Arch Linux User Repository

see the release notes here for why this exists: https://github.com/NVIDIA/nvidia-container-runtime/releases/tag/3.1.0

tl;dr: nvidia-docker is deprecated because docker now has native gpu support, which this package is required to use. :)

Please remove replaces directive. Thank you. (No need to reply via comment; AUR now sends commit notifications.)

@GeorgeRaven After doing that, it works with podman, but not with Docker. How do you make it work with Docker?

$ sudo podman run --rm --device nvidia.com/gpu=all ubuntu nvidia-smi -L
GPU 0: NVIDIA GeForce GTX 1060 3GB

$ docker run --gpus all nvidia/cuda:12.1.1-runtime-ubuntu22.04 nvidia-smi
docker: Error response from daemon: failed to create task for container: failed to create shim task: OCI runtime create failed: runc create failed: unable to start container process: error during container init: error running hook #0: error running hook: exit status 1, stdout: , stderr: Auto-detected mode as 'legacy'
nvidia-container-cli: initialization error: driver rpc error: failed to process request: unknown.
ERRO[0001] error waiting for container:

This is just a note to anyone who is up-to-date and is trying to use the NCT with CDI to pass GPUS via config in /etc/cdi e.g. for pytorch. When trying to use nvidia-ctk resulting in nvidia-ctk: error while loading shared libraries: unexpected PLT reloc type 0x00

You may want to see current discussion at the following, with temporary binary PKGBUILD: https://gitlab.com/nvidia/container-toolkit/container-toolkit/-/issues/17

So that you can properly generate your CDI config:

sudo nvidia-ctk cdi generate --output=/etc/cdi/nvidia.yaml

After this I can now run rootless podman gpu containers:

> podman run --rm --device nvidia.com/gpu=all ubuntu nvidia-smi -L
GPU 0: NVIDIA GeForce GTX 1080 Ti

and

> podman run --rm --device nvidia.com/gpu=0 --userns keep-id -it docker.io/pytorch/pytorch python -c "import torch; print(torch.cuda.get_device_name(0))"
NVIDIA GeForce GTX 1080 Ti

@bitflipper First, that is a build error for libnvidia-container, not for this package. Second, manjaro is not supported on the aur, only archlinux is. I do not know what the output of uname -m produces on your machine, but if it does not match one of x86_64, ppc64le, or aarch64 then it will not be supported by this application as a requirement of the build system from https://github.com/NVIDIA/libnvidia-container, not from this pkgbuild.

I'm attempting to install nvidia-container-toolkit via paru, but receiving the following error:

==> Starting prepare()...
patching file Makefile
patching file mk/common.mk
Hunk #1 succeeded at 27 (offset 1 line).
patching file mk/elftoolchain.mk
Hunk #1 succeeded at 42 (offset 1 line).
patching file deps/src/elftoolchain-0.7.1/mk/elftoolchain.lib.mk
patching file modprobe-utils/nvidia-modprobe-utils.c
patching file modprobe-utils/nvidia-modprobe-utils.h
==> Sources are ready.
libnvidia-container-1.13.5-1 (libnvidia-container libnvidia-container-tools): parsing pkg list...
==> Making package: libnvidia-container 1.13.5-1 (Fri 21 Jul 2023 12:05:37 AM PDT)
==> Checking runtime dependencies...
==> Checking buildtime dependencies...
==> WARNING: Using existing $srcdir/ tree
==> Removing existing $pkgdir/ directory...
==> Starting build()...
/home/redacted/.cache/paru/clone/libnvidia-container/src/libnvidia-container-1.13.5/mk/common.mk:55: *** Unsupported architecture.  Stop.
==> ERROR: A failure occurred in build().
    Aborting...
error: failed to build 'libnvidia-container-1.13.5-1 (libnvidia-container libnvidia-container-tools)': 
error: can't build nvidia-container-toolkit-1.13.5-1, deps not satisfied: libnvidia-container-tools>=1.9.0
error: packages failed to build: libnvidia-container-1.13.5-1 (libnvidia-container libnvidia-container-tools)  nvidia-container-toolkit-1.13.5-1

This is really odd, I'm running Manjaro and my pc is x86_64. Not really sure what else to try here. Anybody know what's going on?

@Sparticuz does this only happen with the latest version (1.13.4)? If so that should not be the problem with this PKGBUILD since we didn't modify the build process at all (and haven't done so for a very long time).

I couldn't install the latest version without upgrading my version of go to 1.20. I had gcc-go installed and had to replace it with just regular go.

Thanks so much to @jshap, @kiendang, and @sebdotv for your work on this! Really appreciate it! :)

Thank you for that sebdotv, applied it just now.

Here's a commit that fixes the "version unknown" issue: https://github.com/sebdotv/nvidia-container-toolkit/commit/515b7ee34df17a7425a06aa036fab71f4eb457fb