chore(deps): bump the pip group across 3 directories with 1 update

Bumps the pip group with 1 update in the /backend/python/sglang directory: torch.
Bumps the pip group with 1 update in the /backend/python/trl directory: torch.
Bumps the pip group with 1 update in the /backend/python/vllm-omni directory: torch.


Updates `torch` from 2.9.0 to 2.12.0+cpu

Updates `torch` from 2.10.0 to 2.12.0+cpu

Updates `torch` from 2.7.0 to 2.12.0+cu130

---
updated-dependencies:
- dependency-name: torch
  dependency-version: 2.12.0+cpu
  dependency-type: direct:production
  dependency-group: pip
- dependency-name: torch
  dependency-version: 2.12.0+cpu
  dependency-type: direct:production
  dependency-group: pip
- dependency-name: torch
  dependency-version: 2.12.0+cu130
  dependency-type: direct:production
  dependency-group: pip
...

Signed-off-by: dependabot[bot] <support@github.com>

.agents/adding-backends.md

@@ -102,6 +102,24 @@ Multi-arch backends are NOT a single matrix entry with `platforms: 'linux/amd64,
 
 Entries whose `dockerfile` is `./backend/Dockerfile.{llama-cpp,ik-llama-cpp,turboquant}` must also set a `builder-base-image` field pointing at a prebuilt base from `quay.io/go-skynet/ci-cache:base-grpc-*` (CI builds these via `.github/workflows/base-images.yml`). The mapping is by `(build-type, platforms)` — see existing entries for the pattern. CI uses these prebuilt bases to skip the gRPC compile (~25–35 min cold). Local `make backends/<name>` ignores `builder-base-image` and uses the from-source path inside the Dockerfile, so you don't need quay access for local builds.
 
+### Cover every OS the project supports (Linux **and** Darwin)
+
+`.github/backend-matrix.yml` has two matrices, and they are the source of truth for which OS a backend ships on:
+
+- `include:` — the **Linux** matrix (x86_64 + arm64; CPU and CUDA / ROCm / SYCL / Vulkan).
+- `includeDarwin:` — the **macOS / Apple Silicon** matrix (arm64; Metal where the engine supports it, otherwise a native arm64 CPU build).
+
+**A new backend must target every OS it can build for — do not ship Linux-only by default.** A backend that appears only under `include:` is silently unavailable on macOS even when its code would run there. Most C/C++/GGML engines build on Darwin out of the box (ggml defaults `GGML_METAL=ON` on Apple, so a plain build is Metal-enabled), and many Python backends do too (CPU / MPS wheels). If a backend genuinely cannot support an OS (e.g. CUDA-only, no CPU variant), state that in the PR description instead of omitting it silently.
+
+Wiring a backend into `includeDarwin:` is more than the matrix entry:
+
+1. **`includeDarwin:` entry** — `tag-suffix: "-metal-darwin-arm64-<backend>"`, `build-type: "metal"`, `lang: "go"` for go+ggml backends; omit `build-type` for the bespoke C++ ones (llama-cpp / ds4 / privacy-filter). Match an existing entry of the same shape.
+2. **`backend/index.yaml`** — add `metal:` to the backend's `capabilities` map (main and `-development`) and concrete `metal-<backend>` / `metal-<backend>-development` image entries pointing at the `-metal-darwin-arm64-<backend>` images.
+3. **C/C++ backends only** — add an `inferBackendPathDarwin` case in `scripts/changed-backends.js` returning `backend/cpp/<backend>/` (the generic fallthrough assumes `backend/<lang>/`, which is wrong for a C++ source tree driven with `lang: go`), and give `run.sh` a Darwin branch that exports `DYLD_LIBRARY_PATH` instead of `LD_LIBRARY_PATH`. If the build is bespoke (single `grpc-server` + dylib bundling), model it on `scripts/build/ds4-darwin.sh` and add a `backends/<backend>-darwin` make target plus a gated step in `.github/workflows/backend_build_darwin.yml`.
+4. **C++ proto gotcha** — if the backend compiles the generated gRPC/protobuf in a separate CMake target (e.g. `hw_grpc_proto`), that target must link `protobuf::libprotobuf` + `gRPC::grpc++` so the Homebrew include dirs propagate; otherwise macOS fails with `google/protobuf/runtime_version.h not found` (Linux hides this because apt headers sit in `/usr/include`).
+
+The CI path filter only builds a backend on a PR when a file under its directory changes, so a darwin-only YAML edit builds nothing — touch a file under `backend/<lang>/<backend>/` (a one-line comment is enough) in the same PR.
+
 ## 3. Add Backend Metadata to `backend/index.yaml`
 
 **Step 3a: Add Meta Definition**
@@ -225,6 +243,7 @@ After adding a new backend, verify:
 
 - [ ] Backend directory structure is complete with all necessary files
 - [ ] Build configurations added to `.github/backend-matrix.yml` for all desired platforms (per-arch entries with `platform-tag` for multi-arch; `builder-base-image` for llama-cpp / ik-llama-cpp / turboquant)
+- [ ] **OS coverage considered**: added to `includeDarwin:` (macOS/Apple Silicon) if the backend can build there — with the `backend/index.yaml` `metal:` capability + `metal-<backend>` image entries, a `run.sh` Darwin/DYLD branch and `inferBackendPathDarwin` case for C++ backends — or the PR explains why an OS is unsupported. Do not ship Linux-only by default.
 - [ ] Meta definition added to `backend/index.yaml` in the `## metas` section
 - [ ] Image entries added to `backend/index.yaml` for all build variants (latest + development)
 - [ ] Tag suffixes match between workflow file and index.yaml

.docker/llama-cpp-compile.sh

@@ -17,19 +17,29 @@ if [[ -n "${CUDA_DOCKER_ARCH:-}" ]]; then
   rm -rf /LocalAI/backend/cpp/llama-cpp-*-build
 fi
 
-if [ "${TARGETARCH}" = "arm64" ] || [ "${BUILD_TYPE}" = "hipblas" ]; then
-  cd /LocalAI/backend/cpp/llama-cpp
-  make llama-cpp-fallback
-  make llama-cpp-grpc
-  make llama-cpp-rpc-server
+cd /LocalAI/backend/cpp/llama-cpp
+if [ -z "${BUILD_TYPE:-}" ]; then
+  # Pure CPU image (BUILD_TYPE empty): one build with ggml CPU_ALL_VARIANTS replaces the
+  # per-microarch binaries (x86: avx/avx2/avx512/fallback; arm64: armv8.x/armv9.x). ggml
+  # dlopens the best libggml-cpu-*.so at runtime by probing host CPU features.
+  #
+  # arm64: the CPU_ALL_VARIANTS table includes armv9.2 SME variants whose -march=...+sme is
+  # rejected by the Ubuntu 24.04 default gcc-13. gcc-14 accepts it, so build the arm64
+  # variants with it (the host never *selects* SME unless it has it, but every variant must
+  # still compile).
+  if [ "${TARGETARCH}" = "arm64" ]; then
+    apt-get update -qq && apt-get install -y -qq gcc-14 g++-14
+    export CC=gcc-14 CXX=g++-14
+  fi
+  make llama-cpp-cpu-all
 else
-  cd /LocalAI/backend/cpp/llama-cpp
-  make llama-cpp-avx
-  make llama-cpp-avx2
-  make llama-cpp-avx512
+  # GPU build (cublas/hipblas/sycl/vulkan/...): the accelerator does the compute, so a
+  # single fallback CPU build is enough - no per-microarch CPU variants needed. (This also
+  # keeps the heavy GPU backend compile from also building the whole CPU variant matrix,
+  # and avoids the gcc-14 apt step on GPU base images such as nvidia l4t.)
   make llama-cpp-fallback
-  make llama-cpp-grpc
-  make llama-cpp-rpc-server
 fi
+make llama-cpp-grpc
+make llama-cpp-rpc-server
 
 ccache -s || true

.docker/turboquant-compile.sh

@@ -19,17 +19,21 @@ fi
 
 cd /LocalAI/backend/cpp/turboquant
 
-if [ "${TARGETARCH}" = "arm64" ] || [ "${BUILD_TYPE}" = "hipblas" ]; then
-  make turboquant-fallback
-  make turboquant-grpc
-  make turboquant-rpc-server
+if [ -z "${BUILD_TYPE:-}" ]; then
+  # Pure CPU image: one ggml CPU_ALL_VARIANTS build replaces the per-microarch binaries.
+  # arm64: the armv9.2 SME variants need gcc-14 (gcc-13 rejects +sme).
+  if [ "${TARGETARCH}" = "arm64" ]; then
+    apt-get update -qq && apt-get install -y -qq gcc-14 g++-14
+    export CC=gcc-14 CXX=g++-14
+  fi
+  make turboquant-cpu-all
 else
-  make turboquant-avx
-  make turboquant-avx2
-  make turboquant-avx512
+  # GPU build (cublas/hipblas/sycl/vulkan/...): single fallback CPU build, the accelerator
+  # does the compute. Keeps the GPU compile from also building the CPU variant matrix and
+  # avoids the gcc-14 apt step on GPU base images such as nvidia l4t.
   make turboquant-fallback
-  make turboquant-grpc
-  make turboquant-rpc-server
 fi
+make turboquant-grpc
+make turboquant-rpc-server
 
 ccache -s || true

.github/backend-matrix.yml

@@ -2,6 +2,28 @@
 # Matrix data for backend container image builds.
 # Consumed by scripts/changed-backends.js for both backend.yml and backend_pr.yml.
 # This file is NOT a workflow — it has no top-level 'on:' or 'jobs:'.
+#
+# OS / platform coverage — READ THIS WHEN ADDING A BACKEND
+# --------------------------------------------------------
+# This file is the source of truth for which OS each backend is built and
+# published for. A backend ships ONLY for the matrices it appears in:
+#   - Linux  -> the `include:` matrix below (x86_64 + arm64; CPU and
+#               CUDA / ROCm / SYCL / Vulkan variants).
+#   - macOS  -> the `includeDarwin:` matrix (Apple Silicon / arm64; Metal where
+#               the engine supports it, otherwise a native arm64 CPU build).
+#
+# New backends must target EVERY OS they can build for, not just Linux. A backend
+# listed only under `include:` is silently unavailable on macOS even when its code
+# would run there. Most C/C++/GGML engines build on Darwin (ggml defaults
+# GGML_METAL=ON on Apple, so a plain build is Metal-enabled), and many Python
+# backends do too (CPU / MPS). If a backend genuinely cannot support an OS, say so
+# in its PR description rather than silently omitting it.
+#
+# Adding a backend to `includeDarwin:` is more than one line — see the darwin
+# checklist in .agents/adding-backends.md (includeDarwin entry, the index.yaml
+# `metal:` capability + `metal-<backend>` image entries, a `run.sh` Darwin/DYLD
+# branch for C/C++ backends, and the inferBackendPathDarwin case in
+# scripts/changed-backends.js so the path filter actually builds it).
 
 # Linux matrix (consumed by backend-jobs).
 include:
@@ -3723,6 +3745,302 @@ include:
     dockerfile: "./backend/Dockerfile.golang"
     context: "./"
     ubuntu-version: '2404'
+  # voice-detect
+  - build-type: 'cublas'
+    cuda-major-version: "12"
+    cuda-minor-version: "8"
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-nvidia-cuda-12-voice-detect'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "voice-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
+  - build-type: 'cublas'
+    cuda-major-version: "13"
+    cuda-minor-version: "0"
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-nvidia-cuda-13-voice-detect'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "voice-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
+  - build-type: 'cublas'
+    cuda-major-version: "13"
+    cuda-minor-version: "0"
+    platforms: 'linux/arm64'
+    skip-drivers: 'false'
+    tag-latest: 'auto'
+    tag-suffix: '-nvidia-l4t-cuda-13-arm64-voice-detect'
+    base-image: "ubuntu:24.04"
+    ubuntu-version: '2404'
+    runs-on: 'ubuntu-24.04-arm'
+    backend: "voice-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+  - build-type: ''
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    platform-tag: 'amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-cpu-voice-detect'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "voice-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
+  - build-type: ''
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/arm64'
+    platform-tag: 'arm64'
+    tag-latest: 'auto'
+    tag-suffix: '-cpu-voice-detect'
+    runs-on: 'ubuntu-24.04-arm'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "voice-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
+  - build-type: 'sycl_f32'
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-intel-sycl-f32-voice-detect'
+    runs-on: 'ubuntu-latest'
+    base-image: "intel/oneapi-basekit:2025.3.0-0-devel-ubuntu24.04"
+    skip-drivers: 'false'
+    backend: "voice-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
+  - build-type: 'sycl_f16'
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-intel-sycl-f16-voice-detect'
+    runs-on: 'ubuntu-latest'
+    base-image: "intel/oneapi-basekit:2025.3.0-0-devel-ubuntu24.04"
+    skip-drivers: 'false'
+    backend: "voice-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
+  - build-type: 'vulkan'
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    platform-tag: 'amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-vulkan-voice-detect'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "voice-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
+  - build-type: 'vulkan'
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/arm64'
+    platform-tag: 'arm64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-vulkan-voice-detect'
+    runs-on: 'ubuntu-24.04-arm'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "voice-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
+  - build-type: 'cublas'
+    cuda-major-version: "12"
+    cuda-minor-version: "0"
+    platforms: 'linux/arm64'
+    skip-drivers: 'false'
+    tag-latest: 'auto'
+    tag-suffix: '-nvidia-l4t-arm64-voice-detect'
+    base-image: "nvcr.io/nvidia/l4t-jetpack:r36.4.0"
+    runs-on: 'ubuntu-24.04-arm'
+    backend: "voice-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2204'
+  - build-type: 'hipblas'
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-rocm-hipblas-voice-detect'
+    base-image: "rocm/dev-ubuntu-24.04:7.2.1"
+    runs-on: 'ubuntu-latest'
+    skip-drivers: 'false'
+    backend: "voice-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
+  # face-detect
+  - build-type: 'cublas'
+    cuda-major-version: "12"
+    cuda-minor-version: "8"
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-nvidia-cuda-12-face-detect'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "face-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
+  - build-type: 'cublas'
+    cuda-major-version: "13"
+    cuda-minor-version: "0"
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-nvidia-cuda-13-face-detect'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "face-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
+  - build-type: 'cublas'
+    cuda-major-version: "13"
+    cuda-minor-version: "0"
+    platforms: 'linux/arm64'
+    skip-drivers: 'false'
+    tag-latest: 'auto'
+    tag-suffix: '-nvidia-l4t-cuda-13-arm64-face-detect'
+    base-image: "ubuntu:24.04"
+    ubuntu-version: '2404'
+    runs-on: 'ubuntu-24.04-arm'
+    backend: "face-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+  - build-type: ''
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    platform-tag: 'amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-cpu-face-detect'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "face-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
+  - build-type: ''
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/arm64'
+    platform-tag: 'arm64'
+    tag-latest: 'auto'
+    tag-suffix: '-cpu-face-detect'
+    runs-on: 'ubuntu-24.04-arm'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "face-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
+  - build-type: 'sycl_f32'
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-intel-sycl-f32-face-detect'
+    runs-on: 'ubuntu-latest'
+    base-image: "intel/oneapi-basekit:2025.3.0-0-devel-ubuntu24.04"
+    skip-drivers: 'false'
+    backend: "face-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
+  - build-type: 'sycl_f16'
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-intel-sycl-f16-face-detect'
+    runs-on: 'ubuntu-latest'
+    base-image: "intel/oneapi-basekit:2025.3.0-0-devel-ubuntu24.04"
+    skip-drivers: 'false'
+    backend: "face-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
+  - build-type: 'vulkan'
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    platform-tag: 'amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-vulkan-face-detect'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "face-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
+  - build-type: 'vulkan'
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/arm64'
+    platform-tag: 'arm64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-vulkan-face-detect'
+    runs-on: 'ubuntu-24.04-arm'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "face-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
+  - build-type: 'cublas'
+    cuda-major-version: "12"
+    cuda-minor-version: "0"
+    platforms: 'linux/arm64'
+    skip-drivers: 'false'
+    tag-latest: 'auto'
+    tag-suffix: '-nvidia-l4t-arm64-face-detect'
+    base-image: "nvcr.io/nvidia/l4t-jetpack:r36.4.0"
+    runs-on: 'ubuntu-24.04-arm'
+    backend: "face-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2204'
+  - build-type: 'hipblas'
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-rocm-hipblas-face-detect'
+    base-image: "rocm/dev-ubuntu-24.04:7.2.1"
+    runs-on: 'ubuntu-latest'
+    skip-drivers: 'false'
+    backend: "face-detect"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
   # acestep-cpp
   - build-type: ''
     cuda-major-version: ""
@@ -4906,6 +5224,14 @@ includeDarwin:
     tag-suffix: "-metal-darwin-arm64-ced"
     build-type: "metal"
     lang: "go"
+  - backend: "voice-detect"
+    tag-suffix: "-metal-darwin-arm64-voice-detect"
+    build-type: "metal"
+    lang: "go"
+  - backend: "face-detect"
+    tag-suffix: "-metal-darwin-arm64-face-detect"
+    build-type: "metal"
+    lang: "go"
   - backend: "acestep-cpp"
     tag-suffix: "-metal-darwin-arm64-acestep-cpp"
     build-type: "metal"
@@ -4922,6 +5248,37 @@ includeDarwin:
     tag-suffix: "-metal-darwin-arm64-vibevoice-cpp"
     build-type: "metal"
     lang: "go"
+  # Vision/utility C++/ggml backends (go+cgo). Their Makefiles already carry a
+  # Darwin/Metal path (GGML_METAL=ON when build-type=metal); this just builds and
+  # publishes the metal image so Apple Silicon can install them.
+  - backend: "depth-anything-cpp"
+    tag-suffix: "-metal-darwin-arm64-depth-anything-cpp"
+    build-type: "metal"
+    lang: "go"
+  - backend: "locate-anything-cpp"
+    tag-suffix: "-metal-darwin-arm64-locate-anything-cpp"
+    build-type: "metal"
+    lang: "go"
+  - backend: "rfdetr-cpp"
+    tag-suffix: "-metal-darwin-arm64-rfdetr-cpp"
+    build-type: "metal"
+    lang: "go"
+  - backend: "sam3-cpp"
+    tag-suffix: "-metal-darwin-arm64-sam3-cpp"
+    build-type: "metal"
+    lang: "go"
+  # privacy-filter (PII/NER) is a C++/ggml backend built by a bespoke darwin
+  # script (make backends/privacy-filter-darwin); ggml defaults Metal ON on Apple
+  # so the build is Metal-enabled. lang=go drives runner/toolchain selection only.
+  - backend: "privacy-filter"
+    tag-suffix: "-metal-darwin-arm64-privacy-filter"
+    lang: "go"
+  # LocalVQE has no Metal path; on Apple Silicon it builds CPU-only (GGML_METAL
+  # OFF) but is still a native arm64 image. Uses the darwin/metal build profile.
+  - backend: "localvqe"
+    tag-suffix: "-metal-darwin-arm64-localvqe"
+    build-type: "metal"
+    lang: "go"
   - backend: "voxtral"
     tag-suffix: "-metal-darwin-arm64-voxtral"
     build-type: "metal"
@@ -4938,9 +5295,6 @@ includeDarwin:
   - backend: "qwen-tts"
     tag-suffix: "-metal-darwin-arm64-qwen-tts"
     build-type: "mps"
-  - backend: "fish-speech"
-    tag-suffix: "-metal-darwin-arm64-fish-speech"
-    build-type: "mps"
   - backend: "voxcpm"
     tag-suffix: "-metal-darwin-arm64-voxcpm"
     build-type: "mps"
@@ -4974,6 +5328,19 @@ includeDarwin:
   - backend: "kitten-tts"
     tag-suffix: "-metal-darwin-arm64-kitten-tts"
     build-type: "mps"
+  # vLLM on Apple Silicon via vllm-metal (MLX). The install is custom
+  # (backend/python/vllm/install.sh has a darwin branch); lang stays python so
+  # backend_build_darwin.yml drives it through build-darwin-python-backend ->
+  # scripts/build/python-darwin.sh, which runs the backend's install.sh.
+  - backend: "vllm"
+    tag-suffix: "-metal-darwin-arm64-vllm"
+    build-type: "mps"
+  - backend: "trl"
+    tag-suffix: "-metal-darwin-arm64-trl"
+    build-type: "mps"
+  - backend: "liquid-audio"
+    tag-suffix: "-metal-darwin-arm64-liquid-audio"
+    build-type: "mps"
   - backend: "piper"
     tag-suffix: "-metal-darwin-arm64-piper"
     build-type: "metal"
@@ -4990,6 +5357,10 @@ includeDarwin:
     tag-suffix: "-metal-darwin-arm64-sherpa-onnx"
     build-type: "metal"
     lang: "go"
+  - backend: "supertonic"
+    tag-suffix: "-metal-darwin-arm64-supertonic"
+    build-type: "metal"
+    lang: "go"
   - backend: "local-store"
     tag-suffix: "-metal-darwin-arm64-local-store"
     build-type: "metal"

