Build fails on aarch64 due to virtual cuDNN package: libcudnn9 needs explicit provider (libcudnn9-cuda-12)

While building isaac_ros_common for Jetson Orin Nano (JetPack r36.4, Ubuntu 22.04, base nvcr.io/nvidia/12.6.11-devel-aarch64-ubuntu22.04) using ./build_images_layser.sh -i aarch64, the Docker build fails at the CUDA install step with “E: Package 'libcudnn9' has no installation candidate.”

In the current Dockerfile.aarch64, libcudnn9 is specified, but on Ubuntu 22.04 with the NVIDIA repos this is a virtual package provided by libcudnn9-cuda-12 and libcudnn9-cuda-13.

Because the image installs CUDA 12.6 (cuda-toolkit-12-6, etc.), apt requires the explicit provider libcudnn9-cuda-12.

Replacing libcudnn9 with libcudnn9-cuda-12 resolves the build immediately; optionally pinning to 9.12.0.46-1 improves reproducibility.

It may also be worth removing the deprecated --force-yes in the same line.

Could the Dockerfile be updated to use libcudnn9-cuda-12 (or a conditional/provider-agnostic approach based on the selected CUDA version) to prevent build breaks on aarch64?

[shawnhanna]

Thanks for posting this fix! This bug has been killing my build system for a while now w/ this error (so others can search)

 - UndefinedVar: Usage of undefined variable '$CMAKE_PREFIX_PATH' (line 277)
Dockerfile.aarch64:312
--------------------
 311 |     # Install CUDA packages
 312 | >>> RUN --mount=type=cache,target=/var/cache/apt \
 313 | >>>     apt-get update && apt-get install -y --no-install-recommends --force-yes \
 314 | >>>     cuda-cudart-12-6 \
 315 | >>>     cuda-libraries-12-6 \
 316 | >>>     cuda-nvml-dev-12-6 \
 317 | >>>     cuda-sanitizer-12-6 \
 318 | >>>     cuda-toolkit-12-6 \
 319 | >>>     libcublas-12-6 \
 320 | >>>     libcudnn9 \
 321 | >>>     libcusparse-12-6 \
 322 | >>>     libnpp-12-6
 323 |     
--------------------