fix(void): report pywhispercpp 1.5.0 and load CPU variants at runtime (#49)

Fixes #43
Fixes #44

packaging/void/README.md

@@ -88,19 +88,35 @@ for the sdist; `tests/test_void_pywhispercpp_template.py` fails otherwise.
   There is no CUDA build: Void ships no CUDA toolkit package. Rebuilding
   locally with a hand-installed toolkit (`GGML_CUDA=1`) is a documented,
   unsupported escape hatch.
-- **Portable CPU flags**: every ggml SIMD option is pinned off
-  (`-DGGML_NATIVE=OFF -DGGML_SSE42=OFF ...`), the plain x86-64 baseline
-  already used by the Flatpak (`#938`, `#957`). The install smoke scans the
-  shipped `libggml-cpu` (`.GCC.command.line`) for `-march=native`,
-  `-mcpu=native` and `-mtune=native` and fails on a match. The scan is a
-  tripwire, not a proof: the proof is the pinned configuration.
+- **CPU variants (`#44`)**: `GGML_BACKEND_DL=ON` and
+  `GGML_CPU_ALL_VARIANTS=ON` build one CPU library per instruction-set
+  level (`libggml-cpu-x64.so` for the plain x86-64 baseline, then
+  `haswell`, `skylakex`, `zen4`, ...). ggml loads the best one the host
+  supports, so a host with AVX2 runs AVX2 code, and the baseline variant
+  ships everywhere. `GGML_NATIVE` stays `OFF`: ggml refuses it with
+  `GGML_BACKEND_DL`, and it keeps `-march=native` out of every variant
+  (`#938`). `patches/load-ggml-backends-beside-module.patch` makes the
+  extension load the variants from its own directory: ggml's default search
+  covers the interpreter's directory and the working directory, and neither
+  holds them, so without the patch no CPU backend is registered. The install
+  smoke checks that `libggml-cpu-x64.so` holds no AVX (`ymm`/`zmm`) code,
+  that `libggml-cpu-haswell.so` does, that no variant carries host-native
+  tuning (`.GCC.command.line`), and that the CPU and Vulkan backends load
+  from the installed package.
+- **Version (`#43`)**: `setup.py` passes no version to `setup()`. The
+  version comes from setuptools-scm, which the sdist declares as a build
+  requirement, so `hostmakedepends` provides `python3-setuptools_scm` and
+  `SETUPTOOLS_SCM_PRETEND_VERSION` pins `1.5.0` without git metadata. Without
+  setuptools-scm, setuptools writes `0.0.0`, which the app reads as an
+  unknown binding version and then skips device selection. The install smoke
+  asserts the installed package reports `1.5.0`.
 - **Shared libraries**: `BUILD_SHARED_LIBS=ON` with an `$ORIGIN` rpath
   (as `install.sh` does), so the extension resolves the whisper.cpp
   libraries installed beside it with no `LD_LIBRARY_PATH`.
 - **Offline build**: the sdist vendors whisper.cpp, pybind11, and the
   ggml sources; the build fetches nothing beyond the declared distfiles.
-  `PYWHISPERCPP_VERSION` is injected because the sdist carries no
-  `version.txt` (its version otherwise comes from git metadata).
+  The sdist carries no git metadata, which is why the version is pinned
+  through setuptools-scm (see Version above).
 
 ## voculinux
 
@@ -202,8 +218,9 @@ host Void install and nothing is published. Stages, in order, each fails the run
    `python3-pysocks` are installed, then `voculinux` installs and upgrades both
    to the carried versions (see [Version floors](#version-floors)),
    `voculinux --version` shows the version, the payload checks pass (desktop
-   entry, icon, marker, no models, no pre-rename names, the Vulkan backend, no
-   host-native CPU tuning), and the imports load (`pynput` through its dummy
+   entry, icon, marker, no models, no pre-rename names, the Vulkan backend, the
+   CPU variants with an AVX-free x86-64 baseline, no host-native CPU tuning),
+   the installed pywhispercpp reports `1.5.0`, and the imports load (`pynput` through its dummy
    backend, `evdev`, `gi` with Gtk 3, `pywhispercpp.model`, `pyaudio`, `socks`,
    and the app's own modules, which keep their retained internal name).
 

