Can't get HW Transcoding working in Docker with Nvidia

[Jackal83 0]

I followed the instructions in Topic https://emby.media/community/index.php?/topic/76937-docker-hwa-nvidia-instructions/ but i just can't get HW transcoding working in Emby Docker (official).

I have two other Docker containers in which the GPU passthrough is working, Tdarr and Jellyfin.

What is missing? What am I doing wrong?

my nvidia-smi output:

 

Logs are attached

embyserver.txt hardware_detection-63729483026.txt

[Luke 42077]

Hi, what exactly is the issue?

[Jackal83 0]

Emby isn't recognizing the Nvidia GPU (GTX 1050 Ti):

 

 

 

 

[Luke 42077]

How exactly did you configure the Docker container?

[Jackal83 0]

I used this:

 

docker run --name emby-server -d \
 --volume /srv/dev-disk-by-label-ssd2/config/emby:/config \
 --volume /srv/dev-disk-by-label-ssd2/appdata/emby:/appdata \
 --volume /srv/61ded521-e5d5-463a-8546-c32e97278133/media:/media \
 --net=host \
 -e UID=1000 \
 -e GID=100 \
 --restart=unless-stopped \
 --runtime=nvidia \
 --env NVIDIA_VISIBLE_DEVICES=all \
 --env NVIDIA_DRIVER_CAPABILITIES=all \
 emby/embyserver:latest

 

[Luke 42077]

@Jackal83 were you able to figure this out?

[Jackal83 0]

sadly no. Tried few different parameters like "--gpus all" and changing "--runtime" but always the same outcome.

what do the errors in the hardware detect log mean? especially these:

"Error": {
        "Number": -1,
        "Message": "Operation not permitted"
    },
    "Log": [
        {
            "Level": 16,
            "Category": 0,
            "Message": "Cannot load libcuda.so.1"
        },
        {
            "Level": 16,
            "Category": 0,
            "Message": "Error loading CUDA functions"
        }

 

What permissions are missing? i know the libcuda.so.1 is present

[Cthalpa 11]

Did you install the cuda drivers for linux?

I think you forgot this

[Cthalpa 11]

[wedgekc 12]

I'm not familiar with the NVIDIA pass through to docker procedure but when I upgraded to Ubuntu 20.04 hardware transcoding would not work with the same docker run command I used in 18.04.  I had to add a second GID to get it to work.  Try adding both numbers to GIDLIST from the following commands:

getent group video | cut -d: -f3

getent group render | cut -d: -f3

Something like "--env GIDLIST=44,109"

Again, not sure what version of Linux you are running or about NVIDIA but I didn't have any hardware transcoding options without adding the second GID.

[Jackal83 0]

17 hours ago, Cthalpa said:

Did you install the cuda drivers for linux?

I think you forgot this

nvidia-smi says CUDA Version 10.2. I even tried installing them again, no luck

 

10 hours ago, wedgekc said:

I'm not familiar with the NVIDIA pass through to docker procedure but when I upgraded to Ubuntu 20.04 hardware transcoding would not work with the same docker run command I used in 18.04.  I had to add a second GID to get it to work.  Try adding both numbers to GIDLIST from the following commands:

getent group video | cut -d: -f3

getent group render | cut -d: -f3

Something like "--env GIDLIST=44,109"

Again, not sure what version of Linux you are running or about NVIDIA but I didn't have any hardware transcoding options without adding the second GID.

tried that, didn't work either. I'm running OMV5 with Debian 10 (Buster)

[Cthalpa 11]

The only suggestion i have for you is to install ubuntu-drivers-common and do ubuntu-drivers autoinstall

[Luke 42077]

Let us know if that helps. Thanks.

[Jackal83 0]

I finally got HW transcoding working, although in the end I didn't use the official Docker container, I used the linuxserver.io container.

Before that I tried installing the nvidia drivers from the official Debian repository but that caused more problems and went back to the nvidia provided drivers.

 

[Cthalpa 11]

Good. Remember, that HW transcoding is not officially supported on docker anyway

[Luke 42077]

On 7/23/2020 at 10:41 AM, Jackal83 said:

I finally got HW transcoding working, although in the end I didn't use the official Docker container, I used the linuxserver.io container.

Before that I tried installing the nvidia drivers from the official Debian repository but that caused more problems and went back to the nvidia provided drivers.

 

How did you configure it?

[Jackal83 0]

57 minutes ago, Luke said:

How did you configure it?

pretty much the same as in my second post. I changed the name and the image. I didn't use the GIDLIST

[Luke 42077]

Thanks for the info. That's strange as so many have had success with the instructions in the other pinned posting.

[EODCrafter 181]

Did something break? This was working and I just checked and it's not now. I followed the tutorial for this docker create file....

Tutorial

sudo docker container create \
--name="emby-server" \
--network="host" \
--volume /opt/docker/volumes/emby/config:/config \
--volume /mnt/media:/mnt/media \
--device /dev/dri:/dev/dri \
--runtime=nvidia \
--env UID=997 \
--env GID=997 \
--env GIDLIST=997 \
--env NVIDIA_VISIBLE_DEVICES=all \
--env NVIDIA_DRIVER_CAPABILITIES=compute,utility,video \
emby/embyserver:latest

 

_emby-server_logs.txt

@softworkz

 

 

hardware_detection-63754180806.txt

Edited April 16, 2021 by EODCrafter

[Luke 42077]

2 hours ago, EODCrafter said:

Did something break? This was working and I just checked and it's not now. I followed the tutorial for this docker create file....

Tutorial

sudo docker container create \
--name="emby-server" \
--network="host" \
--volume /opt/docker/volumes/emby/config:/config \
--volume /mnt/media:/mnt/media \
--device /dev/dri:/dev/dri \
--runtime=nvidia \
--env UID=997 \
--env GID=997 \
--env GIDLIST=997 \
--env NVIDIA_VISIBLE_DEVICES=all \
--env NVIDIA_DRIVER_CAPABILITIES=compute,utility,video \
emby/embyserver:latest

 

_emby-server_logs.txt 10.15 kB · 0 downloads

@softworkz

 

 

hardware_detection-63754180806.txt 123.01 kB · 0 downloads

The minimum cuda version has gone up to accommodate new features. I would ensure you have the latest drivers installed.

[EODCrafter 181]

47 minutes ago, Luke said:

The minimum cuda version has gone up to accommodate new features. I would ensure you have the latest drivers installed.

It's fully updated in Ubuntu 18.04...Don't want 20.04 or whatever the new one is

Edited April 16, 2021 by EODCrafter

[EODCrafter 181]

Think I found solution......https://developer.nvidia.com/cuda-downloads?target_os=Linux&target_arch=x86_64&=Ubuntu&target_version=18.04&target_type=deb_local

[Luke 42077]

Well done.

[EODCrafter 181]

Your the Greatest Luke! 

[softworkz 5066]

I don't want to disturb everybody being happy and satisfied already, but for clarity and future readers:

• CUDA 11.x is not a requirement. The current minimum is CUDA 8.2
  The minimum required driver versions for this are:
  • Linux: 396.24 or newer
  • Windows: 397.93 or newer
• You never need to install the CUDA Toolkit!
  That multi-Gigabyte download is for developers only
• The installation instructions are always the same (like documented in the KB):
  • Go to nvidia.com
  • Go to driver download 
  • Follow the instructions

 

Always follow the instructions on the Nvidia site, even when the installed driver version appears to be sufficient!
Many Linux distributions include a subset of the Nvidia drivers only (e.g. without CUDA), so you cannot judge by the driver version only.