XDG_RUNTIME_DIR not set, + vulkan issues

[ayunami2000]

1) XDG_RUNTIME_DIR is not set. It can be set manually by passing -e XDG_RUNTIME_DIR=/tmp 2) vkcube and vulkaninfo both fail: 3) SteamVR fails, probably due to the above ^ issue with Vulkan, but I'm not sure.

GPU is a 3090 on RunPod. Similar issues on Vast.ai.

[ayunami2000]

Notice: The following was unsuccessful, so do not think it's that one fix you've been looking for, because it does not work:

Vulkan issues were fixed after installing aptitude then running sudo aptitude install nvidia-driver-515=515.48.07-0ubuntu1 libnvidia-gl-515=515.48.07-0ubuntu1 nvidia-dkms-515=515.48.07-0ubuntu1 libnvidia-compute-515=515.48.07-0ubuntu1 libnvidia-extra-515=515.48.07-0ubuntu1 nvidia-compute-utils-515=515.48.07-0ubuntu1 libnvidia-decode-515=515.48.07-0ubuntu1 libnvidia-encode-515=515.48.07-0ubuntu1 nvidia-utils-515=515.48.07-0ubuntu1 xserver-xorg-video-nvidia-515=515.48.07-0ubuntu1 libnvidia-cfg1-515=515.48.07-0ubuntu1 libnvidia-fbc1-515=515.48.07-0ubuntu1 libnvidia-common-515=515.48.07-0ubuntu1 nvidia-kernel-common-515=515.48.07-0ubuntu1

Edit: never mind, now apt hates me Edit 2: I was able to fix apt being mad at me by editing /var/lib/dpkg/status and removing the broken packages from things that depend on them Edit 3: turns out it's using llvmpipe, and vulkan is being held up by mesa-vulkan-drivers ://////

[ehfd]

I use Vulkan a lot of times provided everything was configured in the host properly.

1. So far, I had no issues without XDG_RUNTIME_DIR set. This is not the core issue.
2. You need to use nvidia-docker2 or NVIDIA/k8s-device-plugin for this to properly work. Please confirm you have this applied in the host, outside the container.
3. sudo ubuntu-drivers autoinstall is one way to install NVIDIA drivers properly.

[ayunami2000]

Apparently, applications are supposed to handle when XDG_RUNTIME_DIR is not set, but ALVR crashes (lol) (1). OpenGL works (glxgears) and nvidia-smi works as well. I'm not sure about how they are running it with the GPU (2), but I'll give sudo ubuntu-drivers autoinstall a try (3).

[ayunami2000]

Ok, I've run sudo ubuntu-drivers autoinstall, and I got this.

[ayunami2000]

Update: vulkaninfo works when run like DISPLAY= vulkaninfo (unsetting DISPLAY

Update (this had better not be it): Possibility that I had used the other repo (the egl one) instead. Checking now to be sure.

Edit: Nope, both the egl one and the glx one both act similarly with vulkan. (neither work)

Update: the 18.04 glx one's smoketest just gives me: BUT vulkaninfo on 18.04 WORKS it WORKS OMG

And, final edit for now: Edit: on 18.04 vulkan still doesnt work. i DID dump vulkaninfo output if you want it. lmk

[ehfd]

Ok, I've run sudo ubuntu-drivers autoinstall, and I got this.

I meant outside the container in the host. The host has to be properly configured. Consult this: https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html

[ayunami2000]

The host meets the requirements. It's very likely something wrong with the docker image, not the host. (but I could be wrong)

[ehfd]

DO NOT install your own drivers. The startup script is supposed to install it for you. In a clean, with no modifications, fresh nvidia-glx-desktop container, after startup get the following outputs: glxinfo | head glxgears -info (get the first 50 lines) nvidia-smi Let's start from here. Oh and please come to https://discord.com/invite/wDNGDeSW5F as this might require long troubleshooting.

[ehfd]

https://github.com/NVIDIA/nvidia-container-toolkit/issues/140 This is the issue. In NVIDIA_DRIVER_CAPABILITIES compute,utility,video,graphics,display must all be there or simply use all. display is missing in this case.

[ehfd]

To include in documentation.

[ehfd]

Added to Documentation.