packaging/void/container-release.sh

@@ -106,19 +106,46 @@ payload_voculinux() {
 	echo "payload ok: desktop entry, icon and marker present; no models, wheels or pre-rename names"
 }
 
-# whisper.cpp must ship its Vulkan backend and no host-native CPU tuning.
+# whisper.cpp ships its Vulkan backend and its CPU backend as runtime-loaded
+# variants beside the extension (GGML_BACKEND_DL, #44). Every CPU variant
+# carries its own instruction set, but the x86-64 baseline variant must hold no
+# AVX code (ymm/zmm), the AVX2 variant must hold some (so the variants are real,
+# not one baseline build), and every other shipped native library (the ggml
+# base, the registry, whisper, the extension, Vulkan) must stay AVX-free. No
+# library may carry host-native tuning.
 payload_pywhispercpp() {
 	PKGFILE=$1
 	tar -tf "$PKGFILE" | grep -q "libggml-vulkan" || fail "the Vulkan backend is missing from $PKGFILE"
-	LIBMEMBER=$(tar -tf "$PKGFILE" | grep -E "libggml-cpu\.so" | head -n1)
-	[ -n "$LIBMEMBER" ] || fail "libggml-cpu is missing from $PKGFILE"
-	mkdir -p /flagscan
-	tar -xf "$PKGFILE" -C /flagscan "$LIBMEMBER"
-	LIBSCAN="/flagscan/${LIBMEMBER#./}"
-	if readelf -p .GCC.command.line "$LIBSCAN" 2>/dev/null | grep -E "march=native|mcpu=native|mtune=native"; then
-		fail "host-native CPU tuning in the shipped libggml-cpu"
+	for VARIANT in x64 haswell; do
+		tar -tf "$PKGFILE" | grep -q "libggml-cpu-$VARIANT\.so" ||
+			fail "CPU variant libggml-cpu-$VARIANT.so is missing from $PKGFILE"
+	done
+	rm -rf /flagscan && mkdir -p /flagscan
+	tar -xf "$PKGFILE" -C /flagscan --wildcards '*.so*'
+	while IFS= read -r LIBSCAN; do
+		NAME=${LIBSCAN##*/}
+		if readelf -p .GCC.command.line "$LIBSCAN" 2>/dev/null | grep -E "march=native|mcpu=native|mtune=native"; then
+			fail "host-native CPU tuning in $NAME"
+		fi
+		case "$NAME" in
+		libggml-cpu-*) ;; # each variant is checked below
+		*)
+			if objdump -d "$LIBSCAN" | grep -Eq '%[yz]mm[0-9]'; then
+				fail "$NAME contains AVX code; only the CPU variants may"
+			fi
+			;;
+		esac
+	done <<LIBS
+$(find /flagscan -type f \( -name '*.so' -o -name '*.so.*' \))
+LIBS
+	BASELINE=$(find /flagscan -type f -name 'libggml-cpu-x64.so')
+	AVX2=$(find /flagscan -type f -name 'libggml-cpu-haswell.so')
+	[ -n "$BASELINE" ] && [ -n "$AVX2" ] || fail "the CPU variants were not extracted from $PKGFILE"
+	if objdump -d "$BASELINE" | grep -Eq '%[yz]mm[0-9]'; then
+		fail "the x86-64 baseline CPU variant contains AVX code; it would not run on every x86-64 CPU"
 	fi
-	echo "payload ok: Vulkan backend present; no host-native tuning in libggml-cpu"
+	objdump -d "$AVX2" | grep -Eq '%ymm[0-9]' || fail "the AVX2 CPU variant has no AVX2 code"
+	echo "payload ok: Vulkan backend; CPU variants present; baseline and other libraries free of AVX; no host-native tuning"
 }
 
