cuda

Recipe card from the charly-distros plugin (Images — the deployable catalog).

cuda – CUDA development toolkit

CUDA compiler, cuDNN, ONNX Runtime, and GPU development libraries. Depends on the nvidia candy for runtime support (driver libs, nvidia-ctk).

Candy Properties

Property	Value
Install files	task:, charly.yml
Depends	nvidia, ffmpeg

Environment Variables

Variable	Value
CUDA_HOME	/usr

Packages

RPM: cuda-nvcc, cuda-cudart-devel, cuda-cudart-static, cuda-nvrtc-devel, cuda-cupti-devel, cccl-devel, cuda-cudnn, libcurand-devel, libcufile-devel, onnxruntime, libaio-devel, cpio

PAC: cuda, cudnn, python-onnxruntime-cpu

Note: FFmpeg codec libraries are provided via the ffmpeg dependency candy rather than installed directly. On Fedora, CUDA packages come from the negativo17 fedora-nvidia repo (added by the nvidia candy); on Arch/CachyOS they come from the standard repos.

Multi-distro (Fedora + Arch)

The candy is multi-distro. The distro.arch section installs cuda, cudnn, and python-onnxruntime-cpu from the Arch repos. Arch installs CUDA under /opt/cuda — its toolkit headers under /opt/cuda/targets/x86_64-linux/include/ — while CUDA_HOME is /usr, so the Arch path links every entry of those trees into /usr — files of any extension and subtrees alike, enumerated rather than pattern-matched. nvcc computes its include root relative to its own executable, and /usr/bin/nvcc is a symlink, so anything not reachable under /usr/include cannot be found: cuda_runtime.h’s own #include fails and every CUDA build dies at CMake’s enable_language(CUDA) with “No such file or directory”. Matching on patterns produced both of those failures, on both Arch arms, while every probe in this plan still passed (opencharly/layer-cuda#9): first a *.h glob that linked the top-level headers but not the crt/ and cccl/ subtrees, then the same glob missing the .hpp half of the _functions family once the subtrees were linked. The check cuda-arch-header-root preprocesses a real CUDA translation unit to prove the whole header root resolves.

With that stitch in place, the same nvcc, header, and library paths resolve identically on Fedora and on Arch/CachyOS.

Install tasks

Extracts cuDNN headers from cuda-cudnn-devel RPM (bypasses driver dependency via rpm2cpio) on Fedora; the /opt/cuda → /usr symlink stitch runs on Arch/CachyOS.

Usage

charly.yml

nvidia:
  base: fedora-nonfree
  candy:
    - nvidia
    - cuda

Used In Boxes

• /charly-distros:nvidia (base for all GPU boxes)
• /charly-immich:immich-ml

Related Candies

• /charly-distros:nvidia — NVIDIA GPU runtime (driver libs, CDI toolkit) — required dependency
• /charly-selkies:ffmpeg — FFmpeg multimedia (nonfree codecs) — required dependency
• /charly-distros:rocm — AMD GPU counterpart (ROCm runtime + OpenCL)
• /charly-languages:python-ml — ML Python environment (depends on cuda)
• /charly-jupyter:jupyter — Jupyter notebooks (depends on cuda)
• /charly-ollama:ollama — LLM server (depends on cuda)
• /charly-comfyui:comfyui — image generation (depends on cuda)

Related

• /charly-image:layer — candy authoring reference (charly.yml schema, task verbs, service declarations)
• /charly-check:check — declarative testing (check: block, charly check box, charly check live)