Add Nvidia GPU support

[dmarts05]

Hi, I have been trying to make this work with my Nvidia GPU but it seems to be harder than I thought.

I am using an RTX 2070 with Arch Linux as my host machine and this is the error that it gives at first:
00:00:00.003 [ERROR] [sway/main.c:65] Proprietary Nvidia drivers are NOT supported. Use Nouveau. To launch sway anyway, launch with --unsupported-gpu and DO NOT report issues.

After adding that flag to entrypoint_user.sh it throws another error:
app-1 | 00:00:00.001 [ERROR] [sway/main.c:62] !!! Proprietary Nvidia drivers are in use !!! app-1 | libEGL warning: egl: failed to create dri2 screen app-1 | 00:00:00.014 [ERROR] [wlr] [EGL] command: eglInitialize, error: EGL_NOT_INITIALIZED (0x3001), message: "DRI2: failed to create screen" app-1 | libEGL warning: egl: failed to create dri2 screen app-1 | 00:00:00.017 [ERROR] [wlr] [EGL] command: eglInitialize, error: EGL_NOT_INITIALIZED (0x3001), message: "DRI2: failed to create screen" app-1 | libEGL warning: failed to open /dev/dri/card1: Permission denied app-1 | app-1 | 00:00:00.017 [ERROR] [wlr] [EGL] command: eglInitialize, error: EGL_NOT_INITIALIZED (0x3001), message: "DRI2: failed to load driver" app-1 | 00:00:00.017 [ERROR] [wlr] [EGL] command: eglInitialize, error: EGL_NOT_INITIALIZED (0x3001), message: "eglInitialize" app-1 | 00:00:00.017 [ERROR] [wlr] [render/egl.c:269] Failed to initialize EGL app-1 | 00:00:00.017 [ERROR] [wlr] [render/egl.c:572] Failed to initialize EGL context app-1 | 00:00:00.017 [ERROR] [wlr] [render/gles2/renderer.c:804] Could not initialize EGL

Any help would be appreciated, thanks!

[stephanlensky]

Hey @dmarts05, I don't currently have an Nvidia linux box to test, but this

libEGL warning: failed to open /dev/dri/card1: Permission denied

looks a lot like an error I saw on my AMD machine when the Docker user didn't have permission to access /dev/dri/renderD128.

In my case, this file was owned by the render group (gid 107) and thus I was able to resolve the issue by assigning the Docker user a group with that same gid (107). This is what the RENDER_GROUP_GID environment variable is for (see First Time Setup).

Now, it looks like on your system, instead of (or maybe in addition to?) using the /dev/dri/renderDXXX device, EGL is trying to directly access the /dev/dri/card1 device.

Here's what I would suggest:

1. Figure out what group owns /dev/dri/card1 on your host (probably it's the video group):

   stat /dev/dri/card1 -c "%G %g"
   

2. Set the RENDER_GROUP_GID environment variable in your Docker container to the gid of the group found in the previous step.

There are probably two outcomes after doing this:

1. Everything works 🥳 - this means that on Nvidia systems, you need to give the Docker user permission for /dev/dri/cardX, NOT /dev/dri/renderDXXX
2. You get another permissions error about /dev/dri/renderDXXX. This indicates that on Nvidia systems, the Docker user needs both the render AND video groups. To fix this, we'll need to adjust the Docker image to allow setting GIDs for both of these groups instead of just the render group.

[stephanlensky]

It looks like Nvidia also has some dedicated tools to more easily support using GPUs in Docker containers, but I don't have experience with it. It might be worth looking into though: https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/index.html

[dmarts05]

Hi, thanks for all the support! I'll try this out during the weekend :)

[gokanil]

I’m experiencing the same issue and tested it on Windows using Docker Desktop. Our graphics card is an RTX 3060. I was about to open a new issue, but I wanted to reach out here instead.

Were you able to solve the problem?

Update:
stat /dev/dri/card1 -c "%G %g"

# ls
card0  renderD128
# stat /dev/dri/card0 -c "%G %g"
root 0

[gokanil]

I am currently using Windows 11 and have installed the Docker Desktop application. I have made the following adjustments to the docker-compose and Dockerfile configurations. I was able to successfully display the graphics card information on the screen. However, I am still encountering a permission error.

To resolve this, I used the following command:
run sed -i '/set -e/a chmod 777 /dev/dri/*' /entrypoint.sh.

Despite this, I encountered the following errors:

My primary goal is to use the Pythonnet library with Zendriver Python scripts in a .NET application. I am open to any suggestions or ideas you might have.