 phase_install() {
@@ -179,6 +206,46 @@ phase_install() {
 	python3 -c "import evdev; print('evdev ok')"
 	python3 -c "import gi; gi.require_version('Gtk', '3.0'); from gi.repository import Gtk; print('gi ok: Gtk', Gtk.get_major_version(), Gtk.get_minor_version())"
 	python3 -c "from pywhispercpp.model import Model; print('pywhispercpp.model ok')"
+	echo "== pywhispercpp reports the version it is pinned to (#43)"
+	PWC_VERSION=$(python3 -c "import importlib.metadata as m; print(m.version('pywhispercpp'))") ||
+		fail "pywhispercpp has no installed version"
+	[ "$PWC_VERSION" = "1.5.0" ] || fail "pywhispercpp reports $PWC_VERSION, not 1.5.0 (#43)"
+	echo "pywhispercpp version: $PWC_VERSION"
+	echo "== the extension loads its CPU and Vulkan backends from the package (#44)"
+	python3 - <<'PY' || fail "the dynamic ggml backends did not load from the package"
+import ctypes
+import os
+
+import _pywhispercpp
+import pywhispercpp.model  # noqa: F401  (importing the extension loads the backends)
+
+libdir = os.path.dirname(_pywhispercpp.__file__)
+base = ctypes.CDLL(os.path.join(libdir, "libggml.so.0"))
+base.ggml_backend_reg_count.restype = ctypes.c_size_t
+base.ggml_backend_reg_get.argtypes = [ctypes.c_size_t]
+base.ggml_backend_reg_get.restype = ctypes.c_void_p
+base.ggml_backend_reg_name.argtypes = [ctypes.c_void_p]
+base.ggml_backend_reg_name.restype = ctypes.c_char_p
+names = [
+    base.ggml_backend_reg_name(base.ggml_backend_reg_get(i)).decode()
+    for i in range(base.ggml_backend_reg_count())
+]
+print("ggml backends registered:", ", ".join(names))
+assert "CPU" in names, names
+loaded = set()
+for line in open("/proc/self/maps"):
+    parts = line.split()
+    if len(parts) >= 6 and "libggml-" in parts[-1]:
+        loaded.add(os.path.basename(parts[-1]))
+cpu = sorted(name for name in loaded if name.startswith("libggml-cpu-"))
+assert cpu, "no CPU variant was loaded"
+print("CPU variant loaded:", ", ".join(cpu))
+# The Vulkan module registers only when a GPU driver answers, and this image
+# has none, so prove what it needs instead: its NEEDED libraries (libvulkan,
+# the ggml base) resolve from the installed layout.
+ctypes.CDLL(os.path.join(libdir, "libggml-vulkan.so"))
+print("Vulkan backend library loads from the package")
+PY
 	python3 -c "import pyaudio; print('pyaudio ok')"
 	python3 -c "import socks; print('pysocks ok')"
 	python3 -c "import vocalinux.main, vocalinux.speech_recognition.recognition_manager; print('vocalinux imports ok')"

packaging/void/srcpkgs/python3-pywhispercpp/patches/load-ggml-backends-beside-module.patch

@@ -0,0 +1,39 @@
+--- a/src/main.cpp
++++ b/src/main.cpp
+@@ -16,6 +16,10 @@
+ #include <pybind11/numpy.h>
+ 
+ #include "whisper.h"
++#include "ggml-backend.h"
++
++#include <dlfcn.h>
++#include <string>
+ 
+ 
+ #define STRINGIFY(x) #x
+@@ -850,7 +854,25 @@
+ 
+ ////////////
+ 
++// With GGML_BACKEND_DL the CPU and GPU backends are libraries next to this
++// extension, and whisper.cpp cannot create a context until a CPU device is
++// registered. ggml searches the executable's directory (the interpreter's, for
++// Python) and the working directory, so load them from this module's directory.
++static void load_ggml_backends_beside_module() {
++    Dl_info info;
++    if (dladdr(reinterpret_cast<void *>(&load_ggml_backends_beside_module), &info) == 0 ||
++        info.dli_fname == nullptr) {
++        ggml_backend_load_all();
++        return;
++    }
++    const std::string path(info.dli_fname);
++    const std::size_t slash = path.find_last_of('/');
++    const std::string dir = slash == std::string::npos ? std::string(".") : path.substr(0, slash);
++    ggml_backend_load_all_from_path(dir.c_str());
++}
++
+ PYBIND11_MODULE(_pywhispercpp, m) {
++    load_ggml_backends_beside_module();
+     m.doc() = R"pbdoc(
+         Pywhispercpp: Python binding to whisper.cpp
+         -----------------------

