install: add nvidia legacy

.github/workflows/docker.yml

@@ -21,6 +21,7 @@ jobs:
       matrix:
         BASE: [
           ["noble-nvidia", ".s6", "noble-nvidia", "nvidia"],
+          ["noble-nvidia-legacy", ".s6", "noble-nvidia-legacy", "nvidia-legacy"],
           ["noble-intel", ".s6", "noble-intel", "intel"],
           ["noble-amd", ".s6", "noble-amd", "amd"],
           ["noble-full", ".s6", "noble-full", "full"],

install/docker/Dockerfile.nvidia-legacy

@@ -0,0 +1,11 @@
+ARG BASE="ghcr.io/koush/scrypted-common:20-jammy-full"
+FROM $BASE
+
+ENV SCRYPTED_DOCKER_FLAVOR="nvidia"
+
+ENV NVIDIA_DRIVER_CAPABILITIES=all
+ENV NVIDIA_VISIBLE_DEVICES=all
+
+# nvidia cudnn/libcublas etc.
+# for some reason this is not provided by the nvidia container toolkit
+RUN curl https://raw.githubusercontent.com/koush/scrypted/main/install/docker/install-nvidia-graphics-legacy.sh | bash

install/docker/install-nvidia-graphics-legacy.sh

@@ -0,0 +1,54 @@
+if [ "$(uname -m)" = "x86_64" ]
+then
+    UBUNTU_22_04=$(lsb_release -r | grep "22.04")
+    UBUNTU_24_04=$(lsb_release -r | grep "24.04")
+
+    # needs either ubuntu 22.0.4 or 24.04
+    if [ -z "$UBUNTU_22_04" ] && [ -z "$UBUNTU_24_04" ]
+    then
+        echo "NVIDIA graphics package can not be installed. Ubuntu version could not be detected when checking lsb-release and /etc/os-release."
+        exit 1
+    fi
+
+    if [ -n "$UBUNTU_22_04" ]
+    then
+        distro="ubuntu2204"
+    else
+        distro="ubuntu2404"
+    fi
+
+    echo "Installing NVIDIA graphics packages."
+    apt update -q \
+        && apt install -y wget \
+        && wget -qO /cuda-keyring.deb https://developer.download.nvidia.com/compute/cuda/repos/$distro/$(uname -m)/cuda-keyring_1.1-1_all.deb \
+        && dpkg -i /cuda-keyring.deb \
+        && apt update -q \
+        && apt install -y cuda-nvcc-12-6 libcublas-12-6 libcudnn9-cuda-12 cuda-libraries-12-6;
+
+    if [ "$?" != "0" ]
+    then
+        echo "Error: NVIDIA graphics packages failed to install."
+        exit 1
+    fi
+
+
+    # Update: the libnvidia-opencl.so.1 file is not present in the container image, it is
+    # mounted via the nvidia container runtime. This is why the following check is commented out.
+    # this file is present but for some reason the icd file is not created by nvidia runtime.
+    # if [ ! -f "/usr/lib/x86_64-linux-gnu/libnvidia-opencl.so.1" ]
+    # then
+    #     echo "Error: NVIDIA OpenCL library not found."
+    #     exit 1
+    # fi
+
+    # the container runtime doesn't mount this file for some reason. seems to be a bug.
+    # https://github.com/NVIDIA/nvidia-container-toolkit/issues/682
+    # but the contents are simply the .so file, which is a symlink the nvidia runtime
+    # will mount in.
+    mkdir -p /etc/OpenCL/vendors/
+    echo "libnvidia-opencl.so.1" > /etc/OpenCL/vendors/nvidia.icd
+else
+    echo "NVIDIA graphics will not be installed on this architecture."
+fi
+
+exit 0