.github/bump_vllm_metal.sh

@@ -0,0 +1,55 @@
+#!/bin/bash
+# Bump the single vllm-metal pin (VLLM_METAL_VERSION) in the vLLM backend's
+# darwin (Apple Silicon) install path. The macOS/Metal build
+# (backend/python/vllm/install.sh, Darwin branch) installs vllm-metal, which is
+# version-locked to a specific vLLM source release. install.sh derives that vLLM
+# version at build time from vllm-metal's own installer (`vllm_v=`) at the pinned
+# tag, so there is only ONE value to bump here -- mirroring bump_vllm_wheel.sh,
+# which bumps the Linux cu130 wheel pin.
+#
+# This deliberately tracks vllm-project/vllm-metal, NOT vllm-project/vllm: the
+# darwin build can only use the exact vLLM version vllm-metal supports, so it may
+# lag the Linux pin (requirements-cublas13-after.txt) until vllm-metal catches up.
+set -xe
+REPO=$1   # vllm-project/vllm-metal
+FILE=$2   # backend/python/vllm/install.sh
+VAR=$3    # VLLM_METAL_VERSION (used for the workflow's output file names)
+
+if [ -z "$FILE" ] || [ -z "$REPO" ] || [ -z "$VAR" ]; then
+    echo "usage: $0 <repo> <install-file> <var-name>" >&2
+    exit 1
+fi
+
+# vllm-metal ships frequent dev releases, all flagged as non-prerelease, so
+# /releases/latest returns the newest one (with its cp312 wheel asset).
+LATEST_TAG=$(curl -sS -H "Accept: application/vnd.github+json" \
+    "https://api.github.com/repos/$REPO/releases/latest" \
+    | python3 -c "import json,sys; print(json.load(sys.stdin)['tag_name'])")
+
+# The coupled vLLM source version lives in vllm-metal's installer at that tag.
+NEW_VLLM_VERSION=$(curl -fsSL \
+    "https://raw.githubusercontent.com/$REPO/$LATEST_TAG/install.sh" \
+    | grep -oE 'vllm_v="[0-9]+\.[0-9]+\.[0-9]+"' | head -1 | cut -d'"' -f2)
+
+if [ -z "$LATEST_TAG" ] || [ -z "$NEW_VLLM_VERSION" ]; then
+    echo "Could not resolve vllm-metal tag ($LATEST_TAG) or its vllm_v ($NEW_VLLM_VERSION)." >&2
+    exit 1
+fi
+
+set +e
+CURRENT_TAG=$(grep -oE 'VLLM_METAL_VERSION="[^"]*"' "$FILE" | head -1 | cut -d'"' -f2)
+set -e
+
+# Rewrite the single pin. install.sh derives VLLM_VERSION from this tag at build
+# time, so there is nothing else to touch. peter-evans/create-pull-request opens
+# no PR on a clean tree, so a no-op rewrite (already current) is safe.
+sed -i "$FILE" \
+    -e "s|VLLM_METAL_VERSION=\"[^\"]*\"|VLLM_METAL_VERSION=\"$LATEST_TAG\"|"
+
+if [ -z "$CURRENT_TAG" ]; then
+    echo "Could not find VLLM_METAL_VERSION=\"...\" in $FILE." >&2
+    exit 0
+fi
+
+echo "vllm-metal ${CURRENT_TAG} -> ${LATEST_TAG} (builds vLLM ${NEW_VLLM_VERSION}): https://github.com/$REPO/releases/tag/${LATEST_TAG}" >> "${VAR}_message.txt"
+echo "${LATEST_TAG}" >> "${VAR}_commit.txt"

.github/workflows/backend_build_darwin.yml

@@ -99,6 +99,7 @@ jobs:
             /opt/homebrew/Cellar/xxhash
             /opt/homebrew/Cellar/zstd
             /opt/homebrew/Cellar/nlohmann-json
+            /opt/homebrew/Cellar/opus
           key: brew-${{ runner.os }}-${{ runner.arch }}-v1-${{ hashFiles('.github/workflows/backend_build_darwin.yml') }}
 
       - name: Dependencies
@@ -113,7 +114,12 @@ jobs:
           # nlohmann-json is header-only and required by the ds4 backend
           # (dsml_renderer.cpp includes <nlohmann/json.hpp>); on Linux it comes
           # from the apt-installed nlohmann-json3-dev in the build image.
-          brew install protobuf grpc make protoc-gen-go protoc-gen-go-grpc libomp llvm ccache blake3 fmt hiredis xxhash zstd nlohmann-json
+          # opus + pkg-config are required by the opus go backend: its
+          # Makefile/package.sh call `pkg-config --cflags/--libs opus` to build
+          # libopusshim.dylib and to locate libopus.dylib for bundling. brew's
+          # pkg-config defaults its search path to the Homebrew prefix so the
+          # opus.pc is found.
+          brew install protobuf grpc make protoc-gen-go protoc-gen-go-grpc libomp llvm ccache blake3 fmt hiredis xxhash zstd nlohmann-json opus pkg-config
           # Force-reinstall ccache so brew re-validates its full runtime-dep
           # closure on every run. This is the durable fix: when the upstream
           # ccache formula gains a new transitive dep (as it has multiple times
@@ -132,7 +138,7 @@ jobs:
           # and decides "already installed" without re-linking, so on a cache-
           # hit run the formulas aren't on PATH. Force-link them; --overwrite
           # tolerates pre-existing symlinks from earlier installs.
-          brew link --overwrite protobuf grpc make protoc-gen-go protoc-gen-go-grpc libomp llvm ccache blake3 fmt hiredis xxhash zstd nlohmann-json 2>/dev/null || true
+          brew link --overwrite protobuf grpc make protoc-gen-go protoc-gen-go-grpc libomp llvm ccache blake3 fmt hiredis xxhash zstd nlohmann-json opus pkg-config 2>/dev/null || true
 
       - name: Save Homebrew cache
         if: github.event_name != 'pull_request' && steps.brew-cache.outputs.cache-hit != 'true'
@@ -153,6 +159,7 @@ jobs:
             /opt/homebrew/Cellar/xxhash
             /opt/homebrew/Cellar/zstd
             /opt/homebrew/Cellar/nlohmann-json
+            /opt/homebrew/Cellar/opus
           key: brew-${{ runner.os }}-${{ runner.arch }}-v1-${{ hashFiles('.github/workflows/backend_build_darwin.yml') }}
 
       # ---- ccache for llama.cpp CMake builds ----
@@ -228,8 +235,17 @@ jobs:
         run: |
           make backends/ds4-darwin
 
+      # privacy-filter is a C++/ggml backend like ds4 - a single grpc-server with
+      # otool dylib bundling - so it gets its own bespoke darwin script rather than
+      # the generic build-darwin-go-backend path.
+      - name: Build privacy-filter backend (Darwin Metal)
+        if: inputs.backend == 'privacy-filter'
+        run: |
+          make protogen-go
+          make backends/privacy-filter-darwin
+
       - name: Build ${{ inputs.backend }}-darwin
-        if: inputs.backend != 'llama-cpp' && inputs.backend != 'ds4'
+        if: inputs.backend != 'llama-cpp' && inputs.backend != 'ds4' && inputs.backend != 'privacy-filter'
         run: |
           make protogen-go
           BACKEND=${{ inputs.backend }} BUILD_TYPE=${{ inputs.build-type }} USE_PIP=${{ inputs.use-pip }} make build-darwin-${{ inputs.lang }}-backend

.github/workflows/bump_deps.yaml

@@ -46,6 +46,14 @@ jobs:
             variable: "CED_VERSION"
             branch: "master"
             file: "backend/go/ced/Makefile"
+          - repository: "mudler/voice-detect.cpp"
+            variable: "VOICEDETECT_VERSION"
+            branch: "master"
+            file: "backend/go/voice-detect/Makefile"
+          - repository: "mudler/face-detect.cpp"
+            variable: "FACEDETECT_VERSION"
+            branch: "master"
+            file: "backend/go/face-detect/Makefile"
           - repository: "mudler/depth-anything.cpp"
             variable: "DEPTHANYTHING_VERSION"
             branch: "master"
@@ -154,3 +162,39 @@ jobs:
           branch: "update/VLLM_VERSION"
           body: ${{ steps.bump.outputs.message }}
           signoff: true
+
+  bump-vllm-metal:
+    # The darwin (Apple Silicon) vLLM build installs vllm-metal, which is locked
+    # to a specific vLLM source release. install.sh pins both VLLM_METAL_VERSION
+    # (the wheel release) and VLLM_VERSION (the vLLM it builds against); this job
+    # tracks vllm-project/vllm-metal and rewrites both atomically. Separate from
+    # bump-vllm-wheel because darwin follows vllm-metal, not vllm/vllm latest.
+    if: github.repository == 'mudler/LocalAI'
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v7
+      - name: Bump vllm-metal pin 🔧
+        id: bump
+        run: |
+          bash .github/bump_vllm_metal.sh vllm-project/vllm-metal backend/python/vllm/install.sh VLLM_METAL_VERSION
+          {
+            echo 'message<<EOF'
+            cat "VLLM_METAL_VERSION_message.txt"
+            echo EOF
+          } >> "$GITHUB_OUTPUT"
+          {
+            echo 'commit<<EOF'
+            cat "VLLM_METAL_VERSION_commit.txt"
+            echo EOF
+          } >> "$GITHUB_OUTPUT"
+          rm -rfv VLLM_METAL_VERSION_message.txt VLLM_METAL_VERSION_commit.txt
+      - name: Create Pull Request
+        uses: peter-evans/create-pull-request@v8
+        with:
+          token: ${{ secrets.UPDATE_BOT_TOKEN }}
+          push-to-fork: ci-forks/LocalAI
+          commit-message: ':arrow_up: Update vllm-project/vllm-metal (darwin)'
+          title: 'chore: :arrow_up: Update vllm-metal (darwin) to `${{ steps.bump.outputs.commit }}`'
+          branch: "update/VLLM_METAL_VERSION"
+          body: ${{ steps.bump.outputs.message }}
+          signoff: true

.github/workflows/release.yaml

@@ -24,6 +24,11 @@ jobs:
           args: release --clean
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          MACOS_SIGN_P12: ${{ secrets.MACOS_CERTIFICATE }}
+          MACOS_SIGN_PASSWORD: ${{ secrets.MACOS_CERTIFICATE_PWD }}
+          MACOS_NOTARY_KEY: ${{ secrets.MACOS_NOTARY_KEY }}
+          MACOS_NOTARY_KEY_ID: ${{ secrets.MACOS_NOTARY_KEY_ID }}
+          MACOS_NOTARY_ISSUER_ID: ${{ secrets.MACOS_NOTARY_ISSUER_ID }}
   launcher-build-darwin:
     runs-on: macos-latest
     steps:
@@ -35,9 +40,19 @@ jobs:
         uses: actions/setup-go@v5
         with:
           go-version: 1.23
-      - name: Build launcher for macOS ARM64
-        run: |
-          make build-launcher-darwin
+      - name: Import signing certificate
+        env:
+          MACOS_CERTIFICATE: ${{ secrets.MACOS_CERTIFICATE }}
+          MACOS_CERTIFICATE_PWD: ${{ secrets.MACOS_CERTIFICATE_PWD }}
+          MACOS_CI_KEYCHAIN_PWD: ${{ secrets.MACOS_CI_KEYCHAIN_PWD }}
+        run: bash contrib/macos/sign-and-notarize.sh import-cert
+      - name: Build, sign and notarize the DMG
+        env:
+          MACOS_SIGN_IDENTITY: ${{ secrets.MACOS_SIGN_IDENTITY }}
+          MACOS_NOTARY_KEY: ${{ secrets.MACOS_NOTARY_KEY }}
+          MACOS_NOTARY_KEY_ID: ${{ secrets.MACOS_NOTARY_KEY_ID }}
+          MACOS_NOTARY_ISSUER_ID: ${{ secrets.MACOS_NOTARY_ISSUER_ID }}
+        run: make release-launcher-darwin
       - name: Upload DMG to Release
         uses: softprops/action-gh-release@v3
         with:

.github/workflows/test-extra.yml

@@ -1008,7 +1008,11 @@ jobs:
   # image + working dir.
   tests-vibevoice-cpp-grpc-transcription:
     needs: detect-changes
-    if: needs.detect-changes.outputs.vibevoice-cpp == 'true' || needs.detect-changes.outputs.run-all == 'true'
+    # Skip on release tag pushes: the ASR Q4_K model is ~10 GB and cannot be
+    # pulled from HF within the inner `go test -timeout 30m` budget on a CI
+    # runner, so every tag build hung and timed out. Still runs on PRs/branch
+    # pushes that touch vibevoice-cpp so regressions are caught off the release path.
+    if: (needs.detect-changes.outputs.vibevoice-cpp == 'true' || needs.detect-changes.outputs.run-all == 'true') && !startsWith(github.ref, 'refs/tags/')
     runs-on: bigger-runner
     timeout-minutes: 150
     steps:

.github/workflows/test.yml

@@ -121,3 +121,19 @@ jobs:
           detached: true
           connect-timeout-seconds: 180
           limit-access-to-actor: true
+
+  # Fast standalone unit tests for the backends' pure C++ helpers - currently the
+  # llama-cpp message reconstruction (backend/cpp/llama-cpp/message_content.h),
+  # which guards the OpenAI chat content normalization (mudler/LocalAI#10524,
+  # #7324, #7528). The runner discovers every *_test.cpp under backend/cpp/, so
+  # new pure-C++ unit tests are picked up with no CI changes. These need only the
+  # C++ stdlib + nlohmann/json, so they run on every PR without the full
+  # llama.cpp + gRPC backend build. (The same suite is also wired as an opt-in
+  # CMake/ctest target, -DLLAMA_GRPC_BUILD_TESTS=ON, for in-backend-build runs.)
+  tests-backend-cpp:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Clone
+        uses: actions/checkout@v7
+      - name: Run backend C++ unit tests
+        run: make test-backend-cpp

.gitignore

@@ -94,3 +94,6 @@ core/http/react-ui/test-results/
 
 # SDD / brainstorm scratch (agent-driven development)
 .superpowers/
+
+# Local Apple signing material (never commit)
+.certs/

.goreleaser.yaml

@@ -9,7 +9,8 @@ source:
   enabled: true
   name_template: '{{ .ProjectName }}-{{ .Tag }}-source'
 builds:
-  - main: ./cmd/local-ai
+  - id: local-ai
+    main: ./cmd/local-ai
     env:
       - CGO_ENABLED=0
     ldflags:
@@ -35,3 +36,19 @@ snapshot:
   version_template: "{{ .Tag }}-next"
 changelog:
   use: github-native
+# Sign + notarize the macOS server binary via the quill backend (runs on Linux,
+# no macOS runner needed). Disabled automatically when MACOS_SIGN_P12 is unset
+# (forks / PRs), so those builds stay unsigned and green.
+notarize:
+  macos:
+    - enabled: '{{ isEnvSet "MACOS_SIGN_P12" }}'
+      ids:
+        - local-ai
+      sign:
+        certificate: "{{.Env.MACOS_SIGN_P12}}"
+        password: "{{.Env.MACOS_SIGN_PASSWORD}}"
+      notarize:
+        issuer_id: "{{.Env.MACOS_NOTARY_ISSUER_ID}}"
+        key_id: "{{.Env.MACOS_NOTARY_KEY_ID}}"
+        key: "{{.Env.MACOS_NOTARY_KEY}}"
+        wait: true

AGENTS.md

@@ -43,4 +43,5 @@ LocalAI follows the Linux kernel project's [guidelines for AI coding assistants]
 - **New API endpoints**: LocalAI advertises its capability surface in several independent places — swagger `@Tags`, `/api/instructions` registry, auth `RouteFeatureRegistry`, React UI `capabilities.js`, docs. Read [.agents/api-endpoints-and-auth.md](.agents/api-endpoints-and-auth.md) and follow its checklist — missing any surface means clients, admins, and the UI won't know the endpoint exists.
 - **Admin endpoints → MCP tool**: every admin endpoint that an admin would manage conversationally (install/list/edit/toggle/upgrade) MUST also be exposed as an MCP tool in `pkg/mcp/localaitools/`. The LocalAI Assistant chat modality and the standalone `local-ai mcp-server` consume that package; drift between REST and MCP is a real risk. Read [.agents/localai-assistant-mcp.md](.agents/localai-assistant-mcp.md) — the `TestToolHTTPRouteMappingComplete` test fails until you wire the new tool and update the route map.
 - **Build**: Inspect `Makefile` and `.github/workflows/` — ask the user before running long builds
+- **Backend OS coverage**: a new backend must target every OS it can build for, not just Linux. `.github/backend-matrix.yml` has two matrices — `include:` (Linux) and `includeDarwin:` (macOS / Apple Silicon). Most C/C++/GGML and many Python backends build on Darwin too — wire the `includeDarwin` entry + `backend/index.yaml` `metal:` entries, or say in the PR why an OS is unsupported. See the darwin checklist in [.agents/adding-backends.md](.agents/adding-backends.md).
 - **UI**: The active UI is the React app in `core/http/react-ui/`. The older Alpine.js/HTML UI in `core/http/static/` is pending deprecation — all new UI work goes in the React UI

Makefile

@@ -1,5 +1,5 @@
 # Disable parallel execution for backend builds
-.NOTPARALLEL: backends/diffusers backends/llama-cpp backends/turboquant backends/outetts backends/piper backends/stablediffusion-ggml backends/whisper backends/crispasr backends/parakeet-cpp backends/faster-whisper backends/silero-vad backends/local-store backends/huggingface backends/rfdetr backends/rfdetr-cpp backends/insightface backends/speaker-recognition backends/kitten-tts backends/kokoro backends/chatterbox backends/llama-cpp-darwin backends/neutts build-darwin-python-backend build-darwin-go-backend backends/mlx backends/diffuser-darwin backends/mlx-vlm backends/mlx-audio backends/mlx-distributed backends/stablediffusion-ggml-darwin backends/vllm backends/vllm-omni backends/sglang backends/moonshine backends/pocket-tts backends/qwen-tts backends/faster-qwen3-tts backends/qwen-asr backends/nemo backends/voxcpm backends/whisperx backends/ace-step backends/acestep-cpp backends/fish-speech backends/voxtral backends/opus backends/trl backends/llama-cpp-quantization backends/kokoros backends/sam3-cpp backends/qwen3-tts-cpp backends/omnivoice-cpp backends/vibevoice-cpp backends/localvqe backends/tinygrad backends/sherpa-onnx backends/ds4 backends/ds4-darwin backends/liquid-audio backends/supertonic backends/depth-anything-cpp backends/privacy-filter
+.NOTPARALLEL: backends/diffusers backends/llama-cpp backends/turboquant backends/outetts backends/piper backends/stablediffusion-ggml backends/whisper backends/crispasr backends/parakeet-cpp backends/faster-whisper backends/silero-vad backends/local-store backends/huggingface backends/rfdetr backends/rfdetr-cpp backends/insightface backends/speaker-recognition backends/kitten-tts backends/kokoro backends/chatterbox backends/llama-cpp-darwin backends/neutts build-darwin-python-backend build-darwin-go-backend backends/mlx backends/diffuser-darwin backends/mlx-vlm backends/mlx-audio backends/mlx-distributed backends/stablediffusion-ggml-darwin backends/vllm backends/vllm-omni backends/sglang backends/moonshine backends/pocket-tts backends/qwen-tts backends/faster-qwen3-tts backends/qwen-asr backends/nemo backends/voxcpm backends/whisperx backends/ace-step backends/acestep-cpp backends/fish-speech backends/voxtral backends/opus backends/trl backends/llama-cpp-quantization backends/kokoros backends/sam3-cpp backends/qwen3-tts-cpp backends/omnivoice-cpp backends/vibevoice-cpp backends/localvqe backends/tinygrad backends/sherpa-onnx backends/ds4 backends/ds4-darwin backends/liquid-audio backends/supertonic backends/depth-anything-cpp backends/privacy-filter backends/privacy-filter-darwin
 
 GOCMD=go
 GOTEST=$(GOCMD) test
@@ -103,7 +103,7 @@ COVERAGE_E2E_LABELS?=!real-models
 COVERAGE_EXCLUDE_RE?=grpc/proto/.*[.]pb[.]go
 
 
-.PHONY: all test test-coverage test-coverage-baseline test-coverage-check test-ui test-ui-coverage-baseline test-ui-coverage-check install-hooks build vendor lint lint-all
+.PHONY: all test test-coverage test-coverage-baseline test-coverage-check test-backend-cpp test-ui test-ui-coverage-baseline test-ui-coverage-check install-hooks build vendor lint lint-all
 
 all: help
 
@@ -201,6 +201,13 @@ test: prepare-test
 	OPUS_SHIM_LIBRARY=$(abspath ./pkg/opus/shim/libopusshim.so) \
 	$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --flake-attempts $(TEST_FLAKES) --fail-fast -v -r $(TEST_PATHS)
 
+## Compiles and runs the standalone C++ unit tests for the backends (pure
+## helpers that depend only on the stdlib + nlohmann/json, no full backend
+## build). Discovers every *_test.cpp under backend/cpp/ - see
+## backend/cpp/run-unit-tests.sh. Set NLOHMANN_INCLUDE to skip the header fetch.
+test-backend-cpp:
+	bash backend/cpp/run-unit-tests.sh
+
 ## Runs the core suite ($(TEST_PATHS)) with statement-coverage instrumentation
 ## and writes a merged profile to $(COVERAGE_PROFILE). Deliberately omits
 ## --fail-fast so a single failure doesn't truncate the coverage number, and
@@ -1129,6 +1136,10 @@ backends/ds4-darwin: build
 	bash ./scripts/build/ds4-darwin.sh
 	./local-ai backends install "ocifile://$(abspath ./backend-images/ds4.tar)"
 
+backends/privacy-filter-darwin: build
+	bash ./scripts/build/privacy-filter-darwin.sh
+	./local-ai backends install "ocifile://$(abspath ./backend-images/privacy-filter.tar)"
+
 build-darwin-python-backend: build
 	bash ./scripts/build/python-darwin.sh
 
@@ -1449,13 +1460,32 @@ docs: docs/static/gallery.html
 ########################################################
 
 ## fyne cross-platform build
-build-launcher-darwin: build-launcher
-	go run github.com/tiagomelo/macos-dmg-creator/cmd/createdmg@latest \
-	--appName "LocalAI" \
-	--appBinaryPath "$(LAUNCHER_BINARY_NAME)" \
-	--bundleIdentifier "com.localai.launcher" \
-	--iconPath "core/http/static/logo.png" \
-	--outputDir "dist/"
+# Build LocalAI.app from the launcher via fyne (metadata read from cmd/launcher/FyneApp.toml).
+# Signing happens via contrib/macos/sign-and-notarize.sh, which is a no-op when the signing
+# secrets are unset, so unsigned local/fork builds keep working.
+build-launcher-darwin:
+	rm -rf dist/LocalAI.app cmd/launcher/LocalAI.app
+	mkdir -p dist
+	cd cmd/launcher && go run fyne.io/tools/cmd/fyne@latest package -os darwin -icon ../../core/http/static/logo.png --executable $(LAUNCHER_BINARY_NAME)
+	mv cmd/launcher/LocalAI.app dist/LocalAI.app
+	bash contrib/macos/sign-and-notarize.sh sign dist/LocalAI.app
+
+# Wrap the (signed) app into a drag-to-Applications DMG via hdiutil, then sign the DMG.
+dmg-launcher-darwin: build-launcher-darwin
+	rm -rf dist/dmg dist/LocalAI.dmg
+	mkdir -p dist/dmg
+	cp -R dist/LocalAI.app dist/dmg/LocalAI.app
+	ln -s /Applications dist/dmg/Applications
+	hdiutil create -volname "LocalAI" -srcfolder dist/dmg -ov -format UDZO dist/LocalAI.dmg
+	bash contrib/macos/sign-and-notarize.sh sign dist/LocalAI.dmg
+
+# Submit the DMG to Apple notarization and staple the ticket (no-op without notary secrets).
+notarize-launcher-darwin: dmg-launcher-darwin
+	bash contrib/macos/sign-and-notarize.sh notarize dist/LocalAI.dmg
+
+# Single entrypoint for CI: build -> sign app -> dmg -> sign dmg -> notarize -> staple.
+release-launcher-darwin: notarize-launcher-darwin
+	@echo "dist/LocalAI.dmg is ready"
 
 build-launcher-linux:
-	cd cmd/launcher && go run fyne.io/tools/cmd/fyne@latest package -os linux -icon ../../core/http/static/logo.png --executable $(LAUNCHER_BINARY_NAME)-linux && mv launcher.tar.xz ../../$(LAUNCHER_BINARY_NAME)-linux.tar.xz
+	cd cmd/launcher && go run fyne.io/tools/cmd/fyne@latest package -os linux -icon ../../core/http/static/logo.png --executable $(LAUNCHER_BINARY_NAME)-linux && mv LocalAI.tar.xz ../../$(LAUNCHER_BINARY_NAME)-linux.tar.xz

README.md

@@ -177,6 +177,7 @@ For more details, see the [Getting Started guide](https://localai.io/basics/gett
 
 ## Latest News
 
+- **June 2026**: New native biometric backends from the LocalAI team: [voice-detect.cpp](https://github.com/mudler/voice-detect.cpp) for speaker recognition and voice analysis (ECAPA-TDNN, WeSpeaker, ERes2Net, CAM++, wav2vec2 age/gender/emotion) and [face-detect.cpp](https://github.com/mudler/face-detect.cpp) for face detection, recognition, demographics and anti-spoofing (SCRFD/ArcFace, YuNet/SFace). Both are from-scratch C++/ggml engines with no Python or onnxruntime at inference, self-contained GGUF weights, bit-exact parity with the reference, and GPU cuDNN parity, replacing the heavier Python `insightface` and `speaker-recognition` backends ([PR #10441](https://github.com/mudler/LocalAI/pull/10441)).
 - **June 2026**: New [realtime voice assistant demo](https://github.com/localai-org/localai-realtime-demo) (a tiny Go client for the Realtime API with a full talk-back voice loop and tool calling), plus [streaming of the realtime LLM / TTS / transcription pipeline stages](https://github.com/mudler/LocalAI/pull/10176) and [configurable WebRTC ICE candidates](https://github.com/mudler/LocalAI/pull/10231).
 - **June 2026**: Big speech push: the [parakeet.cpp](https://github.com/mudler/parakeet.cpp) ASR engine gains [NeMo-faithful segment timestamps](https://github.com/mudler/LocalAI/pull/10207), a [multilingual streaming Nemotron-3.5 model](https://github.com/mudler/LocalAI/pull/10199), [dynamic batching for concurrent transcription](https://github.com/mudler/LocalAI/pull/10112) and [CUDA graphs](https://github.com/mudler/LocalAI/pull/10273); the new [CrispASR backend](https://github.com/mudler/LocalAI/pull/10099) adds multi-architecture ASR + TTS, and [60 Piper TTS voices across 42 languages](https://github.com/mudler/LocalAI/pull/10296) land in the gallery (plus [per-request TTS instructions and params](https://github.com/mudler/LocalAI/pull/10172)).
 - **June 2026**: New backends and models: [locate-anything.cpp](https://github.com/mudler/LocalAI/pull/10264) for open-vocabulary object detection via ggml, [Ideogram4 image generation](https://github.com/mudler/LocalAI/pull/10201) in stablediffusion-ggml, [llama.cpp video input](https://github.com/mudler/LocalAI/pull/10216), and the [Gemma 4 QAT family with MTP speculative-decoding pairs](https://github.com/mudler/LocalAI/pull/10215). Plus an [interactive CLI chat mode](https://github.com/mudler/LocalAI/pull/10226) and [RAG source citations in agent responses](https://github.com/mudler/LocalAI/pull/10228).

backend/Dockerfile.golang

@@ -137,7 +137,7 @@ RUN <<EOT bash
             libcusolver-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION}
         if [ "${CUDA_MAJOR_VERSION}" = "13" ] && [ "arm64" = "$TARGETARCH" ]; then
             apt-get install -y --no-install-recommends \
-            libcufile-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} libcudnn9-cuda-${CUDA_MAJOR_VERSION} cuda-cupti-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} libnvjitlink-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION}
+            libcufile-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} libcudnn9-cuda-${CUDA_MAJOR_VERSION} libcudnn9-dev-cuda-${CUDA_MAJOR_VERSION} cuda-cupti-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} libnvjitlink-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION}
         fi
         apt-get clean && \
         rm -rf /var/lib/apt/lists/*

backend/cpp/ik-llama-cpp/CMakeLists.txt

@@ -1,15 +1,6 @@
-## Clip/LLaVA library for multimodal support — built locally from copied sources
-set(TARGET myclip)
-add_library(${TARGET} clip.cpp clip.h llava.cpp llava.h)
-install(TARGETS ${TARGET} LIBRARY)
-target_include_directories(myclip PUBLIC .)
-target_include_directories(myclip PUBLIC ../..)
-target_include_directories(myclip PUBLIC ../../common)
-target_link_libraries(${TARGET} PRIVATE common ggml llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if (NOT MSVC)
-    target_compile_options(${TARGET} PRIVATE -Wno-cast-qual)
-endif()
+## Multimodal support is provided by the in-tree `mtmd` library target
+## (examples/mtmd/), which the grpc-server links and includes below. clip/llava
+## were pruned upstream; the high-level mtmd_* / mtmd_helper_* API is used instead.
 
 set(TARGET grpc-server)
 set(CMAKE_CXX_STANDARD 17)
@@ -67,12 +58,16 @@ add_library(hw_grpc_proto
   ${hw_proto_hdrs} )
 
 add_executable(${TARGET} grpc-server.cpp json.hpp)
-target_link_libraries(${TARGET} PRIVATE common llama myclip ${CMAKE_THREAD_LIBS_INIT} absl::flags hw_grpc_proto
+# mtmd public headers (mtmd.h / mtmd-helper.h) live in examples/mtmd/.
+# Linking the mtmd target also propagates this include dir, but we add it
+# explicitly for clarity.
+target_include_directories(${TARGET} PRIVATE ../mtmd)
+target_link_libraries(${TARGET} PRIVATE common llama mtmd ${CMAKE_THREAD_LIBS_INIT} absl::flags hw_grpc_proto
   absl::flags_parse
   gRPC::${_REFLECTION}
   gRPC::${_GRPC_GRPCPP}
   protobuf::${_PROTOBUF_LIBPROTOBUF})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
 if(TARGET BUILD_INFO)
   add_dependencies(${TARGET} BUILD_INFO)
 endif()

backend/cpp/ik-llama-cpp/Makefile

@@ -1,5 +1,5 @@
 
-IK_LLAMA_VERSION?=6c00e87ac84404af588ad2e65935bd6f079c696f
+IK_LLAMA_VERSION?=f96eaddba8bed6a9a5e628bbf6a566775c70b49c
 LLAMA_REPO?=https://github.com/ikawrakow/ik_llama.cpp
 