Dockerfile

FROM ghcr.io/stephanlensky/swayvnc-chrome:latest

run apt-get update && apt install -y curl gpg

run curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey | gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg \
&& curl -s -L https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list | \
  sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' | \
   tee /etc/apt/sources.list.d/nvidia-container-toolkit.list

run apt-get update && apt-get install -y nvidia-container-toolkit

ENV PYTHONUNBUFFERED=1

COPY --from=ghcr.io/astral-sh/uv:latest /uv /uvx /bin/

ENV UV_COMPILE_BYTECODE=1
ENV UV_LINK_MODE=copy

# Make directory for the app
RUN mkdir /app
RUN chown $DOCKER_USER:$DOCKER_USER /app

# Switch to the non-root user
USER $DOCKER_USER

# Set the working directory
WORKDIR /app

# Install python
RUN uv python install 3.13

# Install the Python project's dependencies using the lockfile and settings
COPY --chown=$DOCKER_USER:$DOCKER_USER pyproject.toml uv.lock /app/
RUN --mount=type=cache,target=/home/$DOCKER_USER/.cache/uv,uid=$PUID,gid=$PGID \
    uv sync --frozen --no-install-project

# Then, add the rest of the project source code and install it
# Installing separately from its dependencies allows optimal layer caching
COPY --chown=$DOCKER_USER:$DOCKER_USER . /app

# Add binaries from the project's virtual environment to the PATH
ENV PATH="/app/.venv/bin:$PATH"

# Sync the project's dependencies and install the project
RUN --mount=type=cache,target=/home/$DOCKER_USER/.cache/uv,uid=$PUID,gid=$PGID \
    uv sync --frozen

USER root

run sed -i '/set -e/a nvidia-smi' /entrypoint.sh
run sed -i '/set -e/a chmod 777 /dev/dri/*' /entrypoint.sh

# Pass custom command to entrypoint script provided by the base image
ENTRYPOINT ["/entrypoint.sh", ".venv/bin/python", "app/main.py"]

docker-compose.yml

services:
  app:
    build:
      context: .
    ports:
      - 5910:5910
    volumes:
      - swayvnc-wayvnc-certs:/certs
    environment:
      - RENDER_GROUP_GID=107 # replace with GID of the group which owns the /dev/dri/renderD128 device
      - SWAY_RESOLUTION=1920x1080
      - WAYVNC_PORT=5910
      - WAYVNC_ENABLE_AUTH=true
      - WAYVNC_USERNAME=wayvnc
      - WAYVNC_PASSWORD=wayvnc
      - NVIDIA_VISIBLE_DEVICES=all
    privileged: true
    runtime: nvidia
volumes:
  swayvnc-wayvnc-certs:

[dmarts05]

I’m experiencing the same issue and tested it on Windows using Docker Desktop. Our graphics card is an RTX 3060. I was about to open a new issue, but I wanted to reach out here instead.

Were you able to solve the problem?

Update: stat /dev/dri/card1 -c "%G %g"

# ls
card0  renderD128
# stat /dev/dri/card0 -c "%G %g"
root 0

I tried to fix it but couldn't make it work. I eventually noticed that I couldn't use Docker for my specific app so I stopped searching for a solution. Good luck!

[stephanlensky]

Hi all, @ewgdg added some instructions in #7 which may help for Nvidia.

Let me know if you have any success with this, or if anything in the newest version of the README is unclear!

[gokanil]

Hello @stephanlensky and @ewgdg ,

I’ve completed my tests with the updated setup. Initially, I encountered the error:
"Error response from daemon: error gathering device information while adding custom device '/dev/kfd': no such file or directory",
so I modified it to #- /dev/kfd.

I also added the --ignore-gpu-blocklist parameter, which allowed some features to be set as Hardware accelerated. I tested both X11 and XWayland configurations. However, I still cannot see my RTX 3060 GPU listed on the chrome://gpu page — it's still operating using the CPU.

I would appreciate your help with resolving this issue.

[ewgdg]

you do need to remove the whole devices section, that is only for AMD GPUs.

your observation is correct, I also noticed chrome report Mesa GPU instead of Nvidia GPU inside the container, and I guess that Mesa GPU is the integrated GPU from my AMD CPU. Also, running nvidia-smi inside the container shows no running program, despite chrome is running.

However I do find a workaround to make Chrome to report Nvidia GPU being used.

1. enable xwayland by following the readme
2. add 2 environment variables to the nvidia compose file, I found them from Arch wiki.

   - GBM_BACKEND=nvidia-drm
   - __GLX_VENDOR_LIBRARY_NAME=nvidia
   

3. add more capabilities to the GPU. Either use environment varariable NVIDIA_DRIVER_CAPABILITIES=all or modify the deploy section in the nvidia compose file:

   capabilities: [gpu, compute, utility, video, display]
   

after that nvidia-smi successfully shows 1 running process, and Chrome reports Nvidia GPU.

[ewgdg]

if you still unable to make it work, that is probably bc wayland does not yet have mature support for Nvidia GPUs, in that case you might be able to find an x11-based VNC as a base image for you.

[ewgdg]

if your goal is to improve the browser performance you might also need to increase the shm size. e.g. shm_size: "2gb".