NVIDIA Variant Architecture#

The dakota-nvidia variant is built by composing a hardware-specific NVIDIA OCI layer on top of the same base oci/bluefin.bst used for the default variant — the two variants share identical GNOME/system sources. NVIDIA-specific packages are never mixed into the base; they live in an isolated parallel element tree that merges cleanly at a single stack element.

The top-level assembly element is elements/oci/bluefin-nvidia.bst, which produces ghcr.io/projectbluefin/dakota-nvidia:latest. Both dakota and dakota-nvidia are released together under :testing → :stable tags via the same promotion pipeline .

Layer Dependency Tree#

The NVIDIA variant's BuildStream element graph mirrors the base layer hierarchy, merging at bluefin-nvidia-stack.bst:

Base branch:
  bluefin-stack.bst
      └── bluefin-runtime.bst
              └── layers/bluefin.bst
                      └── oci/bluefin.bst ← default variant image + parent for NVIDIA

NVIDIA branch:
  bluefin-stack.bst ←── shared
  bluefin-nvidia/deps.bst ←── NVIDIA packages
      │
      └── bluefin-nvidia-stack.bst ← MERGE POINT (also runs depmod)
              └── bluefin-nvidia-runtime.bst
                      └── layers/bluefin-nvidia.bst
                              └── oci/bluefin-nvidia.bst ← NVIDIA variant image

bluefin-nvidia-stack.bst is the sole merge point: it depends on bluefin-stack.bst (the shared base) and adds bluefin-nvidia/deps.bst. It also runs depmod at integration time so the NVIDIA kernel modules are registered alongside the base system's modules .

NVIDIA Component Elements#

All NVIDIA-specific packages are aggregated in elements/bluefin-nvidia/deps.bst :

Element	Purpose
`nvidia-drivers.bst`	Proprietary driver userspace + kernel modules (Turing+ GPUs)
`egl-external-platform.bst`	EGL platform plugin headers
`nvidia-egl-wayland.bst`	Direct EGL/GBM Wayland support for GNOME Shell
`nvidia-kargs.bst`	Kernel arguments: blacklist `nouveau`, enable KMS/framebuffer device
`nvidia-modprobe-config.bst`	`modprobe.d` / `modules-load.d` config for correct driver load order
`nvidia-container-toolkit.bst`	`nvidia-ctk` + `nvidia-cdi-hook` for GPU passthrough in containers
`nvidia-container-toolkit-preset.bst`	Systemd preset enabling the CDI device refresh service

OCI Image Assembly#

oci/bluefin-nvidia.bst calls build-oci with two inputs:

parent: /parent — the full base oci/bluefin.bst image
layer: /layer — the NVIDIA-only compose layer oci/layers/bluefin-nvidia.bst

This produces a derived OCI image where the NVIDIA layer is appended to the base — not a fork. The element also assigns a distinct sysroot-seed UUID (7f9c5d11-...) so deployed NVIDIA and default systems never share machine IDs or partition UUIDs.

Pre-assembly steps run against /layer to prepare the NVIDIA rootfs: prepare-image.sh, systemd-sysusers, glib-compile-schemas, dconf update, and a usr/etc → etc migration required by bootc .

CI Build Matrix#

build.yml defines a two-entry matrix for default and nvidia variants :

Variant	Element	Image suffix	`continue-on-error`
`default`	`oci/bluefin.bst`	(none)	`false`
`nvidia`	`oci/bluefin-nvidia.bst`	`-nvidia`	`true`

Builds run serially (max-parallel: 1) due to the shared remote CAS at cache.projectbluefin.io being rate-limited to a single concurrent build . The nvidia variant has continue: true so a driver build failure doesn't block the base image release. After each build, artifacts are pushed to the remote CAS explicitly with bst artifact push --deps run .

ISO Packaging Note#

The dakota-iso build uses dakota-nvidia:stable as its payload — the live ISO boots the NVIDIA variant and auto-rebases to the non-NVIDIA image on first upgrade if NVIDIA hardware is not detected . This means dakota-nvidia is the universal install medium.