 CMAKE_ARGS?=

backend/cpp/ik-llama-cpp/grpc-server.cpp

@@ -11,8 +11,8 @@
 #include <memory>
 #include <string>
 #include <getopt.h>
-#include "clip.h"
-#include "llava.h"
+#include "mtmd.h"
+#include "mtmd-helper.h"
 #include "log.h"
 #include "common.h"
 #include "json.hpp"
@@ -45,7 +45,9 @@ using backend::HealthMessage;
 
 ///// LLAMA.CPP server code below
 
-using json = nlohmann::json;
+// Match mtmd.h and ik_llama's server/common headers, which all use
+// nlohmann::ordered_json; a plain nlohmann::json alias collides at global scope.
+using json = nlohmann::ordered_json;
 
 struct server_params
 {
@@ -219,6 +221,11 @@ struct llama_client_slot
 
     // multimodal
     std::vector<slot_image> images;
+    // Full prompt with mtmd media markers (mtmd_default_marker()) substituted in
+    // place of the legacy [img-N] tags, covering the text up to and including the
+    // last image. The text after the last image is kept in params.input_suffix and
+    // decoded through the normal token path so the sampling loop is unchanged.
+    std::string mtmd_prompt;
 
     // stats
     size_t sent_count = 0;
@@ -252,14 +259,14 @@ struct llama_client_slot
 
         for (slot_image & img : images)
         {
-            free(img.image_embedding);
-            if (img.img_data) {
-                clip_image_u8_free(img.img_data);
+            if (img.bitmap) {
+                mtmd_bitmap_free(img.bitmap);
+                img.bitmap = nullptr;
             }
-            img.prefix_prompt = "";
         }
 
         images.clear();
+        mtmd_prompt = "";
     }
 
     bool has_budget(gpt_params &global_params) {
@@ -396,46 +403,13 @@ struct llama_metrics {
     }
 };
 
-struct llava_embd_batch {
-    std::vector<llama_pos>      pos;
-    std::vector<int32_t>        n_seq_id;
-    std::vector<llama_seq_id>   seq_id_0;
-    std::vector<llama_seq_id *> seq_ids;
-    std::vector<int8_t>         logits;
-    llama_batch batch;
-    llava_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
-        pos     .resize(n_tokens);
-        n_seq_id.resize(n_tokens);
-        seq_ids .resize(n_tokens + 1);
-        logits  .resize(n_tokens);
-        seq_id_0.resize(1);
-        seq_id_0[0] = seq_id;
-        seq_ids [n_tokens] = nullptr;
-        batch = {
-            /*n_tokens       =*/ n_tokens,
-            /*tokens         =*/ nullptr,
-            /*embd           =*/ embd,
-            /*pos            =*/ pos.data(),
-            /*n_seq_id       =*/ n_seq_id.data(),
-            /*seq_id         =*/ seq_ids.data(),
-            /*logits         =*/ logits.data(),
-        };
-        for (int i = 0; i < n_tokens; i++) {
-            batch.pos     [i] = pos_0 + i;
-            batch.n_seq_id[i] = 1;
-            batch.seq_id  [i] = seq_id_0.data();
-            batch.logits  [i] = false;
-        }
-    }
-};
-
 struct llama_server_context
 {
     llama_model *model = nullptr;
     llama_context *ctx = nullptr;
     const llama_vocab * vocab = nullptr;
 
-    clip_ctx *clp_ctx = nullptr;
+    mtmd_context *mctx = nullptr;
 
     gpt_params params;
 
@@ -491,11 +465,6 @@ struct llama_server_context
         if (!params.mmproj.path.empty()) {
             multimodal = true;
             LOG_INFO("Multi Modal Mode Enabled", {});
-            clp_ctx = clip_model_load(params.mmproj.path.c_str(), /*verbosity=*/ 1);
-            if(clp_ctx == nullptr) {
-                LOG_ERR("unable to load clip model: %s", params.mmproj.path.c_str());
-                return false;
-            }
 
             if (params.n_ctx < 2048) { // request larger context for the image embedding
                 params.n_ctx = 2048;
@@ -512,10 +481,24 @@ struct llama_server_context
         }
 
         if (multimodal) {
-            const int n_embd_clip = clip_n_mmproj_embd(clp_ctx);
-            const int n_embd_llm  = llama_model_n_embd(model);
-            if (n_embd_clip != n_embd_llm) {
-                LOG("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_embd_clip, n_embd_llm);
+            // mtmd_init_from_file requires the already-loaded text model, so it must
+            // run AFTER llama_init_from_gpt_params. It validates the projector
+            // against the model internally and returns nullptr on dim mismatch, so
+            // the explicit clip_n_mmproj_embd check is no longer needed.
+            mtmd_context_params mparams = mtmd_context_params_default();
+            mparams.use_gpu         = params.mmproj_use_gpu;
+            mparams.print_timings   = false;
+            mparams.n_threads       = params.n_threads_mtmd != -1 ? params.n_threads_mtmd
+                                      : params.n_threads_batch != -1 ? params.n_threads_batch
+                                                                     : params.n_threads;
+            mparams.verbosity       = GGML_LOG_LEVEL_INFO;
+            mparams.flash_attn_type = params.flash_attn ? LLAMA_FLASH_ATTN_TYPE_ENABLED
+                                                        : LLAMA_FLASH_ATTN_TYPE_DISABLED;
+            mparams.image_min_tokens = params.image_min_tokens;
+            mparams.image_max_tokens = params.image_max_tokens;
+            mctx = mtmd_init_from_file(params.mmproj.path.c_str(), model, mparams);
+            if (mctx == nullptr) {
+                LOG_ERR("unable to load multimodal projector: %s", params.mmproj.path.c_str());
                 llama_free(ctx);
                 llama_free_model(model);
                 return false;
@@ -865,8 +848,8 @@ struct llama_server_context
 
                     slot_image img_sl;
                     img_sl.id = img.count("id") != 0 ? img["id"].get<int>() : slot->images.size();
-                    img_sl.img_data = clip_image_u8_init();
-                    if (!clip_image_load_from_bytes(image_buffer.data(), image_buffer.size(), img_sl.img_data))
+                    img_sl.bitmap = mtmd_helper_bitmap_init_from_buf(mctx, image_buffer.data(), image_buffer.size());
+                    if (img_sl.bitmap == nullptr)
                     {
                         LOG_ERR("%s: failed to load image, slot_id: %d, img_sl_id: %d",
                              __func__,
@@ -879,50 +862,74 @@ struct llama_server_context
                         {"slot_id",   slot->id},
                         {"img_sl_id", img_sl.id}
                     });
-                    img_sl.request_encode_image = true;
                     slot->images.push_back(img_sl);
                 }
-                // process prompt
-                // example: system prompt [img-102] user [img-103] describe [img-134] -> [{id: 102, prefix: 'system prompt '}, {id: 103, prefix: ' user '}, {id: 134, prefix: ' describe '}]}
+                // Translate the legacy [img-N] tags into mtmd media markers, in
+                // order, and collect the matching bitmaps in marker order so they
+                // line up with the markers passed to mtmd_tokenize(). The text after
+                // the last image stays in input_suffix and is decoded through the
+                // normal token path, so the sampling loop is unchanged.
+                // example: system prompt [img-102] user [img-103] describe [img-134]
                 if (slot->images.size() > 0 && !slot->prompt.is_array())
                 {
+                    const std::string marker = mtmd_default_marker();
                     std::string prompt = slot->prompt.get<std::string>();
-                    size_t pos = 0, begin_prefix = 0;
+                    std::string built_prompt;
+                    std::vector<slot_image> ordered;
+                    size_t pos = 0, copy_from = 0;
                     std::string pattern = "[img-";
-                    while ((pos = prompt.find(pattern, pos)) != std::string::npos) {
-                        size_t end_prefix = pos;
-                        pos += pattern.length();
-                        size_t end_pos = prompt.find(']', pos);
-                        if (end_pos != std::string::npos)
-                        {
-                            std::string image_id = prompt.substr(pos, end_pos - pos);
-                            try
-                            {
-                                int img_id = std::stoi(image_id);
-                                bool found = false;
-                                for (slot_image &img : slot->images)
-                                {
-                                    if (img.id == img_id) {
-                                        found = true;
-                                        img.prefix_prompt = prompt.substr(begin_prefix, end_prefix - begin_prefix);
-                                        begin_prefix = end_pos + 1;
-                                        break;
-                                    }
-                                }
-                                if (!found) {
-                                    LOG("ERROR: Image with id: %i, not found.\n", img_id);
-                                    slot->images.clear();
-                                    return false;
-                                }
-                            } catch (const std::invalid_argument& e) {
-                                LOG("Invalid image number id in prompt\n");
-                                slot->images.clear();
-                                return false;
+
+                    auto free_images = [&]() {
+                        for (slot_image &img : slot->images) {
+                            if (img.bitmap) {
+                                mtmd_bitmap_free(img.bitmap);
+                                img.bitmap = nullptr;
                             }
                         }
+                        slot->images.clear();
+                    };
+
+                    while ((pos = prompt.find(pattern, pos)) != std::string::npos) {
+                        size_t tag_begin = pos;
+                        pos += pattern.length();
+                        size_t end_pos = prompt.find(']', pos);
+                        if (end_pos == std::string::npos) {
+                            break;
+                        }
+                        std::string image_id = prompt.substr(pos, end_pos - pos);
+                        try
+                        {
+                            int img_id = std::stoi(image_id);
+                            bool found = false;
+                            for (slot_image &img : slot->images)
+                            {
+                                if (img.id == img_id) {
+                                    found = true;
+                                    // text before this tag, then the media marker
+                                    built_prompt += prompt.substr(copy_from, tag_begin - copy_from);
+                                    built_prompt += marker;
+                                    copy_from = end_pos + 1;
+                                    ordered.push_back(img);
+                                    break;
+                                }
+                            }
+                            if (!found) {
+                                LOG("ERROR: Image with id: %i, not found.\n", img_id);
+                                free_images();
+                                return false;
+                            }
+                        } catch (const std::invalid_argument& e) {
+                            LOG("Invalid image number id in prompt\n");
+                            free_images();
+                            return false;
+                        }
+                        pos = end_pos + 1;
                     }
+                    // bitmaps are consumed in marker order by mtmd_tokenize()
+                    slot->images = ordered;
+                    slot->mtmd_prompt = built_prompt;
                     slot->prompt = "";
-                    slot->params.input_suffix = prompt.substr(begin_prefix);
+                    slot->params.input_suffix = prompt.substr(copy_from);
                     slot->params.cache_prompt = false; // multimodal doesn't support cache prompt
                 }
             }
@@ -1176,21 +1183,10 @@ struct llama_server_context
 
     bool process_images(llama_client_slot &slot) const
     {
-        for (slot_image &img : slot.images)
-        {
-            if (!img.request_encode_image)
-            {
-                continue;
-            }
-
-            if (!llava_image_embed_make_with_clip_img(clp_ctx, params.n_threads, img.img_data, &img.image_embedding, &img.image_tokens)) {
-                LOG("Error processing the given image");
-                return false;
-            }
-
-            img.request_encode_image = false;
-        }
-
+        // With the mtmd pipeline, image encoding is no longer eager: the bitmaps
+        // are tokenized and encoded together with the surrounding text inside
+        // ingest_images() via mtmd_tokenize() + mtmd_helper_eval_chunks(). This
+        // just reports whether the slot carries any images to process.
         return slot.images.size() > 0;
     }
 
@@ -1435,69 +1431,70 @@ struct llama_server_context
         }
     }
 
-    // for multiple images processing
+    // Tokenize the multimodal prompt (text interleaved with media markers) together
+    // with the slot's bitmaps, then decode the resulting chunks into the llama
+    // context via the high-level mtmd helper. The helper runs llama_decode() on the
+    // text chunks and mtmd_encode() + llama_decode() on the image chunks, handling
+    // batching and any pre/post decode setup (e.g. non-causal attention for gemma3).
+    // Advances slot.n_past by the number of positions consumed, then leaves the
+    // post-image suffix tokens in `batch` so the normal decode + sampling loop
+    // produces the first generated token.
     bool ingest_images(llama_client_slot &slot, int n_batch)
     {
-        int image_idx = 0;
-
-        while (image_idx < (int) slot.images.size())
+        if (mctx == nullptr)
         {
-            slot_image &img = slot.images[image_idx];
+            LOG("%s : multimodal context is not initialized\n", __func__);
+            return false;
+        }
 
-            // process prefix prompt
-            for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch)
-            {
-                const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
-                llama_batch batch_view = {
-                    n_tokens,
-                    batch.token    + i,
-                    nullptr,
-                    batch.pos      + i,
-                    batch.n_seq_id + i,
-                    batch.seq_id   + i,
-                    batch.logits   + i,
-                };
-                if (llama_decode(ctx, batch_view))
-                {
-                    LOG("%s : failed to eval\n", __func__);
-                    return false;
-                }
-            }
+        // bitmaps stay owned by slot.images (freed on reset()); pass non-owning ptrs
+        std::vector<const mtmd_bitmap *> bitmaps;
+        bitmaps.reserve(slot.images.size());
+        for (const slot_image &img : slot.images)
+        {
+            bitmaps.push_back(img.bitmap);
+        }
 
-            // process image with llm
-            for (int i = 0; i < img.image_tokens; i += n_batch)
-            {
-                int n_eval = img.image_tokens - i;
-                if (n_eval > n_batch)
-                {
-                    n_eval = n_batch;
-                }
+        mtmd_input_text inp_txt;
+        inp_txt.text          = slot.mtmd_prompt.c_str();
+        inp_txt.add_special   = add_bos_token;
+        inp_txt.parse_special = true;
 
-                const int n_embd = llama_model_n_embd(model);
-                float * embd = img.image_embedding + i * n_embd;
-                llava_embd_batch llava_batch = llava_embd_batch(embd, n_eval, slot.n_past, 0);
-                if (llama_decode(ctx, llava_batch.batch))
-                {
-                    LOG("%s : failed to eval image\n", __func__);
-                    return false;
-                }
-                slot.n_past += n_eval;
-            }
-            image_idx++;
+        mtmd::input_chunks chunks(mtmd_input_chunks_init());
+        int32_t res = mtmd_tokenize(mctx,
+                                    chunks.ptr.get(),
+                                    &inp_txt,
+                                    bitmaps.data(),
+                                    bitmaps.size());
+        if (res != 0)
+        {
+            LOG("%s : failed to tokenize multimodal prompt, res = %d\n", __func__, res);
+            return false;
+        }
 
-            common_batch_clear(batch);
+        const llama_pos start_pos = (llama_pos) system_tokens.size() + slot.n_past;
+        llama_pos new_n_past = start_pos;
+        if (mtmd_helper_eval_chunks(mctx,
+                                    ctx,
+                                    chunks.ptr.get(),
+                                    start_pos,
+                                    slot.id,
+                                    n_batch,
+                                    /*logits_last=*/ false,
+                                    &new_n_past) != 0)
+        {
+            LOG("%s : failed to eval multimodal chunks\n", __func__);
+            return false;
+        }
+        slot.n_past += (int32_t) (new_n_past - start_pos);
 
-            // append prefix of next image
-            const auto json_prompt = (image_idx >= (int) slot.images.size()) ?
-                slot.params.input_suffix : // no more images, then process suffix prompt
-                (json)(slot.images[image_idx].prefix_prompt);
-
-            std::vector<llama_token> append_tokens = tokenize(json_prompt, false); // has next image
-            for (int i = 0; i < (int) append_tokens.size(); ++i)
-            {
-                common_batch_add(batch, append_tokens[i], system_tokens.size() + slot.n_past, { slot.id }, true);
-                slot.n_past += 1;
-            }
+        // queue the post-image suffix text for the normal decode + sampling path
+        common_batch_clear(batch);
+        std::vector<llama_token> suffix_tokens = tokenize(slot.params.input_suffix, false);
+        for (llama_token tok : suffix_tokens)
+        {
+            common_batch_add(batch, tok, system_tokens.size() + slot.n_past, { slot.id }, false);
+            slot.n_past += 1;
         }
 
         return true;
@@ -1884,8 +1881,11 @@ struct llama_server_context
 
                     const bool has_images = process_images(slot);
 
