ggml-cuda-git

• Vendored under adapted/ 2026-09-22 (via scripts/add_aur_package.sh adapted ggml-cuda-git) after being found built-but-untracked on eurobuild6 during an archived-build-tree cleanup. Not maintained by us upstream – adapted, not maintained/.
• Built + published cleanly on 2026-09-22: 1:b11102.r6.f95b0d9539-1, x86_64 only (+ -debug). CUDA toolkit (/opt/cuda, cuda 13.4.2-1) already present on euronuc; build succeeded without a live GPU present (compiling CUDA code doesn’t need one, only running it would).
• aarch64 (also in arch=()) deliberately not attempted: none of the configured remote ARM boards (eurobuild4/5/14, see remote_arm_build_machines.md) have a CUDA toolchain. Restrict to x86_64 via REPO_BUILD_ONLY_ARCH=x86_64 for any future rebuild too, unless an ARM+CUDA board is ever added.
• whisper.cpp-cuda (see its own memory file) depends on this package and currently fails to build against it – an upstream ggml_ssm_scan() API change in this -git package’s live HEAD broke whisper.cpp-cuda’s vendored talk-llama example code. Not a bug in this package; flagging here since it’s the moving half of that incompatibility.

2026-10-03: removed from arch/ tracking a second time – CUDA+CPU-arch make it a local-only build

The 2026-09-22 build (1:b11102.r6.f95b0d9539-1) was compiled with GGML_NATIVE=ON (the PKGBUILD’s default branch, toggled by aur_ggml_build_universal), baking in the build host’s (euronuc) exact CPU features – this is what caused whisper-server’s ggml_cpu_init/ggml_backend_cpu_reg SIGILL-class crash on eurobuild6 (Ryzen 7 2700X, lacks euronuc’s AVX-512), same root-cause family as the ladybird -march=native saga. The PKGBUILD already has the correct fix available (aur_ggml_build_universal=true -> GGML_NATIVE=OFF + GGML_CPU_ALL_VARIANTS=ON + GGML_BACKEND_DL=ON, portable runtime CPU dispatch) – but getting that env var through repo_build.sh’s sudo --preserve-env into the chroot has no existing support (only MAKEFLAGS/MAKEPKG_CONF are threaded through today).

User’s call 2026-10-03, before that plumbing question got resolved: don’t bother – this package (CUDA + a CPU-arch-sensitive runtime dispatch choice) is “so special” it should just be built and installed locally per-machine, not published generically through this project’s shared repo. Re-removed adapted/ggml-cuda-git (had been restored earlier the same session) – git submodule deinit + removed the stray uncommitted .gitmodules addition, confirmed back to a fully clean state (as if never restored). The already-published 1:b11102.r6.f95b0d9539-1 package is untouched on the repo (not asked to unpublish it, just to stop tracking/rebuilding it here). whisper.cpp-cuda (depends on this) removed the same way, same reasoning – see its own memory file.

2026-10-08: confirmed still out of scope

User confirmed: CUDA build now happens locally on the machine that actually uses it (GPU present there), not through this tooling at all. Came up while comparing memory/site/arch/ for gaps – correctly absent from arch/ and from scripts/install_only_packages; not a tracking gap, don’t re-add.