- Extend pkgs/llama-cpp to support CUDA backend (backend="cuda") for NVIDIA GPUs
- Add llama-cpp-cuda variant to modules/core/default.nix overlay
- Create hosts/hl-bigbox1/services/llama-cpp.nix with RTX 3060 tuning:
* gpuLayers=30 (conserve 12GB VRAM for context)
* parallel=2 (handle concurrent requests on smaller VRAM)
* Same Qwen3.8-27B + DFlash2 drafter as fwdesk
- Wire service into bigbox1 services/default.nix
- Disabled by default (weights download manual)
v0.8.0 is the first upstream release with DFlash2 drafters; nixpkgs has 0.6.0 (unstable) and an older tree (stable), so the engine is built here. HIP build for Strix Halo. The NixOS module exposes llama-server behind skyg.nixos.server.services.llama-cpp; fwdesk wires it to Qwen3.8-27B + its DFlash2 drafter, disabled until the weights are on disk.
The patchelf package ships a setup hook that runs 'patchelf --shrink-rpath' on every ELF during fixup, which rewrites DT_RPATH as DT_RUNPATH. RUNPATH is not inherited by the bundled libs, so libXau.so.6 and friends were not found. Set dontPatchELF to keep the RPATH, and set the interpreter to the NixOS glibc instead of relying on /lib64.
Delta ships a self-contained tarball with no stable download URL (the release API returns a short-lived signed R2 link), so src is a fixed-output derivation that resolves the API at build time. Patch the interpreter/RPATH, add the dlopen'd graphics libs (libglvnd, vulkan-loader, wayland) via a wrapper, and set XKB_CONFIG_ROOT for the bundled libxkbcommon.
Registered in the core/lib overlays, exposed as a flake package, and installed on fwbook.