-                    // process the prefix of first image
-                    std::vector<llama_token> prefix_tokens = has_images ? tokenize(slot.images[0].prefix_prompt, add_bos_token) : prompt_tokens;
+                    // For the multimodal path the whole pre-image / inter-image text is
+                    // tokenized and decoded inside ingest_images() via mtmd, so no prefix
+                    // tokens are queued here; the post-image suffix is appended by
+                    // ingest_images() for the normal decode + sampling loop.
+                    std::vector<llama_token> prefix_tokens = has_images ? std::vector<llama_token>() : prompt_tokens;
 
                     int32_t slot_npast = slot.n_past_se > 0 ? slot.n_past_se : slot.n_past;
 

backend/cpp/ik-llama-cpp/patches/0002-clip-ggml-quantize-chunk-user-data.patch

@@ -1,11 +0,0 @@
---- a/examples/llava/clip.cpp
-+++ b/examples/llava/clip.cpp
-@@ -2494,7 +2494,7 @@
-             }
-             new_data = work.data();
-
--            new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, n_elms/cur->ne[0], cur->ne[0], nullptr);
-+            new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, n_elms/cur->ne[0], cur->ne[0], nullptr, nullptr);
-         } else {
-             new_type = cur->type;
-             new_data = cur->data;

backend/cpp/ik-llama-cpp/prepare.sh

@@ -17,28 +17,9 @@ cp -r grpc-server.cpp llama.cpp/examples/grpc-server/
 cp -r utils.hpp llama.cpp/examples/grpc-server/
 cp -rfv llama.cpp/vendor/nlohmann/json.hpp llama.cpp/examples/grpc-server/
 
-## Copy clip/llava files for multimodal support (built as myclip library)
-cp -rfv llama.cpp/examples/llava/clip.h llama.cpp/examples/grpc-server/clip.h
-cp -rfv llama.cpp/examples/llava/clip.cpp llama.cpp/examples/grpc-server/clip.cpp
-cp -rfv llama.cpp/examples/llava/llava.cpp llama.cpp/examples/grpc-server/llava.cpp
-# Prepend llama.h include to llava.h
-echo '#include "llama.h"' > llama.cpp/examples/grpc-server/llava.h
-cat llama.cpp/examples/llava/llava.h >> llama.cpp/examples/grpc-server/llava.h
-# Copy clip-impl.h if it exists
-if [ -f llama.cpp/examples/llava/clip-impl.h ]; then
-    cp -rfv llama.cpp/examples/llava/clip-impl.h llama.cpp/examples/grpc-server/clip-impl.h
-fi
-# Copy stb_image.h
-if [ -f llama.cpp/vendor/stb/stb_image.h ]; then
-    cp -rfv llama.cpp/vendor/stb/stb_image.h llama.cpp/examples/grpc-server/stb_image.h
-elif [ -f llama.cpp/common/stb_image.h ]; then
-    cp -rfv llama.cpp/common/stb_image.h llama.cpp/examples/grpc-server/stb_image.h
-fi
-
-## Fix API compatibility in llava.cpp (llama_n_embd -> llama_model_n_embd)
-if [ -f llama.cpp/examples/grpc-server/llava.cpp ]; then
-    sed -i 's/llama_n_embd(/llama_model_n_embd(/g' llama.cpp/examples/grpc-server/llava.cpp
-fi
+## Multimodal support is provided by the `mtmd` library target (examples/mtmd/),
+## which the grpc-server links and includes directly. No source copy is needed:
+## clip/llava were pruned upstream and the high-level mtmd_* API is used instead.
 
 set +e
 if grep -q "grpc-server" llama.cpp/examples/CMakeLists.txt; then

backend/cpp/ik-llama-cpp/run.sh

@@ -2,7 +2,7 @@
 set -ex
 
 # Get the absolute current dir where the script is located
-CURDIR=$(dirname "$(realpath $0)")
+CURDIR=$(dirname "$(realpath "$0")")
 
 cd /
 
@@ -13,28 +13,28 @@ grep -e "flags" /proc/cpuinfo | head -1
 # ik_llama.cpp requires AVX2 — default to avx2 binary
 BINARY=ik-llama-cpp-avx2
 
-if [ -e $CURDIR/ik-llama-cpp-fallback ] && ! grep -q -e "\savx2\s" /proc/cpuinfo ; then
+if [ -e "$CURDIR"/ik-llama-cpp-fallback ] && ! grep -q -e "\savx2\s" /proc/cpuinfo ; then
 	echo "CPU:    AVX2   NOT found, using fallback"
 	BINARY=ik-llama-cpp-fallback
 fi
 
 # Extend ld library path with the dir where this script is located/lib
 if [ "$(uname)" == "Darwin" ]; then
-	export DYLD_LIBRARY_PATH=$CURDIR/lib:$DYLD_LIBRARY_PATH
-	#export DYLD_FALLBACK_LIBRARY_PATH=$CURDIR/lib:$DYLD_FALLBACK_LIBRARY_PATH
+	export DYLD_LIBRARY_PATH="$CURDIR"/lib:$DYLD_LIBRARY_PATH
+	#export DYLD_FALLBACK_LIBRARY_PATH="$CURDIR"/lib:$DYLD_FALLBACK_LIBRARY_PATH
 else
-	export LD_LIBRARY_PATH=$CURDIR/lib:$LD_LIBRARY_PATH
+	export LD_LIBRARY_PATH="$CURDIR"/lib:$LD_LIBRARY_PATH
 fi
 
 # If there is a lib/ld.so, use it
-if [ -f $CURDIR/lib/ld.so ]; then
+if [ -f "$CURDIR"/lib/ld.so ]; then
 	echo "Using lib/ld.so"
 	echo "Using binary: $BINARY"
-	exec $CURDIR/lib/ld.so $CURDIR/$BINARY "$@"
+	exec "$CURDIR"/lib/ld.so "$CURDIR"/$BINARY "$@"
 fi
 
 echo "Using binary: $BINARY"
-exec $CURDIR/$BINARY "$@"
+exec "$CURDIR"/$BINARY "$@"
 
 # We should never reach this point, however just in case we do, run fallback
-exec $CURDIR/ik-llama-cpp-fallback "$@"
+exec "$CURDIR"/ik-llama-cpp-fallback "$@"

backend/cpp/ik-llama-cpp/utils.hpp

@@ -11,9 +11,12 @@
 
 #include "json.hpp"
 
-#include "clip.h"
+#include "mtmd.h"
 
-using json = nlohmann::json;
+// mtmd.h and ik_llama's entire server/common stack (chat.h, server-common.h,
+// server-task.h, ...) declare `using json = nlohmann::ordered_json`, so match it
+// here: a plain `nlohmann::json` alias collides with mtmd.h's at global scope.
+using json = nlohmann::ordered_json;
 
 extern bool server_verbose;
 
@@ -111,13 +114,12 @@ struct slot_image
 {
     int32_t id;
 
-    bool request_encode_image = false;
-    float * image_embedding = nullptr;
-    int32_t image_tokens = 0;
-
-    clip_image_u8 * img_data;
-
-    std::string prefix_prompt; // before of this image
+    // mtmd bitmap (image/audio) decoded from the request buffer. Owned by the
+    // slot; freed via mtmd_bitmap_free() on reset. The high-level mtmd pipeline
+    // (mtmd_tokenize + mtmd_helper_eval_chunks) consumes these directly, so the
+    // legacy eager-encode fields (embedding/tokens) and per-image prefix prompt
+    // are no longer needed.
+    mtmd_bitmap * bitmap = nullptr;
 };
 
 // completion token output with probabilities

backend/cpp/llama-cpp/CMakeLists.txt

@@ -50,8 +50,13 @@ add_custom_command(
         "${hw_proto}"
       DEPENDS "${hw_proto}")
 
-# hw_grpc_proto
-add_library(hw_grpc_proto
+# hw_grpc_proto: force STATIC. Under the CPU_ALL_VARIANTS build BUILD_SHARED_LIBS=ON
+# (ggml/llama become shared), which would otherwise make this glue library a DSO. As a
+# DSO it references the hidden-visibility symbols in the static libprotobuf.a, which the
+# linker cannot satisfy ("hidden symbol ... in libprotobuf.a is referenced by DSO").
+# Keeping it STATIC links protobuf/gRPC directly into the grpc-server executable while
+# only ggml/llama stay shared. No effect on the static variants (already BUILD_SHARED_LIBS=OFF).
+add_library(hw_grpc_proto STATIC
   ${hw_grpc_srcs}
   ${hw_grpc_hdrs}
   ${hw_proto_srcs}
@@ -82,3 +87,18 @@ target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)
   add_dependencies(${TARGET} BUILD_INFO)
 endif()
+
+# Unit test for the message-content normalization helper (message_content.h).
+# Off by default so the normal backend build is untouched; enable with
+# -DLLAMA_GRPC_BUILD_TESTS=ON and run via ctest. It reuses llama.cpp's vendored
+# <nlohmann/json.hpp> (propagated by the common helpers library) so it has no
+# extra dependency beyond what the backend already builds against.
+option(LLAMA_GRPC_BUILD_TESTS "Build grpc-server unit tests" OFF)
+if(LLAMA_GRPC_BUILD_TESTS)
+    enable_testing()
+    add_executable(message_content_test message_content_test.cpp message_content.h)
+    target_include_directories(message_content_test PRIVATE ${CMAKE_CURRENT_SOURCE_DIR})
+    target_link_libraries(message_content_test PRIVATE ${_LLAMA_COMMON_TARGET})
+    target_compile_features(message_content_test PRIVATE cxx_std_17)
+    add_test(NAME message_content_test COMMAND message_content_test)
+endif()

backend/cpp/llama-cpp/Makefile

@@ -1,5 +1,5 @@
 
-LLAMA_VERSION?=73618f27a801c0b8614ceaf3547d3c2a99baae14
+LLAMA_VERSION?=0ed235ea2c17a19fc8238668653946721ed136fd
 LLAMA_REPO?=https://github.com/ggerganov/llama.cpp
 
 CMAKE_ARGS?=
@@ -10,8 +10,16 @@ TARGET?=--target grpc-server
 JOBS?=$(shell nproc 2>/dev/null || sysctl -n hw.ncpu 2>/dev/null || echo 1)
 ARCH?=$(shell uname -m)
 
-# Disable Shared libs as we are linking on static gRPC and we can't mix shared and static
-CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF -DLLAMA_CURL=OFF
+# Shared libs default to OFF: we link static gRPC and the avx/avx2/avx512/fallback
+# variants are fully static. The CPU_ALL_VARIANTS build flips SHARED_LIBS=ON (ggml/llama
+# become shared so the dynamic CPU backends work; gRPC stays static via its imported
+# targets). SHARED_LIBS is a make variable, not an appended -D, so it survives the
+# recursive sub-make into the VARIANT build dir (which re-parses this Makefile) instead
+# of being re-clobbered by a second -DBUILD_SHARED_LIBS=OFF. EXTRA_CMAKE_ARGS is the hook
+# the CPU_ALL_VARIANTS target uses to inject -DGGML_BACKEND_DL/-DGGML_CPU_ALL_VARIANTS.
+SHARED_LIBS?=OFF
+EXTRA_CMAKE_ARGS?=
+CMAKE_ARGS+=-DBUILD_SHARED_LIBS=$(SHARED_LIBS) -DLLAMA_CURL=OFF $(EXTRA_CMAKE_ARGS)
 
 CURRENT_MAKEFILE_DIR := $(dir $(abspath $(lastword $(MAKEFILE_LIST))))
 ifeq ($(NATIVE),false)
@@ -120,15 +128,39 @@ llama-cpp-fallback: llama.cpp
 	CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off" $(MAKE) VARIANT="llama-cpp-fallback-build" build-llama-cpp-grpc-server
 	cp -rfv $(CURRENT_MAKEFILE_DIR)/../llama-cpp-fallback-build/grpc-server llama-cpp-fallback
 
+# Single-build CPU backend using ggml's CPU_ALL_VARIANTS. Produces ONE grpc-server
+# plus a set of dlopen-able libggml-cpu-*.so (sandybridge/haswell/skylakex/...) that
+# ggml's backend registry selects from at runtime by probing host CPU features.
+# Replaces the avx/avx2/avx512/fallback multi-binary build on x86.
+#
+# CPU_ALL_VARIANTS requires GGML_BACKEND_DL, which requires BUILD_SHARED_LIBS=ON, so we
+# pass SHARED_LIBS=ON and the DL flags as make variables (NOT pre-expanded into the
+# CMAKE_ARGS env string): command-line make variables propagate through every recursive
+# sub-make, so the deepest VARIANT-dir build computes BUILD_SHARED_LIBS=ON consistently.
+# Only ggml/llama go shared - gRPC is found via its static imported targets, so the
+# grpc-server binary keeps static gRPC and only dynamically links ggml.
+#
+# TARGET adds "ggml": the per-microarch backends are runtime-dlopened, not link deps of
+# grpc-server, so they only build because each is an add_dependencies() of the ggml target.
+llama-cpp-cpu-all: llama.cpp
+	cp -rf $(CURRENT_MAKEFILE_DIR)/../llama-cpp $(CURRENT_MAKEFILE_DIR)/../llama-cpp-cpu-all-build
+	$(MAKE) -C $(CURRENT_MAKEFILE_DIR)/../llama-cpp-cpu-all-build purge
+	$(info ${GREEN}I llama-cpp build info:cpu-all-variants${RESET})
+	$(MAKE) SHARED_LIBS=ON EXTRA_CMAKE_ARGS="-DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON" TARGET="--target grpc-server --target ggml" VARIANT="llama-cpp-cpu-all-build" build-llama-cpp-grpc-server
+	cp -rfv $(CURRENT_MAKEFILE_DIR)/../llama-cpp-cpu-all-build/grpc-server llama-cpp-cpu-all
+	rm -rf ggml-shared-libs && mkdir -p ggml-shared-libs
+	find $(CURRENT_MAKEFILE_DIR)/../llama-cpp-cpu-all-build/llama.cpp/build \( -name '*.so*' -o -name '*.dylib' \) -exec cp -av {} ggml-shared-libs/ \;
+	@echo "Collected ggml shared backends:" && ls -la ggml-shared-libs/
+
 llama-cpp-grpc: llama.cpp
 	cp -rf $(CURRENT_MAKEFILE_DIR)/../llama-cpp $(CURRENT_MAKEFILE_DIR)/../llama-cpp-grpc-build
 	$(MAKE) -C $(CURRENT_MAKEFILE_DIR)/../llama-cpp-grpc-build purge
 	$(info ${GREEN}I llama-cpp build info:grpc${RESET})
-	CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_RPC=ON -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off" TARGET="--target grpc-server --target rpc-server" $(MAKE) VARIANT="llama-cpp-grpc-build" build-llama-cpp-grpc-server
+	CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_RPC=ON -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off" TARGET="--target grpc-server --target ggml-rpc-server" $(MAKE) VARIANT="llama-cpp-grpc-build" build-llama-cpp-grpc-server
 	cp -rfv $(CURRENT_MAKEFILE_DIR)/../llama-cpp-grpc-build/grpc-server llama-cpp-grpc
 
 llama-cpp-rpc-server: llama-cpp-grpc
-	cp -rf $(CURRENT_MAKEFILE_DIR)/../llama-cpp-grpc-build/llama.cpp/build/bin/rpc-server llama-cpp-rpc-server
+	cp -rf $(CURRENT_MAKEFILE_DIR)/../llama-cpp-grpc-build/llama.cpp/build/bin/ggml-rpc-server llama-cpp-rpc-server
 
 llama.cpp:
 	mkdir -p llama.cpp

backend/cpp/llama-cpp/grpc-server.cpp

@@ -30,6 +30,19 @@
 #define LOCALAI_HAS_SERVER_SCHEMA 1
 #include "server-schema.cpp"
 #endif
+// server-stream.cpp exists only in llama.cpp after the upstream refactor that
+// added the SSE stream-resumption layer (stream_session/stream_pipe_producer).
+// server-context.cpp calls into it (spipe->cleanup(), stream_aware_should_stop,
+// stream_session_attach_pipe), so its definitions must be part of this
+// translation unit or the link fails with "undefined reference to
+// stream_pipe_producer::cleanup()". The file is self-contained (its only
+// external symbols come from server-common, already pulled in above) and the
+// http route-handler factories it also defines are unused here but harmless.
+// __has_include keeps the source compatible with older pins/forks that predate
+// the split.
+#if __has_include("server-stream.cpp")
+#include "server-stream.cpp"
+#endif
 #include "server-context.cpp"
 
 // LocalAI
@@ -37,7 +50,9 @@
 #include "backend.pb.h"
 #include "backend.grpc.pb.h"
 #include "common.h"
+#include "arg.h"
 #include "chat-auto-parser.h"
+#include "message_content.h"
 #include <getopt.h>
 #include <grpcpp/ext/proto_server_reflection_plugin.h>
 #include <grpcpp/grpcpp.h>
@@ -592,6 +607,10 @@ static void params_parse(server_context& /*ctx_server*/, const backend::ModelOpt
     params.checkpoint_min_step = 256;
 #endif
 
+    // Raw upstream llama-server flags collected from any option entry that
+    // starts with '-'. Applied once after the loop via common_params_parse.
+    std::vector<std::string> extra_argv;
+
      // decode options. Options are in form optname:optvale, or if booleans only optname.
     for (int i = 0; i < request->options_size(); i++) {
         std::string opt = request->options(i);
@@ -1080,6 +1099,31 @@ static void params_parse(server_context& /*ctx_server*/, const backend::ModelOpt
                 } catch (...) {}
             }
 
