Adding support for headless opengl

[mmartial]

As discussed with @Tremeschin tonight (thank you)

To support hardware OpenGL, "otherwise it'll use llvmpipe for rendering shaders (cpu rendering)"

# Add libEGL ICD loaders and libraries
RUN apt install -y libglvnd0 libglvnd-dev libegl1-mesa-dev && \
    mkdir -p /usr/share/glvnd/egl_vendor.d && \
    echo '{"file_format_version":"1.0.0","ICD":{"library_path":"libEGL_nvidia.so.0"}}' > \
    /usr/share/glvnd/egl_vendor.d/10_nvidia.json

# Add Vulkan ICD loaders and libraries
RUN apt install -y libvulkan1 libvulkan-dev && \
    mkdir -p /usr/share/vulkan/icd.d && \
    echo '{"file_format_version":"1.0.0","ICD":{"library_path":"libGLX_nvidia.so.0","api_version":"1.3"}}' > \
    /usr/share/vulkan/icd.d/nvidia_icd.json

• For Windows use add -v /usr/lib/wsl:/usr/lib/wsl "to 'pass through' the nvidia drivers at least related to opengl"

[Tremeschin]

Minimal reproducible examples and explanation here! 🙂

OpenGL + EGL

Projects like Manim or DepthFlow uses ModernGL for rendering shaders, a task GPUs excels at. CPU rendering (aka software rendering) is known to be very slow compared to it, yielding seconds per frame performance instead of FPS!

We need both libglvnd and egl to find the GPU as a opengl capable device both natively and in containers. The two files below are a minimal example of the packages/configuration required for getting proper support in containers (non-trivial):

# docker-egl.py
import moderngl

ctx = moderngl.create_standalone_context(backend="egl")

print(ctx.info['GL_RENDERER'])

# docker-egl.dockerfile
FROM python:3.12-slim-bookworm
ENV DEBIAN_FRONTEND="noninteractive"
RUN apt update

ENV LD_LIBRARY_PATH="/usr/lib/wsl/lib"
ENV NVIDIA_DRIVER_CAPABILITIES="all"
ENV NVIDIA_VISIBLE_DEVICES="all"

# These two commands below can be combined as seen above,
# splitting for testing purposes in removing either (see below)
RUN apt install -y libglvnd0 libglvnd-dev libegl1-mesa-dev
RUN mkdir -p /usr/share/glvnd/egl_vendor.d && \
    echo '{"file_format_version":"1.0.0","ICD":{"library_path":"libEGL_nvidia.so.0"}}' > \
    /usr/share/glvnd/egl_vendor.d/10_nvidia.json

RUN pip install moderngl
COPY docker-egl.py /docker-egl.py
CMD ["python", "/docker-egl.py"]

Ensure you have the NVIDIA Container Toolkit installed on the host if Linux (never in the container), and then run it with:

docker build -t docker-egl -f docker-egl.dockerfile . && docker run --rm --gpus all docker-egl

All of the packages, env vars and ICD loaders are 100% required as well as --gpus all flag

Removing any of those (eg. commenting the RUN mkdir command) will make the GPU invisible to EGL, and will fallback to llvmpipe (software rendering). This is kind of the "secret" of nvidia/glvnd images, see how we started in a minimal image!

As for Docker Compose and/or Podman, the following/similar is needed in docker-compose.yml:

services:
  image-name:
    
    # (...)

    # Windows WSL mapping
    volumes:
      - /usr/lib/wsl:/usr/lib/wsl

    # Docker "--gpus all"
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              capabilities: [gpu]
              count: all

    # Podman "--gpus all"
    devices:
      - nvidia.com/gpu=all

Windows users might need to go to Settings > Resources > WSL Integration > Default distro if they are using Docker Desktop, could be outdated device or not need, I'm not on Windows often :^). IIRC the libEGL_nvidia.so.0 files are available in WSL at /usr/lib/wsl, so the mapping is needed.

Vulkan

The Vulkan libraries and ICD loaders follows a similar story. A good reason to support it is that some projects like ollama may have support for Vulkan compute backend in the future (broader support). Some projects, mainly upscalers, like upscayl, realesrgan depends 100% on it, and can be part of a pipeline of image processing in ComfyUI!

All of this (no surprises in opengl + vulkan working) shouldn't be more than ~100 MB in size

Happy to contribute, only took 100h of sanity across a few weeks to figure out last year 😅

[mmartial]

Also looking at https://github.com/ComfyDock/ComfyDock-Docker/blob/feature/docker-non-root/Dockerfile

[mmartial]

Test in 202503wip branch

[mmartial]

Current test build out mmartial/comfyui-nvidia-docker:latest-beta

[mmartial]

Confirmed as functional

Tested: docker run --rm -it --gpus all mmartial/comfyui-nvidia-docker:latest-beta /bin/bash

In container:

sudo su -l comfy
python3 -m venv venv
source venv/bin/activate
pip3 install moderngl
[...]
Successfully installed glcontext-3.0.0 moderngl-5.12.0
python3
>>> import moderngl
>>> ctx = moderngl.create_standalone_context(backend="egl")
>>> print(ctx.info['GL_RENDERER'])
NVIDIA GeForce RTX 3090/PCIe/SSE2

[mmartial]

Awaiting for confirmation that other new features added to container (FORCE_CHOWN) are functional before extending README.md and making a new release

[mmartial]

Closing with release 20250320