glvnd container with cuda

[Besler]

I'm running a torch model through trame in a docker container. Could a glvnd-cuda tag be provided through the nvidia/cuda:11.2.0-cudnn8-runtime-ubuntu20.04 base image? Alternatively, I could install the cuda/cudnn dependencies over the glvnd layer or manage the trame image myself?

[jourdain]

If you can install the cuda/cudnn dependencies over the glvnd that would be great.

If not, we might be able to add yet another image but if we can prevent that, it would be great... I don't want to blow the matrix of builds...

Out of curiosity, are you using the rendering capability of the image with VTK or so, or not at all?

[Besler]

I've always had a hard time installing cuda directly, so I switched up the root build image and built the trame image myself.

echo "Building ${TRAME_DOCKER_TAG}"
docker build \
    --progress=plain \
    --build-arg BASE_IMAGE=nvidia/cuda:11.2.0-cudnn8-runtime-ubuntu20.04 \
    -t ondiag/trame:common \
    ${TRAME_DIR}/docker \
    -f "${TRAME_DIR}/docker/Dockerfile.common"
docker build \
    --progress=plain \
    --build-arg BASE_IMAGE=ondiag/trame:common \
    -t "${TRAME_DOCKER_TAG}" \
    ${TRAME_DIR}/docker \
    -f "${TRAME_DIR}/docker/Dockerfile.pip"

${TRAME_DIR} points to a trame git submodule. Then I use ${TRAME_DOCKER_TAG} as the base image. I render with VTK OSMESA. So far runs together (can infer on GPU, can render with VTK). I haven't tried EGL.

[jourdain]

Thanks for sharing your command lines and it does not seems to bad doing it yourself... ;-)