+        // --- main model MoE on CPU (upstream --cpu-moe / --n-cpu-moe) ---
+        } else if (!strcmp(optname, "cpu_moe")) {
+            // Bool-style flag: keep all MoE expert weights on CPU.
+            const bool enable = (optval == NULL) ||
+                optval_str == "true" || optval_str == "1" || optval_str == "yes" ||
+                optval_str == "on" || optval_str == "enabled";
+            if (enable) {
+                params.tensor_buft_overrides.push_back(llm_ffn_exps_cpu_override());
+            }
+        } else if (!strcmp(optname, "n_cpu_moe")) {
+            if (optval != NULL) {
+                try {
+                    int n = std::stoi(optval_str);
+                    if (n < 0) n = 0;
+                    // Keep override-name storage alive for the lifetime of the
+                    // params struct (mirrors upstream arg.cpp's function-local static).
+                    static std::list<std::string> buft_overrides_main;
+                    for (int i = 0; i < n; ++i) {
+                        buft_overrides_main.push_back(llm_ffn_exps_block_regex(i));
+                        params.tensor_buft_overrides.push_back(
+                            {buft_overrides_main.back().c_str(), ggml_backend_cpu_buffer_type()});
+                    }
+                } catch (...) {}
+            }
+
         // --- draft model tensor buffer overrides (upstream --spec-draft-override-tensor) ---
         } else if (!strcmp(optname, "draft_override_tensor") || !strcmp(optname, "spec_draft_override_tensor")) {
             // Format: <tensor regex>=<buffer type>,<tensor regex>=<buffer type>,...
@@ -1111,6 +1155,30 @@ static void params_parse(server_context& /*ctx_server*/, const backend::ModelOpt
                 else { cur.push_back(c); }
             }
             if (!cur.empty()) flush(cur);
+
+        // --- generic passthrough: any entry starting with '-' is a raw
+        //     upstream llama-server flag, forwarded verbatim to the parser. ---
+        } else if (optname[0] == '-') {
+            std::string flag = optname;
+            // These flags make upstream's parser exit() (printing usage /
+            // completion), which would kill the backend process. Skip them.
+            if (flag == "-h" || flag == "--help" || flag == "--usage" ||
+                flag == "--version" || flag == "--license" ||
+                flag == "--list-devices" || flag == "-cl" ||
+                flag == "--cache-list" ||
+                flag.rfind("--completion", 0) == 0) {
+                fprintf(stderr,
+                    "[llama-cpp] ignoring passthrough flag that would exit: %s\n",
+                    flag.c_str());
+            } else {
+                extra_argv.push_back(flag);
+                // Preserve the whole value after the first ':' so embedded
+                // colons (e.g. host:port) survive strtok's truncation of optval.
+                auto colon = opt.find(':');
+                if (colon != std::string::npos) {
+                    extra_argv.push_back(opt.substr(colon + 1));
+                }
+            }
         }
     }
 
@@ -1146,27 +1214,6 @@ static void params_parse(server_context& /*ctx_server*/, const backend::ModelOpt
         }
     }
 
-    if (!params.kv_overrides.empty()) {
-        params.kv_overrides.emplace_back();
-        params.kv_overrides.back().key[0] = 0;
-    }
-
-    // tensor_buft_overrides sentinel termination (mirrors upstream common/arg.cpp).
-    // Real entries are pushed during option parsing; here we pad/terminate so the
-    // model loader sees back().pattern == nullptr (GGML_ASSERT at common.cpp:1543)
-    // and so llama_params_fit has the placeholder slots it requires.
-    {
-        const size_t ntbo = llama_max_tensor_buft_overrides();
-        while (params.tensor_buft_overrides.size() < ntbo) {
-            params.tensor_buft_overrides.push_back({nullptr, nullptr});
-        }
-    }
-    // Terminate the draft tensor_buft_overrides list with a sentinel, mirroring
-    // the main-model handling above.
-    if (!params.speculative.draft.tensor_buft_overrides.empty()) {
-        params.speculative.draft.tensor_buft_overrides.push_back({nullptr, nullptr});
-    }
-
     // TODO: Add yarn
 
     if (!request->tensorsplit().empty()) {
@@ -1259,6 +1306,69 @@ static void params_parse(server_context& /*ctx_server*/, const backend::ModelOpt
             params.sampling.grammar_triggers.push_back(std::move(trigger));
         }
     }
+
+    // Apply any raw upstream flags last so an explicit passthrough flag wins
+    // over the LocalAI-resolved field it maps to (e.g. --ctx-size beats
+    // context_size). This is the same parser llama-server itself uses.
+    if (!extra_argv.empty()) {
+        // common_params_parser_init resets a few fields for the SERVER example
+        // (n_parallel -> -1, use_color). Snapshot n_parallel so an unrelated
+        // passthrough flag can't silently clobber LocalAI's resolved value.
+        const int saved_n_parallel = params.n_parallel;
+
+        std::vector<char *> argv;
+        std::string prog = "llama-server";
+        argv.push_back(prog.data());
+        for (auto & a : extra_argv) {
+            argv.push_back(a.data());
+        }
+
+        // ctx_arg.params is a reference, so this overlays the given flags onto
+        // `params` in place. Returns false on a recoverable parse error (and
+        // self-restores params); may exit() on a hard error, exactly as
+        // passing the same bad flag to llama-server would.
+        if (!common_params_parse((int)argv.size(), argv.data(), params,
+                                 LLAMA_EXAMPLE_SERVER)) {
+            fprintf(stderr,
+                "[llama-cpp] failed to parse passthrough options; ignoring them\n");
+        }
+
+        // Restore n_parallel unless a passthrough flag explicitly set it
+        // (parser_init's reset sentinel for SERVER is -1).
+        if (params.n_parallel == -1) {
+            params.n_parallel = saved_n_parallel;
+        }
+    }
+
+    // Terminate/pad the override vectors only after BOTH the named-option loop
+    // and the generic passthrough (common_params_parse above) have pushed their
+    // real entries, so back() is the null sentinel the model loader asserts on.
+    // Running these before the passthrough let a passthrough flag (--cpu-moe,
+    // --override-tensor, --override-kv, ...) append a real entry after the
+    // sentinel: a GGML_ASSERT crash for tensor_buft_overrides, a silent drop for
+    // kv_overrides. Double-termination is harmless (the while is a no-op if the
+    // passthrough parse already padded; an extra trailing null is ignored).
+
+    if (!params.kv_overrides.empty()) {
+        params.kv_overrides.emplace_back();
+        params.kv_overrides.back().key[0] = 0;
+    }
+
+    // tensor_buft_overrides sentinel termination (mirrors upstream common/arg.cpp).
+    // Real entries are pushed during option parsing; here we pad/terminate so the
+    // model loader sees back().pattern == nullptr (GGML_ASSERT at common.cpp:1543)
+    // and so llama_params_fit has the placeholder slots it requires.
+    {
+        const size_t ntbo = llama_max_tensor_buft_overrides();
+        while (params.tensor_buft_overrides.size() < ntbo) {
+            params.tensor_buft_overrides.push_back({nullptr, nullptr});
+        }
+    }
+    // Terminate the draft tensor_buft_overrides list with a sentinel, mirroring
+    // the main-model handling above.
+    if (!params.speculative.draft.tensor_buft_overrides.empty()) {
+        params.speculative.draft.tensor_buft_overrides.push_back({nullptr, nullptr});
+    }
 }
 
 
@@ -1520,242 +1630,20 @@ public:
 
                 for (int i = 0; i < request->messages_size(); i++) {
                     const auto& msg = request->messages(i);
-                    json msg_json;
-                    msg_json["role"] = msg.role();
-
-                    bool is_last_user_msg = (i == last_user_msg_idx);
-                    bool has_images_or_audio = (request->images_size() > 0 || request->audios_size() > 0 || request->videos_size() > 0);
-
-                    // Handle content - can be string, null, or array
-                    // For multimodal content, we'll embed images/audio from separate fields
-                    if (!msg.content().empty()) {
-                        // Try to parse content as JSON to see if it's already an array
-                        json content_val;
-                        try {
-                            content_val = json::parse(msg.content());
-                            // Handle null values - convert to empty string to avoid template errors
-                            if (content_val.is_null()) {
-                                content_val = "";
-                            }
-                        } catch (const json::parse_error&) {
-                            // Not JSON, treat as plain string
-                            content_val = msg.content();
-                        }
-
-                        // If content is an object (e.g., from tool call failures), convert to string
-                        if (content_val.is_object()) {
-                            content_val = content_val.dump();
-                        }
-
-                        // If content is a string and this is the last user message with images/audio, combine them
-                        if (content_val.is_string() && is_last_user_msg && has_images_or_audio) {
-                            json content_array = json::array();
-                            // Add text first
-                            content_array.push_back({{"type", "text"}, {"text", content_val.get<std::string>()}});
-                            // Add images
-                            if (request->images_size() > 0) {
-                                for (int j = 0; j < request->images_size(); j++) {
-                                    json image_chunk;
-                                    image_chunk["type"] = "image_url";
-                                    json image_url;
-                                    image_url["url"] = "data:image/jpeg;base64," + request->images(j);
-                                    image_chunk["image_url"] = image_url;
-                                    content_array.push_back(image_chunk);
-                                }
-                            }
-                            // Add audios
-                            if (request->audios_size() > 0) {
-                                for (int j = 0; j < request->audios_size(); j++) {
-                                    json audio_chunk;
-                                    audio_chunk["type"] = "input_audio";
-                                    json input_audio;
-                                    input_audio["data"] = request->audios(j);
-                                    input_audio["format"] = "wav"; // default, could be made configurable
-                                    audio_chunk["input_audio"] = input_audio;
-                                    content_array.push_back(audio_chunk);
-                                }
-                            }
-                            if (request->videos_size() > 0) {
-                                for (int j = 0; j < request->videos_size(); j++) {
-                                    json video_chunk;
-                                    video_chunk["type"] = "input_video";
-                                    json input_video;
-                                    input_video["data"] = request->videos(j);
-                                    video_chunk["input_video"] = input_video;
-                                    content_array.push_back(video_chunk);
-                                }
-                            }
-                            msg_json["content"] = content_array;
-                        } else {
-                            // Use content as-is (already array or not last user message)
-                            // Ensure null values are converted to empty string
-                            if (content_val.is_null()) {
-                                msg_json["content"] = "";
-                            } else {
-                                msg_json["content"] = content_val;
-                            }
-                        }
-                    } else if (is_last_user_msg && has_images_or_audio) {
-                        // If no content but this is the last user message with images/audio, create content array
-                        json content_array = json::array();
-                        if (request->images_size() > 0) {
-                            for (int j = 0; j < request->images_size(); j++) {
-                                json image_chunk;
-                                image_chunk["type"] = "image_url";
-                                json image_url;
-                                image_url["url"] = "data:image/jpeg;base64," + request->images(j);
-                                image_chunk["image_url"] = image_url;
-                                content_array.push_back(image_chunk);
-                            }
-                        }
-                        if (request->audios_size() > 0) {
-                            for (int j = 0; j < request->audios_size(); j++) {
-                                json audio_chunk;
-                                audio_chunk["type"] = "input_audio";
-                                json input_audio;
-                                input_audio["data"] = request->audios(j);
-                                input_audio["format"] = "wav"; // default, could be made configurable
-                                audio_chunk["input_audio"] = input_audio;
-                                content_array.push_back(audio_chunk);
-                            }
-                        }
-                        if (request->videos_size() > 0) {
-                            for (int j = 0; j < request->videos_size(); j++) {
-                                json video_chunk;
-                                video_chunk["type"] = "input_video";
-                                json input_video;
-                                input_video["data"] = request->videos(j);
-                                video_chunk["input_video"] = input_video;
-                                content_array.push_back(video_chunk);
-                            }
-                        }
-                        msg_json["content"] = content_array;
-                    } else if (msg.role() == "tool") {
-                        // Tool role messages must have content field set, even if empty
-                        // Jinja templates expect content to be a string, not null or object
-                        SRV_INF("[CONTENT DEBUG] PredictStream: Message %d is tool role, content_empty=%d\n", i, msg.content().empty() ? 1 : 0);
-                        if (msg.content().empty()) {
-                            msg_json["content"] = "";
-                            SRV_INF("[CONTENT DEBUG] PredictStream: Message %d (tool): empty content, set to empty string\n", i);
-                        } else {
-                            SRV_INF("[CONTENT DEBUG] PredictStream: Message %d (tool): content exists: %s\n",
-                                    i, msg.content().substr(0, std::min<size_t>(200, msg.content().size())).c_str());
-                            // Content exists, parse and ensure it's a string
-                            json content_val;
-                            try {
-                                content_val = json::parse(msg.content());
-                                SRV_INF("[CONTENT DEBUG] PredictStream: Message %d (tool): parsed JSON, type=%s\n",
-                                        i, content_val.is_null() ? "null" :
-                                           content_val.is_object() ? "object" :
-                                           content_val.is_string() ? "string" :
-                                           content_val.is_array() ? "array" : "other");
-                                // Handle null values - Jinja templates expect content to be a string, not null
-                                if (content_val.is_null()) {
-                                    msg_json["content"] = "";
-                                    SRV_INF("[CONTENT DEBUG] PredictStream: Message %d (tool): null content, converted to empty string\n", i);
-                                } else if (content_val.is_object()) {
-                                    // If content is an object (e.g., from tool call failures/errors), convert to string
-                                    msg_json["content"] = content_val.dump();
-                                    SRV_INF("[CONTENT DEBUG] PredictStream: Message %d (tool): object content, converted to string: %s\n",
-                                            i, content_val.dump().substr(0, std::min<size_t>(200, content_val.dump().size())).c_str());
-                                } else if (content_val.is_string()) {
-                                    msg_json["content"] = content_val.get<std::string>();
-                                    SRV_INF("[CONTENT DEBUG] PredictStream: Message %d (tool): string content, using as-is\n", i);
-                                } else {
-                                    // For arrays or other types, convert to string
-                                    msg_json["content"] = content_val.dump();
-                                    SRV_INF("[CONTENT DEBUG] PredictStream: Message %d (tool): %s content, converted to string\n",
-                                            i, content_val.is_array() ? "array" : "other type");
-                                }
-                            } catch (const json::parse_error&) {
-                                // Not JSON, treat as plain string
-                                msg_json["content"] = msg.content();
-                                SRV_INF("[CONTENT DEBUG] PredictStream: Message %d (tool): not JSON, using as string\n", i);
-                            }
-                        }
-                    } else {
-                        // Ensure all messages have content set (fallback for any unhandled cases)
-                        // Jinja templates expect content to be present, default to empty string if not set
-                        if (!msg_json.contains("content")) {
-                            SRV_INF("[CONTENT DEBUG] PredictStream: Message %d (role=%s): no content field, adding empty string\n",
-                                    i, msg.role().c_str());
-                            msg_json["content"] = "";
-                        }
+                    llama_grpc::ReconstructedMessageInput rin;
+                    rin.role = msg.role();
+                    rin.content = msg.content();
+                    rin.name = msg.name();
+                    rin.tool_call_id = msg.tool_call_id();
+                    rin.reasoning_content = msg.reasoning_content();
+                    rin.tool_calls = msg.tool_calls();
+                    rin.is_last_user_msg = (i == last_user_msg_idx);
+                    if (rin.is_last_user_msg) {
+                        for (int j = 0; j < request->images_size(); j++) rin.images.push_back(request->images(j));
+                        for (int j = 0; j < request->audios_size(); j++) rin.audios.push_back(request->audios(j));
+                        for (int j = 0; j < request->videos_size(); j++) rin.videos.push_back(request->videos(j));
                     }
-
-                    // Add optional fields for OpenAI-compatible message format
-                    if (!msg.name().empty()) {
-                        msg_json["name"] = msg.name();
-                    }
-                    if (!msg.tool_call_id().empty()) {
-                        msg_json["tool_call_id"] = msg.tool_call_id();
-                    }
-                    if (!msg.reasoning_content().empty()) {
-                        msg_json["reasoning_content"] = msg.reasoning_content();
-                    }
-                    if (!msg.tool_calls().empty()) {
-                        // Parse tool_calls JSON string and add to message
-                        try {
-                            json tool_calls = json::parse(msg.tool_calls());
-                            msg_json["tool_calls"] = tool_calls;
-                            SRV_INF("[TOOL CALLS DEBUG] PredictStream: Message %d has tool_calls: %s\n", i, tool_calls.dump().c_str());
-                            // IMPORTANT: If message has tool_calls but content is empty or not set,
-                            // set content to space " " instead of empty string "", because llama.cpp's
-                            // common_chat_msgs_to_json_oaicompat converts empty strings to null (line 312),
-                            // which causes template errors when accessing message.content[:tool_start_length]
-                            if (!msg_json.contains("content") || (msg_json.contains("content") && msg_json["content"].is_string() && msg_json["content"].get<std::string>().empty())) {
-                                SRV_INF("[CONTENT DEBUG] PredictStream: Message %d has tool_calls but empty content, setting to space\n", i);
-                                msg_json["content"] = " ";
-                            }
-                            // Log each tool call with name and arguments
-                            if (tool_calls.is_array()) {
-                                for (size_t tc_idx = 0; tc_idx < tool_calls.size(); tc_idx++) {
-                                    const auto& tc = tool_calls[tc_idx];
-                                    std::string tool_name = "unknown";
-                                    std::string tool_args = "{}";
-                                    if (tc.contains("function")) {
-                                        const auto& func = tc["function"];
-                                        if (func.contains("name")) {
-                                            tool_name = func["name"].get<std::string>();
-                                        }
-                                        if (func.contains("arguments")) {
-                                            tool_args = func["arguments"].is_string() ?
-                                                func["arguments"].get<std::string>() :
-                                                func["arguments"].dump();
-                                        }
-                                    } else if (tc.contains("name")) {
-                                        tool_name = tc["name"].get<std::string>();
-                                        if (tc.contains("arguments")) {
-                                            tool_args = tc["arguments"].is_string() ?
-                                                tc["arguments"].get<std::string>() :
-                                                tc["arguments"].dump();
-                                        }
-                                    }
-                                    SRV_INF("[TOOL CALLS DEBUG] PredictStream: Message %d, tool_call %zu: name=%s, arguments=%s\n",
-                                            i, tc_idx, tool_name.c_str(), tool_args.c_str());
-                                }
-                            }
-                        } catch (const json::parse_error& e) {
-                            SRV_WRN("Failed to parse tool_calls JSON: %s\n", e.what());
-                        }
-                    }
-
-                    // Debug: Log final content state before adding to array
-                    if (msg_json.contains("content")) {
-                        if (msg_json["content"].is_null()) {
-                            SRV_INF("[CONTENT DEBUG] PredictStream: Message %d FINAL STATE: content is NULL - THIS WILL CAUSE ERROR!\n", i);
-                        } else {
-                            SRV_INF("[CONTENT DEBUG] PredictStream: Message %d FINAL STATE: content type=%s, has_value=%d\n",
-                                    i, msg_json["content"].is_string() ? "string" :
-                                       msg_json["content"].is_array() ? "array" :
-                                       msg_json["content"].is_object() ? "object" : "other",
-                                    msg_json["content"].is_null() ? 0 : 1);
-                        }
-                    } else {
-                        SRV_INF("[CONTENT DEBUG] PredictStream: Message %d FINAL STATE: NO CONTENT FIELD - THIS WILL CAUSE ERROR!\n", i);
-                    }
-
-                    messages_json.push_back(msg_json);
+                    messages_json.push_back(llama_grpc::build_reconstructed_message(rin));
                 }
 