packaging/void/srcpkgs/python3-pywhispercpp/template

@@ -3,7 +3,7 @@ pkgname=python3-pywhispercpp
 version=1.5.0
 revision=1
 build_style=python3-module
-hostmakedepends="cmake pkg-config python3-setuptools python3-wheel $(vopt_if vulkan shaderc)"
+hostmakedepends="cmake pkg-config python3-setuptools python3-setuptools_scm python3-wheel $(vopt_if vulkan shaderc)"
 makedepends="python3-devel $(vopt_if vulkan vulkan-loader-devel)"
 depends="python3-numpy python3-requests python3-tqdm python3-platformdirs $(vopt_if vulkan vulkan-loader)"
 short_desc="Python bindings for whisper.cpp speech recognition"
@@ -20,21 +20,23 @@ build_options="vulkan"
 build_options_default="vulkan"
 
 pre_build() {
-	# The sdist derives its version from git metadata, which is absent in
-	# the xbps build root, so inject the pinned version explicitly. This is
-	# the same version install.sh (PYWHISPERCPP_VERSION) and uv.lock pin.
-	export PYWHISPERCPP_VERSION=${version}
-	# Portable x86-64 baseline: GGML_NATIVE is ggml's default and compiles
-	# libggml-cpu with -march=native, so the package would inherit the
-	# builder's instruction set and SIGILL on older CPUs (#938).
-	# GGML_NATIVE=OFF alone is not enough: while the native default is on,
-	# ggml turns every GGML_* SIMD option on and still compiles -mavx2
-	# -mfma (#957). Each option is pinned off, mirroring the Flatpak build.
-	# Shared libraries with an $ORIGIN rpath let the extension resolve the
-	# whisper.cpp libraries installed beside it (as install.sh does), with
-	# no LD_LIBRARY_PATH at runtime. (setup.py install never builds a
-	# wheel, so NO_REPAIR is meaningless here and deliberately unset.)
-	export CMAKE_ARGS='-DCMAKE_INSTALL_RPATH=$ORIGIN -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON -DBUILD_SHARED_LIBS=ON -DGGML_NATIVE=OFF -DGGML_SSE42=OFF -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_BMI2=OFF -DGGML_FMA=OFF -DGGML_F16C=OFF'
+	# Version (#43). setup.py passes no version to setup(); the version comes
+	# from setuptools-scm, which the sdist declares as a build requirement and
+	# which hostmakedepends provides. The xbps build root has no git metadata,
+	# so pin the version through setuptools-scm. Without it setuptools falls
+	# back to 0.0.0. The sdist's PYWHISPERCPP_VERSION hook is never called.
+	export SETUPTOOLS_SCM_PRETEND_VERSION=${version}
+	# CPU (#44). GGML_BACKEND_DL builds each CPU variant (x64 baseline,
+	# sandybridge, haswell, icelake, zen4, ...) as its own library. The
+	# extension, through patches/, loads the variants that sit beside it, and
+	# ggml picks the best one the host CPU supports. The x64 baseline always
+	# ships, so the package still runs on any x86-64 CPU, and hosts with AVX2
+	# use it. GGML_NATIVE must stay OFF: ggml refuses it with GGML_BACKEND_DL,
+	# and it keeps -march=native out of every variant (#938).
+	# Shared libraries with an $ORIGIN rpath let each variant resolve
+	# libggml-base beside it, with no LD_LIBRARY_PATH at runtime. (setup.py
+	# install never builds a wheel, so NO_REPAIR is meaningless here.)
+	export CMAKE_ARGS='-DCMAKE_INSTALL_RPATH=$ORIGIN -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON -DBUILD_SHARED_LIBS=ON -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON'
 	if [ -n "$build_option_vulkan" ]; then
 		# Vulkan acceleration for every GPU vendor. There is no CUDA
 		# build: Void ships no CUDA toolkit package.
@@ -42,6 +44,20 @@ pre_build() {
 	fi
 }
 
