whisper.cpp-cuda

Build times (auto-updated)

• x86_64: 5m6s (2026-09-22 19:41 UTC, MAKEFLAGS=-j6)

---

• Vendored under adapted/ 2026-09-22 (via scripts/add_aur_package.sh adapted whisper.cpp-cuda) after being found built-but-untracked on eurobuild6 during an archived-build-tree cleanup. Not maintained by us upstream – adapted, not maintained/.
• Depends on ggml-cuda-git (see its memory file) via -DWHISPER_USE_SYSTEM_GGML=ON in the PKGBUILD’s build() – links against whatever ggml is currently published, not a bundled copy.

2026-09-22: build FAILS against current ggml-cuda-git – not fixed this pass

ggml-cuda-git was built+published fresh the same day (live git HEAD, as always for a -git package). whisper.cpp-cuda 1.9.1’s build then failed:

examples/talk-llama/llama-model-loader.cpp:992:42: error: too few arguments
to function 'ggml_tensor* ggml_ssm_scan(ggml_context*, ggml_tensor*,
ggml_tensor*, ggml_tensor*, ggml_tensor*, ggml_tensor*, ggml_tensor*,
ggml_tensor*, int64_t)'

Root cause: upstream ggml added a parameter to ggml_ssm_scan() since whisper.cpp 1.9.1 was cut. examples/talk-llama/ is whisper.cpp’s own vendored copy of llama.cpp model-loading code, not maintained in lockstep with ggml’s live API – a real upstream incompatibility between two independently-versioned rolling/pinned projects, not a packaging mistake.

talk-llama can’t be disabled without also losing stream/command/lsp – examples/CMakeLists.txt gates all four together under one if (WHISPER_SDL2) block, no separate per-example toggle.

Not fixed as part of this pass. Options identified but not chosen yet (needs a decision, this is non-routine):

1. Patch examples/CMakeLists.txt to split talk-llama into its own (default-off) guard, keeping stream/command/lsp working.
2. Patch the vendored llama-model-loader.cpp call site to add the missing argument, matching current ggml_ssm_scan() – would need re-patching whenever whisper.cpp’s vendored copy or ggml’s signature drifts again.
3. Report upstream, leave unpublished until whisper.cpp’s own next release catches up to the ggml API (likely the more sustainable fix, given this is exactly the kind of churn a -git-tracked dependency causes for a pinned-version consumer).

Nothing published this session for this package. from_eurobuild6/whisper.cpp-cuda (the archived build that motivated vendoring this) should stay until a working local build actually supersedes it.

2026-09-22 (2): fixed – option 1, published

Went with option 1: added disable-talk-llama.patch (new WHISPER_BUILD_TALK_LLAMA CMake option, defaults ON upstream, gates just the add_subdirectory(talk-llama) call so stream/command/lsp are unaffected) and passed -DWHISPER_BUILD_TALK_LLAMA=OFF in build(). pkgrel 2 -> 3. Built and published cleanly, x86_64: 1.9.1-3. Confirmed indexed in the repo db, not just uploaded as a file.

Chose this over patching the vendored llama-model-loader.cpp call site (would need re-patching on every future ggml/whisper.cpp drift) and over waiting on upstream (no telling how long whisper.cpp’s next release takes) – losing talk-llama specifically is an acceptable tradeoff, the core whisper transcription functionality and the other SDL2 examples are intact.

from_eurobuild6/whisper.cpp-cuda is now genuinely superseded – deleted.

2026-09-24: upstream AUR package appears delisted

check_aur_feedback.sh now flags this MISSING (AUR RPC info returns 0 results, and https://aur.archlinux.org/packages/whisper.cpp-cuda 404s). All commits here are from the real upstream maintainers (Bink, robertfoster, Gianluca Boiano – never us, this is adapted/, not pushed to by this tooling), so this isn’t our doing. The git backend itself is still reachable though: git ls-remote origin (authenticated SSH) still resolves refs/heads/master at 7c32310 (the same commit we have locally) – so the repo hasn’t been deleted outright, just delisted from AUR’s public search/RPC/web listing (same phenomenon already documented in memory/android-file-transfer.md for a different, long-dead package name). Practical effect: no further upstream updates will ever come from this remote. Our local copy (pinned at 1.9.1-3 with the disable-talk-llama.patch fix) is now effectively frozen/orphaned upstream – still fine to keep building and publishing from as-is, just won’t get new pkgver bumps for free anymore via a plain git pull.

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.