in the repository, and we have hashcat. Hashcat in this case however requires CUDA. Just "bumping" this, but with a context at least.

[jjc224]

Literally came here for the same reason as @soanvig. Would be great to have this.

[Juggernaut-Coder]

Is there any possibility of getting CUDA on Repository?

[namgo]

To anyone who stumbles across this or #31984 it might be helpful to look at what slackbuilds does https://github.com/Ponce/slackbuilds/blob/f7519a395ed941f7ae6fa59051a1bcd770821b1a/development/cudatoolkit/cudatoolkit.SlackBuild which is to add --noexec to the .run and then unpack everything manually?

I've got a void x86_64-glibc machine that I would very much like to have CUDA on, I'm a little hesitant to try anything that provides breaking changes to it today, but if nobody gets on this in the next few days, I will.

[mcneb10]

Also to anyone installing this, the installer needs ncurses 5. I had to make the symlinks for version 5 from the version of ncurses I had installed (6). Other than that the driver works flawlessly so far. I had to install libtinfo as well make symlinks for that as well

[mcneb10]

The installer (extracted from the installer package) dynamically loads libncursesw, libformw, and libtinfo so one must make version 5 symlinks for those libs

[mcneb10]

I also had to add myself to the video group to get OpenGL to work

[namgo]

If anyone's still struggling with this (as I am) the cuda installer when extracted provides directories that can be merged into an "outdir"

NVCC=/opt/cuda-extract/outdir/bin/nvcc LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/cuda-extract/outdir/lib64/ INCLUDES="-I /opt/cuda-extract/outdir/include/ " PATH=$PATH:/opt/cuda-extract/outdir/bin/ NVVMIR_LIBRARY_DIR=/opt/cuda-extract/outdir/nvvm/libdevice/ CUDA_PATH=/opt/cuda-extract/outdir/targets/x86_64-linux/ (command) at least points nvcc to the right place.

This is isn't directly relevant to packaging as I don't think we can legally redistribute extracted cuda binaries.

Edit: maybe we can?? https://archlinux.org/packages/extra/x86_64/cuda/

edit: I managed to get a python package to compile its cuda-required code successfully using the env vars above!

[mcneb10]

Perhaps I could write and package and a script that automatically downloads the driver and prompts the user to accept the license agreement

[namgo]

@mcneb10 I'm asking on irc today, but given how archlinux packages cuda, it MIGHT be possible to distribute the binaries/libraries/headers as their own package.

One nice thing about the CUDA bins/libs/headers is that they're really just that. It appears that the driver that void-linux ships with is cuda-compatible. (note: this is not a well-tested statement and I'm questioning myself - see next comment).

I'll update when I hear back from others on how to proceed.

The EULA is confusing to me, otoh we're legally able to distribute the SDK, but only if it's used by our "application". This makes me think I should contact one of the arch linux packagers to see what they do.

• I sent off an email to the arch linux package maintainer responsible for CUDA, I'll hopefully hear back soon.
• if I hear back with an easy licensing option, I'll describe it here
• If it requires someone signing off on a legalities email, I'll let one of the void-linux maintainers know that they can do that.

Meanwhile, I'll spend some time and figure out the least ugly hack to re-package CUDA under void's templating. Even if it ends up being a licensing nightmare, it'd be good to document how to do it.

[namgo]

This comment is less important than the former, but:

https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html#id5 is an important reference for driver versions to cuda versions if we're bundling them separately. Nvidia driver is 535.129(.03?) as of Dec 5, 2023 which corresponds to CUDA 12.2 - I had downloaded 12.3 and was encountering:

cuLinkAddData(): the provided PTX was compiled with an unsupported toolchain. running hashcat as a generally available quick test.

I think this is important for matching CUDA versions to drivers if we package (https://developer.nvidia.com/cuda-toolkit-archive).

We can also always package a nvidia-cuda-download-helper script if the cuda licensing becomes an issue.

Here's an example of how nvidia-cuda-download-helper might look: https://gist.github.com/namgo/9aeeee2025b341779be85e596822893f - I've tested the result with hashcat briefly but haven't tried building pytorch with it yet. If CUDA licensing is really an issue, a restricted package might be necessary then? Still waiting to hear back, but the above script seems to be a start.

[atisharma]

I would LOVE to have cuda packaged but at the moment cuda SDK requires gcc <= 12 and gcc 13 seems to be in the repos. How would that be handled?

[namgo]

@atisharma I didn't consider that, good point!

I think that a temporary solution for those desperate would be to install a non-system-wide gcc12 and a non-system-wide cuda, making use of environment variables?

So:

• download the nvidia cuda installer
• download gcc12 sources
• compile gcc12 with DESTDIR set somewhere like /opt/gcc12
• ./cuda_12.2.2_535.104.05_linux.run --extract=/opt/cuda-12.2.2/extract
• (possibly link/move things around like my gist above)
• set LD_LIBRARY_PATH and PATH to tell gcc and cuda where to find things

But that's not something that one can/should package, so I'm pretty stuck still.

[jlhamilton777]

It looks like GCC 13.2 is supported now https://docs.nvidia.com/cuda/cuda-installation-guide-linux/#host-compiler-support-policy