+pre_install() {
+	# The backend modules (libggml-cpu-*.so, libggml-vulkan.so) are written to
+	# build/bin, a CMake directory whose whisper.cpp setup fixes, and setup.py
+	# installs only the build/lib tree that holds the extension. Copy them
+	# beside the extension, where patches/ loads them from.
+	local backend found=""
+	for backend in build/temp.*/_pywhispercpp/bin/libggml-*.so; do
+		[ -f "$backend" ] || continue
+		cp "$backend" build/lib.*/
+		found=1
+	done
+	[ -n "$found" ] || msg_error "no ggml backend modules were built; the CPU variants are missing\n"
+}
+
 post_install() {
 	vlicense LICENSE
 }

src/vocalinux/speech_recognition/recognition_manager.py

@@ -199,7 +199,11 @@ def _pywhispercpp_distribution_version() -> Optional[tuple[int, ...]]:
             parts.append(int(token))
         else:
             break
-    return tuple(parts) if parts else None
+    # setuptools writes 0.0.0 when a build had no version (#43). That means
+    # unknown, not old, so the feature probe in the caller decides instead.
+    if not parts or tuple(parts) == (0, 0, 0):
+        return None
+    return tuple(parts)
 
 
 def _select_ggml_gpu_device(

src/vocalinux/utils/pywhispercpp_loader.py

@@ -83,6 +83,18 @@ def find_shared_library_dirs() -> list[str]:
     return library_dirs
 
 
+def _is_runtime_cpu_variant(path: Path) -> bool:
+    """True for a ggml CPU variant that ggml's own loader selects at run time.
+
+    GGML_CPU_ALL_VARIANTS builds one ``libggml-cpu-<level>.so`` per instruction
+    set level. The variants export the same symbols, so preloading them with
+    RTLD_GLOBAL would let the first one loaded interpose the others. ggml
+    loads the one the host needs. A non-variant build has a single
+    ``libggml-cpu.so``, which the extension still needs preloaded.
+    """
+    return path.name.startswith("libggml-cpu-")
+
+
 def preload_shared_libraries() -> None:
     """Preload bundled pywhispercpp shared libraries for source-built installs."""
     if _PRELOADED_LIBS:
@@ -91,7 +103,9 @@ def preload_shared_libraries() -> None:
     libraries: list[Path] = []
     for library_dir in find_shared_library_dirs():
         root = Path(library_dir)
-        libraries.extend(sorted(root.glob("libggml*.so*")))
+        libraries.extend(
+            path for path in sorted(root.glob("libggml*.so*")) if not _is_runtime_cpu_variant(path)
+        )
         libraries.extend(sorted(root.glob("libwhisper*.so*")))
 
     if not libraries:

src/vocalinux/utils/whispercpp_model_info.py

@@ -361,7 +361,9 @@ def detect_ggml_gpu_devices() -> Optional[list[dict[str, Any]]]:
     if version_match is None:
         return None
     installed_version = tuple(int(part or 0) for part in version_match.groups())
-    if installed_version < (1, 5, 0):
+    # 0.0.0 is what setuptools writes when a build had no version (#43). That means
+    # unknown, not old, so the runtime probe below decides whether devices are listed.
+    if installed_version != (0, 0, 0) and installed_version < (1, 5, 0):
         return None
 
     try:

