[Bug]: CDI spec missing core driver libraries (libnvidia-eglcore/glcore/glsi/tls) on Debian-based systems — getVersionLibs() only searches libcuda.so's subdirectory

[silent-zgh]

Describe the bug
On Debian-based systems, nvidia-ctk cdi generate (and jit-cdi at container startup) produces an incomplete CDI specification that omits 4 core driver libraries: libnvidia-eglcore, libnvidia-glcore, libnvidia-glsi, and libnvidia-tls. This causes EGL initialization to fail silently at runtime, falling back to Mesa software rendering without any error. nvidia-smi shows zero GPU utilization for affected containers.
To Reproduce

# On any Debian/Ubuntu/Deepin host with NVIDIA .deb driver installed
nvidia-ctk cdi generate --output /tmp/test-cdi.yaml

# Verify core dependencies are missing
grep "containerPath" /tmp/test-cdi.yaml | grep -E "eglcore|glcore|glsi|tls"
# → empty output

# Verify the files exist on the host
ls -1 /usr/lib/x86_64-linux-gnu/libnvidia-eglcore* \
      /usr/lib/x86_64-linux-gnu/libnvidia-glcore* \
      /usr/lib/x86_64-linux-gnu/libnvidia-glsi* \
      /usr/lib/x86_64-linux-gnu/libnvidia-tls*
# → all 4 files present, but not included in the CDI spec

Screenshots

Expected behavior

All driver libraries matching *.so.<driver_version> should be included in the CDI spec regardless of whether they reside in the same subdirectory as libcuda.so. EGL/GLX initialization should succeed and use GPU rendering.
Environment

• nvidia-container-toolkit version: 1.18.2

• NVIDIA Driver Version: 535.161.10 (installed via .deb)

• Host OS: UOS Desktop 20 Professional (Debian-based, derivative of Deepin)

• Kernel Version: 6.1.32-amd64-desktop-hwe

• Container Runtime Version: containerd (bundled with k3s v1.34.3+k3s1)

• CPU Architecture: x86_64

• GPU Model: NVIDIA Graphics Device (RTX series)

• RPM-based systems (RHEL, Fedora) are not affected — all driver libraries are flat in /usr/lib64/, so filepath.Dir(libcuda) captures everything.

• Workaround: container init script manually symlinks missing libraries, but this should not be necessary.

• Related: CDI Spec Generation is missing driver libraries #1215 (broader class of CDI spec missing driver libraries), Missed libraries in container compared to what GPU driver installs kubernetes-sigs/dra-driver-nvidia-gpu#446.

[cdesiniotis]

Hi @silent-zgh this appears to be a duplicate of #1559 which should be fixed by #1820 -- this was recently merged into main and will get included in our next release.

If you want to verify that top of main resolves this issue, you can follow the instructions from #1599 (comment) for extracting packages built by our GitHub Actions and installing them. To pull the packages you can use:

./hack/pull-packages.sh \
    ghcr.io/nvidia/container-toolkit:69c285d7-packaging \
    test-package-69c285d7

This will create a folder named test-package-69c285d7 including the rpm and deb packages. 69c285d7 is the current HEAD of main.

[silent-zgh]

@cdesiniotis thinks for your help。I will pull the packages。