Cross-arch building¶

tailor targets amd64 by default. Build arm64 by declaring it — nothing is inferred from the host, so a workspace builds the same set on any machine.

Build one arm64 image¶

Add an arch axis to the image:

# image.yaml
name: gizmo
matrix:
  arch: [arm64]
base:
  oci:
    uri: "registry.example/gizmo/base:edge"
    platform: "linux/${arch}"      # ${arch} → linux/arm64
outputs:
  - format: cosi
config:
  os:
    hostname: gizmo

tailor build gizmo          # → gizmo_arm64_cosi.cosi

Build both arches¶

List both values; tailor expands one cell per arch:

matrix:
  arch: [amd64, arm64]       # → gizmo_amd64_cosi, gizmo_arm64_cosi
base:
  path: ./bases/gizmo-${arch}.img

Amd64 is the default¶

With no arch axis and no workspace override, an image builds a single amd64 cell:

name: gizmo
base:
  path: ./bases/gizmo.img      # → gizmo_amd64_cosi

Give a local base its own arch¶

A local path base can declare its own arch, which supplies the cell arch when there is no arch axis:

# image.yaml
base:
  path: ./bases/gizmo-arm64.img
  arch: arm64                  # this image builds one arm64 cell

A baseImages: catalogue slot's arch works the same way for base: { ref: <name> }.

Platform must match the arch¶

The arch component of an oci.platform must equal the cell's arch. Always write linux/${arch} so each cell pulls its own manifest; a fixed platform: linux/arm64 on an amd64 cell fails at validate before any pull. path and azureLinux bases declare no arch, so the cell arch decides.

See Target architectures for the full model.