tests/test_pywhispercpp_loader.py

@@ -0,0 +1,50 @@
+"""Preloading must leave ggml's runtime-selected CPU variants to ggml (#44)."""
+
+from pathlib import Path
+from unittest.mock import patch
+
+from vocalinux.utils import pywhispercpp_loader
+
+
+def _preload_names(tmp_path: Path, *names: str) -> list[str]:
+    """Run the preloader over ``names`` in one directory and return what it dlopens."""
+    for name in names:
+        (tmp_path / name).write_bytes(b"")
+    loaded: list[str] = []
+
+    def record(path: str, mode: int = 0) -> object:
+        loaded.append(Path(path).name)
+        return object()
+
+    with (
+        patch.object(pywhispercpp_loader, "find_shared_library_dirs", return_value=[str(tmp_path)]),
+        patch.object(pywhispercpp_loader, "_PRELOADED_LIBS", []),
+        patch.object(pywhispercpp_loader.ctypes, "CDLL", side_effect=record),
+    ):
+        pywhispercpp_loader.preload_shared_libraries()
+    return loaded
+
+
+def test_preload_leaves_runtime_cpu_variants_to_ggml(tmp_path: Path) -> None:
+    """Variants share one symbol table; preloading them RTLD_GLOBAL would interpose them."""
+    loaded = _preload_names(
+        tmp_path,
+        "libggml.so.0",
+        "libggml-base.so.0",
+        "libggml-cpu-x64.so",
+        "libggml-cpu-haswell.so",
+        "libggml-vulkan.so",
+        "libwhisper.so.1",
+    )
+
+    assert not [name for name in loaded if name.startswith("libggml-cpu-")]
+    assert {"libggml.so.0", "libggml-base.so.0", "libggml-vulkan.so", "libwhisper.so.1"} <= set(
+        loaded
+    )
+
+
+def test_preload_keeps_the_single_cpu_library_of_a_non_variant_build(tmp_path: Path) -> None:
+    """Without GGML_CPU_ALL_VARIANTS the CPU backend is a dependency and must be preloaded."""
+    loaded = _preload_names(tmp_path, "libggml.so.0", "libggml-cpu.so")
+
+    assert "libggml-cpu.so" in loaded

tests/test_recognition_manager_device_config.py

@@ -1967,6 +1967,40 @@ class TestWhispercppGpuDeviceSelection(unittest.TestCase):
         assert len(self.ContextParamsModel.calls) == 1
         assert "context_params" not in self.ContextParamsModel.calls[0][1]
 
+    def test_placeholder_distribution_version_is_unknown_not_old(self):
+        """setuptools writes 0.0.0 when a build had no version; that is not 'too old' (#43)."""
+        from vocalinux.speech_recognition import recognition_manager
+
+        with patch("importlib.metadata.version", return_value="0.0.0"):
+            assert recognition_manager._pywhispercpp_distribution_version() is None
+        with patch("importlib.metadata.version", return_value="1.5.0"):
+            assert recognition_manager._pywhispercpp_distribution_version() == (1, 5, 0)
+
+    def test_gpu_device_kept_when_distribution_version_is_placeholder(self):
+        """A 0.0.0 build still has context_params, so device selection must stay on (#43)."""
+        manager = _make_manager(engine="whisper_cpp", whispercpp_gpu_device=1)
+        manager.model_size = "tiny"
+
+        mock_pywhispercpp = MagicMock()
+        self.ContextParamsModel.calls = []
+        mock_pywhispercpp.Model = self.ContextParamsModel
+
+        with patch.dict(
+            "sys.modules",
+            {
+                "pywhispercpp": mock_pywhispercpp,
+                "pywhispercpp.model": mock_pywhispercpp,
+            },
+        ):
+            with patch("importlib.metadata.version", return_value="0.0.0"):
+                result = manager._load_model_with_compatible_params(
+                    "/fake/model.bin", {}, gpu_device=1
+                )
+
+        assert isinstance(result, self.ContextParamsModel)
+        assert len(self.ContextParamsModel.calls) == 1
+        assert self.ContextParamsModel.calls[0][1].get("context_params") == {"gpu_device": 1}
+
     def test_host_cuda_label_runtime_vulkan_selects_discrete(self):
         """Device selection follows bundled Vulkan libs, not a host CUDA label (#604)."""
         manager = _make_manager(engine="whisper_cpp", whispercpp_gpu_device=None)

tests/test_void_pywhispercpp_template.py

@@ -3,7 +3,8 @@
 Spec #16 packages four carried templates in void-packages layout under
 ``packaging/void``; this ticket owns the hard one, ``python3-pywhispercpp``,
 built from the pinned PyPI source distribution with the Vulkan backend,
-shared libraries, and portable CPU flags. Ticket #28 builds the container
+shared libraries, runtime-selected CPU variants (#44), and a version pinned
+through setuptools-scm (#43). Ticket #28 builds the container
 smoke gate on top; everything asserted here runs offline from sources
 already in the repo.
 