                 // Final safety check: Ensure no message has null content (Jinja templates require strings)
@@ -1976,36 +1864,7 @@ public:
                 if (body_json.contains("messages") && body_json["messages"].is_array()) {
                     SRV_INF("[CONTENT DEBUG] PredictStream: Before oaicompat_chat_params_parse - checking %zu messages\n", body_json["messages"].size());
                     for (size_t idx = 0; idx < body_json["messages"].size(); idx++) {
-                        auto& msg = body_json["messages"][idx];
-                        std::string role_str = msg.contains("role") ? msg["role"].get<std::string>() : "unknown";
-                        if (msg.contains("content")) {
-                            if (msg["content"].is_null()) {
-                                SRV_INF("[CONTENT DEBUG] PredictStream: BEFORE TEMPLATE - Message %zu (role=%s) has NULL content - FIXING!\n", idx, role_str.c_str());
-                                msg["content"] = ""; // Fix null content
-                            } else if (role_str == "tool" && msg["content"].is_array()) {
-                                // Tool messages must have string content, not array
-                                // oaicompat_chat_params_parse expects tool messages to have string content
-                                SRV_INF("[CONTENT DEBUG] PredictStream: BEFORE TEMPLATE - Message %zu (role=tool) has array content, converting to string\n", idx);
-                                msg["content"] = msg["content"].dump();
-                            } else if (!msg["content"].is_string() && !msg["content"].is_array()) {
-                                // If content is object or other non-string type, convert to string for templates
-                                SRV_INF("[CONTENT DEBUG] PredictStream: BEFORE TEMPLATE - Message %zu (role=%s) content is not string/array, converting\n", idx, role_str.c_str());
-                                if (msg["content"].is_object()) {
-                                    msg["content"] = msg["content"].dump();
-                                } else {
-                                    msg["content"] = "";
-                                }
-                            } else {
-                                SRV_INF("[CONTENT DEBUG] PredictStream: BEFORE TEMPLATE - Message %zu (role=%s): content type=%s\n",
-                                        idx, role_str.c_str(),
-                                        msg["content"].is_string() ? "string" :
-                                        msg["content"].is_array() ? "array" :
-                                        msg["content"].is_object() ? "object" : "other");
-                            }
-                        } else {
-                            SRV_INF("[CONTENT DEBUG] PredictStream: BEFORE TEMPLATE - Message %zu (role=%s) MISSING content field - ADDING!\n", idx, role_str.c_str());
-                            msg["content"] = ""; // Add missing content
-                        }
+                        llama_grpc::normalize_template_message(body_json["messages"][idx]);
                     }
                 }
 
@@ -2337,264 +2196,20 @@ public:
                 SRV_INF("[CONTENT DEBUG] Predict: Processing %d messages\n", request->messages_size());
                 for (int i = 0; i < request->messages_size(); i++) {
                     const auto& msg = request->messages(i);
-                    json msg_json;
-                    msg_json["role"] = msg.role();
-
-                    SRV_INF("[CONTENT DEBUG] Predict: Message %d: role=%s, content_empty=%d, content_length=%zu\n",
-                            i, msg.role().c_str(), msg.content().empty() ? 1 : 0, msg.content().size());
-                    if (!msg.content().empty()) {
-                        SRV_INF("[CONTENT DEBUG] Predict: Message %d content (first 200 chars): %s\n",
-                                i, msg.content().substr(0, std::min<size_t>(200, msg.content().size())).c_str());
+                    llama_grpc::ReconstructedMessageInput rin;
+                    rin.role = msg.role();
+                    rin.content = msg.content();
+                    rin.name = msg.name();
+                    rin.tool_call_id = msg.tool_call_id();
+                    rin.reasoning_content = msg.reasoning_content();
+                    rin.tool_calls = msg.tool_calls();
+                    rin.is_last_user_msg = (i == last_user_msg_idx);
+                    if (rin.is_last_user_msg) {
+                        for (int j = 0; j < request->images_size(); j++) rin.images.push_back(request->images(j));
+                        for (int j = 0; j < request->audios_size(); j++) rin.audios.push_back(request->audios(j));
+                        for (int j = 0; j < request->videos_size(); j++) rin.videos.push_back(request->videos(j));
                     }
-
-                    bool is_last_user_msg = (i == last_user_msg_idx);
-                    bool has_images_or_audio = (request->images_size() > 0 || request->audios_size() > 0 || request->videos_size() > 0);
-
-                    // Handle content - can be string, null, or array
-                    // For multimodal content, we'll embed images/audio from separate fields
-                    if (!msg.content().empty()) {
-                        // Try to parse content as JSON to see if it's already an array
-                        json content_val;
-                        try {
-                            content_val = json::parse(msg.content());
-                            // Handle null values - convert to empty string to avoid template errors
-                            if (content_val.is_null()) {
-                                SRV_INF("[CONTENT DEBUG] Predict: Message %d parsed JSON is null, converting to empty string\n", i);
-                                content_val = "";
-                            }
-                        } catch (const json::parse_error&) {
-                            // Not JSON, treat as plain string
-                            content_val = msg.content();
-                        }
-
-                        // If content is an object (e.g., from tool call failures), convert to string
-                        if (content_val.is_object()) {
-                            SRV_INF("[CONTENT DEBUG] Predict: Message %d content is object, converting to string\n", i);
-                            content_val = content_val.dump();
-                        }
-
-                        // If content is a string and this is the last user message with images/audio, combine them
-                        if (content_val.is_string() && is_last_user_msg && has_images_or_audio) {
-                            json content_array = json::array();
-                            // Add text first
-                            content_array.push_back({{"type", "text"}, {"text", content_val.get<std::string>()}});
-                            // Add images
-                            if (request->images_size() > 0) {
-                                for (int j = 0; j < request->images_size(); j++) {
-                                    json image_chunk;
-                                    image_chunk["type"] = "image_url";
-                                    json image_url;
-                                    image_url["url"] = "data:image/jpeg;base64," + request->images(j);
-                                    image_chunk["image_url"] = image_url;
-                                    content_array.push_back(image_chunk);
-                                }
-                            }
-                            // Add audios
-                            if (request->audios_size() > 0) {
-                                for (int j = 0; j < request->audios_size(); j++) {
-                                    json audio_chunk;
-                                    audio_chunk["type"] = "input_audio";
-                                    json input_audio;
-                                    input_audio["data"] = request->audios(j);
-                                    input_audio["format"] = "wav"; // default, could be made configurable
-                                    audio_chunk["input_audio"] = input_audio;
-                                    content_array.push_back(audio_chunk);
-                                }
-                            }
-                            if (request->videos_size() > 0) {
-                                for (int j = 0; j < request->videos_size(); j++) {
-                                    json video_chunk;
-                                    video_chunk["type"] = "input_video";
-                                    json input_video;
-                                    input_video["data"] = request->videos(j);
-                                    video_chunk["input_video"] = input_video;
-                                    content_array.push_back(video_chunk);
-                                }
-                            }
-                            msg_json["content"] = content_array;
-                        } else {
-                            // Use content as-is (already array or not last user message)
-                            // Ensure null values are converted to empty string
-                            if (content_val.is_null()) {
-                                SRV_INF("[CONTENT DEBUG] Predict: Message %d content_val was null, setting to empty string\n", i);
-                                msg_json["content"] = "";
-                            } else {
-                                msg_json["content"] = content_val;
-                                SRV_INF("[CONTENT DEBUG] Predict: Message %d content set, type=%s\n",
-                                        i, content_val.is_string() ? "string" :
-                                           content_val.is_array() ? "array" :
-                                           content_val.is_object() ? "object" : "other");
-                            }
-                        }
-                    } else if (is_last_user_msg && has_images_or_audio) {
-                        // If no content but this is the last user message with images/audio, create content array
-                        json content_array = json::array();
-                        if (request->images_size() > 0) {
-                            for (int j = 0; j < request->images_size(); j++) {
-                                json image_chunk;
-                                image_chunk["type"] = "image_url";
-                                json image_url;
-                                image_url["url"] = "data:image/jpeg;base64," + request->images(j);
-                                image_chunk["image_url"] = image_url;
-                                content_array.push_back(image_chunk);
-                            }
-                        }
-                        if (request->audios_size() > 0) {
-                            for (int j = 0; j < request->audios_size(); j++) {
-                                json audio_chunk;
-                                audio_chunk["type"] = "input_audio";
-                                json input_audio;
-                                input_audio["data"] = request->audios(j);
-                                input_audio["format"] = "wav"; // default, could be made configurable
-                                audio_chunk["input_audio"] = input_audio;
-                                content_array.push_back(audio_chunk);
-                            }
-                        }
-                        if (request->videos_size() > 0) {
-                            for (int j = 0; j < request->videos_size(); j++) {
-                                json video_chunk;
-                                video_chunk["type"] = "input_video";
-                                json input_video;
-                                input_video["data"] = request->videos(j);
-                                video_chunk["input_video"] = input_video;
-                                content_array.push_back(video_chunk);
-                            }
-                        }
-                        msg_json["content"] = content_array;
-                        SRV_INF("[CONTENT DEBUG] Predict: Message %d created content array with media\n", i);
-                    } else if (!msg.tool_calls().empty()) {
-                        // Tool call messages may have null content, but templates expect string
-                        // IMPORTANT: Set to space " " instead of empty string "", because llama.cpp's
-                        // common_chat_msgs_to_json_oaicompat converts empty strings to null (line 312),
-                        // which causes template errors when accessing message.content[:tool_start_length]
-                        SRV_INF("[CONTENT DEBUG] Predict: Message %d has tool_calls, setting content to space (not empty string)\n", i);
-                        msg_json["content"] = " ";
-                    } else if (msg.role() == "tool") {
-                        // Tool role messages must have content field set, even if empty
-                        // Jinja templates expect content to be a string, not null or object
-                        SRV_INF("[CONTENT DEBUG] Predict: Message %d is tool role, content_empty=%d\n", i, msg.content().empty() ? 1 : 0);
-                        if (msg.content().empty()) {
-                            msg_json["content"] = "";
-                            SRV_INF("[CONTENT DEBUG] Predict: Message %d (tool): empty content, set to empty string\n", i);
-                        } else {
-                            SRV_INF("[CONTENT DEBUG] Predict: Message %d (tool): content exists: %s\n",
-                                    i, msg.content().substr(0, std::min<size_t>(200, msg.content().size())).c_str());
-                            // Content exists, parse and ensure it's a string
-                            json content_val;
-                            try {
-                                content_val = json::parse(msg.content());
-                                SRV_INF("[CONTENT DEBUG] Predict: Message %d (tool): parsed JSON, type=%s\n",
-                                        i, content_val.is_null() ? "null" :
-                                           content_val.is_object() ? "object" :
-                                           content_val.is_string() ? "string" :
-                                           content_val.is_array() ? "array" : "other");
-                                // Handle null values - Jinja templates expect content to be a string, not null
-                                if (content_val.is_null()) {
-                                    msg_json["content"] = "";
-                                    SRV_INF("[CONTENT DEBUG] Predict: Message %d (tool): null content, converted to empty string\n", i);
-                                } else if (content_val.is_object()) {
-                                    // If content is an object (e.g., from tool call failures/errors), convert to string
-                                    msg_json["content"] = content_val.dump();
-                                    SRV_INF("[CONTENT DEBUG] Predict: Message %d (tool): object content, converted to string: %s\n",
-                                            i, content_val.dump().substr(0, std::min<size_t>(200, content_val.dump().size())).c_str());
-                                } else if (content_val.is_string()) {
-                                    msg_json["content"] = content_val.get<std::string>();
-                                    SRV_INF("[CONTENT DEBUG] Predict: Message %d (tool): string content, using as-is\n", i);
-                                } else {
-                                    // For arrays or other types, convert to string
-                                    msg_json["content"] = content_val.dump();
-                                    SRV_INF("[CONTENT DEBUG] Predict: Message %d (tool): %s content, converted to string\n",
-                                            i, content_val.is_array() ? "array" : "other type");
-                                }
-                            } catch (const json::parse_error&) {
-                                // Not JSON, treat as plain string
-                                msg_json["content"] = msg.content();
-                                SRV_INF("[CONTENT DEBUG] Predict: Message %d (tool): not JSON, using as string\n", i);
-                            }
-                        }
-                    } else {
-                        // Ensure all messages have content set (fallback for any unhandled cases)
-                        // Jinja templates expect content to be present, default to empty string if not set
-                        if (!msg_json.contains("content")) {
-                            SRV_INF("[CONTENT DEBUG] Predict: Message %d (role=%s): no content field, adding empty string\n",
-                                    i, msg.role().c_str());
-                            msg_json["content"] = "";
-                        }
-                    }
-
-                    // Add optional fields for OpenAI-compatible message format
-                    if (!msg.name().empty()) {
-                        msg_json["name"] = msg.name();
-                    }
-                    if (!msg.tool_call_id().empty()) {
-                        msg_json["tool_call_id"] = msg.tool_call_id();
-                    }
-                    if (!msg.reasoning_content().empty()) {
-                        msg_json["reasoning_content"] = msg.reasoning_content();
-                    }
-                    if (!msg.tool_calls().empty()) {
-                        // Parse tool_calls JSON string and add to message
-                        try {
-                            json tool_calls = json::parse(msg.tool_calls());
-                            msg_json["tool_calls"] = tool_calls;
-                            SRV_INF("[TOOL CALLS DEBUG] Predict: Message %d has tool_calls: %s\n", i, tool_calls.dump().c_str());
-                            // IMPORTANT: If message has tool_calls but content is empty or not set,
-                            // set content to space " " instead of empty string "", because llama.cpp's
-                            // common_chat_msgs_to_json_oaicompat converts empty strings to null (line 312),
-                            // which causes template errors when accessing message.content[:tool_start_length]
-                            if (!msg_json.contains("content") || (msg_json.contains("content") && msg_json["content"].is_string() && msg_json["content"].get<std::string>().empty())) {
-                                SRV_INF("[CONTENT DEBUG] Predict: Message %d has tool_calls but empty content, setting to space\n", i);
-                                msg_json["content"] = " ";
-                            }
-                            // Log each tool call with name and arguments
-                            if (tool_calls.is_array()) {
-                                for (size_t tc_idx = 0; tc_idx < tool_calls.size(); tc_idx++) {
-                                    const auto& tc = tool_calls[tc_idx];
-                                    std::string tool_name = "unknown";
-                                    std::string tool_args = "{}";
-                                    if (tc.contains("function")) {
-                                        const auto& func = tc["function"];
-                                        if (func.contains("name")) {
-                                            tool_name = func["name"].get<std::string>();
-                                        }
-                                        if (func.contains("arguments")) {
-                                            tool_args = func["arguments"].is_string() ?
-                                                func["arguments"].get<std::string>() :
-                                                func["arguments"].dump();
-                                        }
-                                    } else if (tc.contains("name")) {
-                                        tool_name = tc["name"].get<std::string>();
-                                        if (tc.contains("arguments")) {
-                                            tool_args = tc["arguments"].is_string() ?
-                                                tc["arguments"].get<std::string>() :
-                                                tc["arguments"].dump();
-                                        }
-                                    }
-                                    SRV_INF("[TOOL CALLS DEBUG] Predict: Message %d, tool_call %zu: name=%s, arguments=%s\n",
-                                            i, tc_idx, tool_name.c_str(), tool_args.c_str());
-                                }
-                            }
-                        } catch (const json::parse_error& e) {
-                            SRV_WRN("Failed to parse tool_calls JSON: %s\n", e.what());
-                        }
-                    }
-
-                    // Debug: Log final content state before adding to array
-                    if (msg_json.contains("content")) {
-                        if (msg_json["content"].is_null()) {
-                            SRV_INF("[CONTENT DEBUG] Predict: Message %d FINAL STATE: content is NULL - THIS WILL CAUSE ERROR!\n", i);
-                        } else {
-                            SRV_INF("[CONTENT DEBUG] Predict: Message %d FINAL STATE: content type=%s, has_value=%d\n",
-                                    i, msg_json["content"].is_string() ? "string" :
-                                       msg_json["content"].is_array() ? "array" :
-                                       msg_json["content"].is_object() ? "object" : "other",
-                                    msg_json["content"].is_null() ? 0 : 1);
-                        }
-                    } else {
-                        SRV_INF("[CONTENT DEBUG] Predict: Message %d FINAL STATE: NO CONTENT FIELD - THIS WILL CAUSE ERROR!\n", i);
-                    }
-
-                    messages_json.push_back(msg_json);
+                    messages_json.push_back(llama_grpc::build_reconstructed_message(rin));
                 }
 
