Aaron Jackson

You can build and publish a complete signed Llama Manager appliance release (.deb packages, APT repo, ISO) with atomic publish and download-site verification, and separately build or update the ROCm llama.cpp llama-server engine for the AMD Strix Halo iGPU, including new model architecture support.

They target the maintainer and release engineers of the Llama Manager appliance—developers comfortable with Debian packaging, distrobox containers, ROCm builds, and rsync-based deployment who need to cut releases or bump the inference engine version.

The unattended release runner builds signed packages and the ISO from main, stages pinned inputs from the asset mirror, performs an atomic publish, then runs a two-pass rsync to the thromgar host and verifies public URLs. A --no-sync local loop with serve-local.sh supports iteration.

Use build-llama-cpp to update or rebuild the ROCm llama.cpp llama-server engine itself—version bumps, new model architectures, or engine-level build fixes. Use build-and-deploy-release only to cut and publish a full appliance release; it does not rebuild llama.cpp.

The llama.cpp build requires a pinned checkout and a distrobox ROCm build environment on AMD Strix Halo hardware. The release flow stages pinned inputs from an asset mirror; signing may run unattended or require a human depending on the step, per the skill's documented gotchas.