@@ -24,21 +25,27 @@ BUILD_SH = VOID_DIR / "build-release.sh"
 INSTALL_SH = REPO_ROOT / "install.sh"
 UV_LOCK = REPO_ROOT / "uv.lock"
 
-#: ggml SIMD options pinned off for the plain x86-64 baseline (#957).
-#: GGML_NATIVE=OFF alone is not portable: while ggml's native default is on,
-#: every GGML_* SIMD option stays on and still compiles -mavx2 -mfma.
-#: Mirrors PORTABLE_GGML_OPTIONS in tests/test_flatpak_packaging.py and the
-#: CMAKE_ARGS line in packaging/flatpak/python3-dependencies.yaml.
-PORTABLE_GGML_OPTIONS = (
-    "GGML_NATIVE",
-    "GGML_SSE42",
-    "GGML_AVX",
-    "GGML_AVX2",
-    "GGML_BMI2",
-    "GGML_FMA",
-    "GGML_F16C",
+#: CMake arguments of the CPU backend (#44): one library per CPU level, loaded
+#: at run time, with the x86-64 baseline always shipped. GGML_NATIVE must stay
+#: OFF: ggml refuses it with GGML_BACKEND_DL, and it would compile the builder's
+#: instruction set into every variant (#938).
+REQUIRED_CMAKE_ARGS = (
+    "-DGGML_NATIVE=OFF",
+    "-DGGML_BACKEND_DL=ON",
+    "-DGGML_CPU_ALL_VARIANTS=ON",
+    "-DBUILD_SHARED_LIBS=ON",
 )
 
+#: The binding patch that loads the backend modules from the extension's directory.
+BINDING_PATCH = (
+    VOID_DIR
+    / "srcpkgs"
+    / "python3-pywhispercpp"
+    / "patches"
+    / "load-ggml-backends-beside-module.patch"
+)
+INSTALL_SMOKE = VOID_DIR / "container-release.sh"
+
 #: install_requires of pywhispercpp's setup.py, mapped to the Void binary
 #: package names. Each name was verified to exist as srcpkgs/<name>/template
 #: in current void-packages (2026-11-05); the container build re-verifies via
@@ -126,21 +133,57 @@ def test_build_style_is_python3_module() -> None:
     assert re.search(r"^build_style=python3-module$", _template(), re.M)
 
 
-def test_portable_cpu_flags_pin_every_ggml_simd_option_off() -> None:
-    """No host-native tuning: the plain x86-64 baseline, like the Flatpak."""
-    text = _template()
-    for flag in PORTABLE_GGML_OPTIONS:
-        assert f"-D{flag}=OFF" in text, (
-            f"template does not pin {flag} off; the package inherits the "
-            "builder's instruction set and SIGILLs on older CPUs (#938, #957)"
-        )
-
-
 def _code_lines() -> str:
     """Template with comment lines stripped; flags live in code, not prose."""
     return "\n".join(line for line in _template().splitlines() if not line.lstrip().startswith("#"))
 
