Snap: llama-cpp
Requested tracks: rocm, cuda
Rationale: llama-cpp ships GPU-backend variants of the same upstream codebase. Downstream need to pull a specific backend as a stage-snaps/build-snaps, which requires the variant to be addressable as llama-cpp//. Components were considered but aren’t consumable by stage-snaps. Separate snaps were considered but fragment the namespace for what’s one upstream project.
Cadence: each track receives a new revision per upstream b#### tag we package, with the corresponding backend toolchain. The latest track will continue to track CPU-only builds.