 
+def test_cpu_backend_is_built_as_runtime_selected_variants() -> None:
+    """#44: every CPU level is its own library, and the x86-64 baseline ships."""
+    text = _code_lines()
+    for arg in REQUIRED_CMAKE_ARGS:
+        assert arg in text, f"template does not pass {arg}"
+
+
+def test_no_simd_option_is_forced_on_for_the_whole_build() -> None:
+    """A global SIMD switch would give every variant one instruction set (#938, #957)."""
+    text = _code_lines()
+    for feature in ("NATIVE", "SSE42", "AVX", "AVX2", "BMI2", "FMA", "F16C", "AVX512"):
+        assert f"-DGGML_{feature}=ON" not in text, f"GGML_{feature} forced on for every variant"
+
+
+def test_backends_are_copied_beside_the_extension_before_install() -> None:
+    """#44: whisper.cpp writes the modules to build/bin, which setup.py does not install."""
+    text = _template()
+    assert "pre_install()" in text
+    assert "build/temp.*/_pywhispercpp/bin/libggml-*.so" in text
+    assert "build/lib.*/" in text
+    assert "msg_error" in text, "the build must fail loudly when no backend module was built"
+
+
+def test_binding_loads_the_backends_beside_the_extension() -> None:
+    """#44: ggml's default search misses the package directory, so the binding loads it."""
+    assert BINDING_PATCH.is_file(), f"missing binding patch: {BINDING_PATCH}"
+    text = BINDING_PATCH.read_text(encoding="utf-8")
+    assert "--- a/src/main.cpp" in text and "+++ b/src/main.cpp" in text
+    assert " PYBIND11_MODULE(_pywhispercpp, m) {" in text
+    assert "+    load_ggml_backends_beside_module();" in text
+    assert "+    ggml_backend_load_all_from_path(dir.c_str());" in text
+
+
+def test_install_smoke_checks_version_and_cpu_variants() -> None:
+    """#43 and #44: the installed binding reports the pinned version, and its variants load."""
+    text = INSTALL_SMOKE.read_text(encoding="utf-8")
+    version = re.search(r"^version=(\S+)", _template(), re.M)
+    assert version, "template declares no version"
+    assert (
+        f'[ "$PWC_VERSION" = "{version.group(1)}" ]' in text
+    ), "the install smoke must assert the installed pywhispercpp reports the template version"
+    assert "libggml-cpu-x64.so" in text and "libggml-cpu-haswell.so" in text
+    assert "objdump -d" in text, "the baseline variant must be checked for AVX code"
+    assert "libggml.so.0" in text, "the backend registry lives in libggml"
+
+
 def test_no_native_tuning_or_cuda_anywhere() -> None:
     """Void has no CUDA toolkit package; portable means portable."""
     text = _code_lines()
@@ -181,15 +224,31 @@ def test_build_time_deps_cover_setup_py() -> None:
     text = _template()
     host = re.search(r"^hostmakedepends=(.*)$", text, re.M)
     assert host, "template declares no hostmakedepends"
-    for dep in ("cmake", "pkg-config", "python3-setuptools", "python3-wheel"):
+    for dep in (
+        "cmake",
+        "pkg-config",
+        "python3-setuptools",
+        "python3-setuptools_scm",
+        "python3-wheel",
+    ):
         assert dep in host.group(1), f"{dep} missing from hostmakedepends"
     make = re.search(r"^makedepends=(.*)$", text, re.M)
     assert make and "python3-devel" in make.group(1), "compiled extension needs python3-devel"
 
 
-def test_version_is_injected_for_the_offline_build() -> None:
-    """The sdist carries no version.txt; setup.py reads PYWHISPERCPP_VERSION."""
-    assert "PYWHISPERCPP_VERSION" in _template()
+def test_version_is_pinned_through_setuptools_scm() -> None:
+    """#43: setup() gets no version; setuptools-scm supplies it, so the build must provide it."""
+    host = re.search(r"^hostmakedepends=(.*)$", _template(), re.M)
+    assert host and "python3-setuptools_scm" in host.group(
+        1
+    ), "without setuptools-scm the build reports version 0.0.0"
+    assert (
+        "SETUPTOOLS_SCM_PRETEND_VERSION=${version}" in _code_lines()
+    ), "the xbps build root has no git metadata; pin the template version"
+    installer = (REPO_ROOT / "requirements" / "installer-build.txt").read_text(encoding="utf-8")
+    assert re.search(
+        r"^setuptools-scm==", installer, re.M
+    ), "the installer build tools pin setuptools-scm as well"
 
 
 def test_build_fetches_nothing_beyond_the_sdist() -> None:

tests/test_whispercpp_model_info_ext.py

@@ -580,6 +580,14 @@ class TestDetectGgmlGpuDevices(unittest.TestCase):
         assert devices is not None
         self.assertEqual(len(devices), 1)
 
+    def test_probe_trusts_runtime_when_version_is_placeholder(
+        self: "TestDetectGgmlGpuDevices",
+    ) -> None:
+        """setuptools' 0.0.0 means the build had no version (#43): probe the runtime."""
+        devices = _run_ggml_probe([("Vulkan0", "GPU", 1)], version_string="0.0.0")
+        assert devices is not None
+        self.assertEqual(len(devices), 1)
+
 
 class TestPreferDiscreteVulkanDeviceMore(unittest.TestCase):
     """Exercise additional Vulkan selection edge cases."""