                 // Final safety check: Ensure no message has null content (Jinja templates require strings)
@@ -2815,36 +2430,7 @@ public:
                 if (body_json.contains("messages") && body_json["messages"].is_array()) {
                     SRV_INF("[CONTENT DEBUG] Predict: Before oaicompat_chat_params_parse - checking %zu messages\n", body_json["messages"].size());
                     for (size_t idx = 0; idx < body_json["messages"].size(); idx++) {
-                        auto& msg = body_json["messages"][idx];
-                        std::string role_str = msg.contains("role") ? msg["role"].get<std::string>() : "unknown";
-                        if (msg.contains("content")) {
-                            if (msg["content"].is_null()) {
-                                SRV_INF("[CONTENT DEBUG] Predict: BEFORE TEMPLATE - Message %zu (role=%s) has NULL content - FIXING!\n", idx, role_str.c_str());
-                                msg["content"] = ""; // Fix null content
-                            } else if (role_str == "tool" && msg["content"].is_array()) {
-                                // Tool messages must have string content, not array
-                                // oaicompat_chat_params_parse expects tool messages to have string content
-                                SRV_INF("[CONTENT DEBUG] Predict: BEFORE TEMPLATE - Message %zu (role=tool) has array content, converting to string\n", idx);
-                                msg["content"] = msg["content"].dump();
-                            } else if (!msg["content"].is_string() && !msg["content"].is_array()) {
-                                // If content is object or other non-string type, convert to string for templates
-                                SRV_INF("[CONTENT DEBUG] Predict: BEFORE TEMPLATE - Message %zu (role=%s) content is not string/array, converting\n", idx, role_str.c_str());
-                                if (msg["content"].is_object()) {
-                                    msg["content"] = msg["content"].dump();
-                                } else {
-                                    msg["content"] = "";
-                                }
-                            } else {
-                                SRV_INF("[CONTENT DEBUG] Predict: BEFORE TEMPLATE - Message %zu (role=%s): content type=%s\n",
-                                        idx, role_str.c_str(),
-                                        msg["content"].is_string() ? "string" :
-                                        msg["content"].is_array() ? "array" :
-                                        msg["content"].is_object() ? "object" : "other");
-                            }
-                        } else {
-                            SRV_INF("[CONTENT DEBUG] Predict: BEFORE TEMPLATE - Message %zu (role=%s) MISSING content field - ADDING!\n", idx, role_str.c_str());
-                            msg["content"] = ""; // Add missing content
-                        }
+                        llama_grpc::normalize_template_message(body_json["messages"][idx]);
                     }
                 }
 

backend/cpp/llama-cpp/message_content.h

@@ -0,0 +1,192 @@
+#pragma once
+
+#include <string>
+#include <vector>
+
+#include <nlohmann/json.hpp>
+
+namespace llama_grpc {
+
+// Normalizes a proto message's content string into the JSON value used when
+// reconstructing OpenAI-format messages for the tokenizer (jinja) template.
+//
+// Shared by the streaming (PredictStream) and non-streaming (Predict) message
+// reconstruction paths so the two cannot drift.
+//
+// LocalAI's Go layer (schema.Messages.ToProto) always sends content as a plain
+// text string; multimodal media travels in separate proto fields, never inside
+// content. So user/system/developer content is *only ever* opaque text and must
+// NOT be JSON-sniffed: a prompt that merely looks like JSON (e.g. an ingredient
+// list ["1/4 cup sugar", ...]) would otherwise be reinterpreted as structured
+// content parts and rejected by oaicompat_chat_params_parse with
+// "unsupported content[].type" (https://github.com/mudler/LocalAI/issues/10524).
+// (developer is OpenAI's modern system alias - same "human-authored text" nature.)
+//
+// For assistant/tool messages we still collapse a literal JSON null/object
+// (tool-call bookkeeping) to a string, but we never turn a plain string into an
+// array/scalar. The array defense is therefore role-independent (arrays/scalars
+// fall through for every role); the role gate only governs the null/object case.
+inline nlohmann::ordered_json normalize_message_content(const std::string& role,
+                                                        const std::string& content) {
+    nlohmann::ordered_json content_val = content;
+    if (role != "user" && role != "system" && role != "developer") {
+        try {
+            nlohmann::ordered_json parsed = nlohmann::ordered_json::parse(content);
+            if (parsed.is_null()) {
+                content_val = "";
+            } else if (parsed.is_object()) {
+                content_val = parsed.dump();
+            }
+            // arrays / scalars: keep the original plain-text string as-is
+        } catch (const nlohmann::ordered_json::parse_error&) {
+            // Not JSON, already the plain string
+        }
+    }
+    return content_val;
+}
+
+// Final safety pass applied to each reconstructed OpenAI message right before it
+// is handed to oaicompat_chat_params_parse (jinja templating). Jinja templates
+// assume content is a string: a literal null breaks slicing such as
+// message.content[:N] (#7324), and a tool message with array content is rejected
+// (#7528). A multimodal user message legitimately carries a typed-part array
+// ({type:text}, {type:image_url}, ...), which must be left intact. Shared by the
+// streaming and non-streaming paths so this invariant cannot drift between them.
+inline void normalize_template_message(nlohmann::ordered_json& msg) {
+    if (!msg.contains("content")) {
+        msg["content"] = ""; // templates expect the field to exist
+        return;
+    }
+    nlohmann::ordered_json& content = msg["content"];
+    const std::string role = (msg.contains("role") && msg["role"].is_string())
+                                 ? msg["role"].get<std::string>()
+                                 : std::string();
+    if (content.is_null()) {
+        content = ""; // #7324: null would crash content[:N] slicing
+    } else if (role == "tool" && content.is_array()) {
+        content = content.dump(); // #7528: tool messages must have string content
+    } else if (!content.is_string() && !content.is_array()) {
+        if (content.is_object()) {
+            content = content.dump(); // tool-call bookkeeping object -> string
+        } else {
+            content = ""; // other scalar (number/bool) -> empty
+        }
+    }
+    // string, or a non-tool (multimodal) typed-part array: leave untouched
+}
+
+// One proto message's data, flattened to plain types so the reconstruction logic
+// can be shared and unit-tested without protobuf. The streaming and non-streaming
+// predict paths both populate this from proto::Message + the request's media.
+struct ReconstructedMessageInput {
+    std::string role;
+    std::string content;            // proto.Message.content (always a plain string)
+    std::string name;
+    std::string tool_call_id;
+    std::string reasoning_content;
+    std::string tool_calls;         // tool_calls as a JSON string, or empty
+    bool is_last_user_msg = false;  // attach request media to this message
+    std::vector<std::string> images; // base64 (jpeg)
+    std::vector<std::string> audios; // base64 (wav)
+    std::vector<std::string> videos; // base64
+};
+
+// Appends the request's media as OpenAI typed content parts. Imperative (not
+// brace-init) to avoid nlohmann's object-vs-array initializer-list ambiguity.
+inline void append_media_parts(nlohmann::ordered_json& content_array,
+                               const std::vector<std::string>& images,
+                               const std::vector<std::string>& audios,
+                               const std::vector<std::string>& videos) {
+    for (const auto& img : images) {
+        nlohmann::ordered_json image_chunk;
+        image_chunk["type"] = "image_url";
+        nlohmann::ordered_json image_url;
+        image_url["url"] = "data:image/jpeg;base64," + img;
+        image_chunk["image_url"] = image_url;
+        content_array.push_back(image_chunk);
+    }
+    for (const auto& aud : audios) {
+        nlohmann::ordered_json audio_chunk;
+        audio_chunk["type"] = "input_audio";
+        nlohmann::ordered_json input_audio;
+        input_audio["data"] = aud;
+        input_audio["format"] = "wav"; // default; could be made configurable
+        audio_chunk["input_audio"] = input_audio;
+        content_array.push_back(audio_chunk);
+    }
+    for (const auto& vid : videos) {
+        nlohmann::ordered_json video_chunk;
+        video_chunk["type"] = "input_video";
+        nlohmann::ordered_json input_video;
+        input_video["data"] = vid;
+        video_chunk["input_video"] = input_video;
+        content_array.push_back(video_chunk);
+    }
+}
+
+// Reconstructs a single OpenAI-format message (the object fed to
+// oaicompat_chat_params_parse) from a proto message. Shared by PredictStream and
+// Predict so the content/multimodal/tool_calls handling cannot drift between the
+// two stream modes (it previously lived as two ~150-line copies with a redundant
+// Predict-only tool_calls->" " branch). Guarantees content is always a string or
+// a typed-part array, never null/missing.
+inline nlohmann::ordered_json build_reconstructed_message(const ReconstructedMessageInput& in) {
+    nlohmann::ordered_json msg_json;
+    msg_json["role"] = in.role;
+    const bool has_media = !in.images.empty() || !in.audios.empty() || !in.videos.empty();
+
+    if (!in.content.empty()) {
+        nlohmann::ordered_json content_val = normalize_message_content(in.role, in.content);
+        if (content_val.is_string() && in.is_last_user_msg && has_media) {
+            // Last user message + media: build a typed-part array (text first).
+            nlohmann::ordered_json content_array = nlohmann::ordered_json::array();
+            nlohmann::ordered_json text_part;
+            text_part["type"] = "text";
+            text_part["text"] = content_val.get<std::string>();
+            content_array.push_back(text_part);
+            append_media_parts(content_array, in.images, in.audios, in.videos);
+            msg_json["content"] = content_array;
+        } else if (content_val.is_null()) {
+            msg_json["content"] = "";
+        } else {
+            msg_json["content"] = content_val;
+        }
+    } else if (in.is_last_user_msg && has_media) {
+        // No text but media on the last user message: media-only typed array.
+        nlohmann::ordered_json content_array = nlohmann::ordered_json::array();
+        append_media_parts(content_array, in.images, in.audios, in.videos);
+        msg_json["content"] = content_array;
+    } else {
+        // Empty content (any role, incl. tool/assistant): templates need a string.
+        msg_json["content"] = "";
+    }
+
+    if (!in.name.empty()) {
+        msg_json["name"] = in.name;
+    }
+    if (!in.tool_call_id.empty()) {
+        msg_json["tool_call_id"] = in.tool_call_id;
+    }
+    if (!in.reasoning_content.empty()) {
+        msg_json["reasoning_content"] = in.reasoning_content;
+    }
+    if (!in.tool_calls.empty()) {
+        try {
+            nlohmann::ordered_json tool_calls = nlohmann::ordered_json::parse(in.tool_calls);
+            msg_json["tool_calls"] = tool_calls;
+            // tool_calls + empty/blank content: use " " not "", because llama.cpp's
+            // common_chat_msgs_to_json_oaicompat turns "" into null, which breaks
+            // templates that slice message.content[:tool_start_length] (#7324).
+            if (!msg_json.contains("content") ||
+                (msg_json["content"].is_string() && msg_json["content"].get<std::string>().empty())) {
+                msg_json["content"] = " ";
+            }
+        } catch (const nlohmann::ordered_json::parse_error&) {
+            // Malformed tool_calls JSON: leave content as-is (prior behavior).
+        }
+    }
+
+    return msg_json;
+}
+
+}  // namespace llama_grpc

backend/cpp/llama-cpp/message_content_test.cpp

@@ -0,0 +1,234 @@
+// Unit tests for the shared message-reconstruction helpers (message_content.h).
+//
+// Build & run standalone (nlohmann/json single header on the include path):
+//   g++ -std=c++17 -I<dir-with-nlohmann> message_content_test.cpp -o t && ./t
+// or via CMake: -DLLAMA_GRPC_BUILD_TESTS=ON then ctest.
+//
+// Regression coverage for:
+//   #10524 - a user/system prompt that is itself a JSON-array string must stay
+//            plain text, never be reinterpreted as OpenAI structured parts.
+//   #7324  - assistant/tool null content -> "" (templates slice content[:N]);
+//            assistant+tool_calls+empty content -> " " (not "", which becomes null).
+//   #7528  - tool message array content must reach the template as a string.
+//   multimodal - last user message text + media -> typed-part array, media kept.
+
+#include <cassert>
+#include <iostream>
+#include <string>
+
+#include "message_content.h"
+
+using nlohmann::ordered_json;
+using llama_grpc::normalize_message_content;
+using llama_grpc::normalize_template_message;
+using llama_grpc::build_reconstructed_message;
+using llama_grpc::ReconstructedMessageInput;
+
+static int failures = 0;
+
+static void check(bool ok, const std::string& name, const std::string& detail = "") {
+    if (!ok) {
+        std::cerr << "FAIL " << name << (detail.empty() ? "" : ": " + detail) << "\n";
+        failures++;
+    }
+}
+
+// ---- normalize_message_content -------------------------------------------
+
+static void expect_norm_string(const char* name, const std::string& role,
+                               const std::string& content, const std::string& want) {
+    auto got = normalize_message_content(role, content);
+    if (!got.is_string()) {
+        check(false, name, "expected a JSON string, got " +
+                               std::string(got.is_array() ? "array" : got.is_object() ? "object" : "other") +
+                               " (" + got.dump() + ")");
+        return;
+    }
+    check(got.get<std::string>() == want, name, "expected \"" + want + "\", got \"" + got.get<std::string>() + "\"");
+}
+
+static void test_normalize() {
+    const std::string ingredients = R"(["1/4 cup brown sugar, packed","1 pound ground beef"])";
+
+    // #10524 - JSON-array text must stay a string. Role-INDEPENDENT array defense.
+    for (const char* role : {"user", "system", "developer", "function", "assistant", "tool"}) {
+        expect_norm_string((std::string("json_array_stays_text:") + role).c_str(), role, ingredients, ingredients);
+    }
+
+    // #10524 - user/system/developer JSON-object text stays verbatim (NOT re-dumped).
+    expect_norm_string("user_json_object_verbatim", "user", R"({"a":1})", R"({"a":1})");
+    expect_norm_string("system_json_object_verbatim", "system", R"({"a":1})", R"({"a":1})");
+    expect_norm_string("developer_json_object_verbatim", "developer", R"({"a":1})", R"({"a":1})");
+
+    // Plain text unchanged for all roles.
+    expect_norm_string("user_plain_text", "user", "hello world", "hello world");
+    expect_norm_string("assistant_non_json_text_kept", "assistant", "hi [unclosed", "hi [unclosed");
+
+    // #7324 boundary - user/system/developer literal "null" preserved (never parsed).
+    expect_norm_string("user_literal_null_stays", "user", "null", "null");
+    expect_norm_string("system_literal_null_stays", "system", "null", "null");
+    expect_norm_string("developer_literal_null_stays", "developer", "null", "null");
+
+    // #7324 - assistant/tool literal null collapses to empty string.
+    expect_norm_string("assistant_null_to_empty", "assistant", "null", "");
+    expect_norm_string("tool_null_to_empty", "tool", "null", "");
+
+    // #7324/#7528 - assistant/tool object bookkeeping stringified (stays a string).
+    check(normalize_message_content("assistant", R"({"tool":"x"})").is_string(), "assistant_object_stringified");
+    check(normalize_message_content("tool", R"({"error":"boom"})").is_string(), "tool_object_stringified");
+
+    // #10524-family - a bare scalar that parses as a JSON number stays the string.
+    expect_norm_string("assistant_scalar_number_stays_string", "assistant", "42", "42");
+
+    // baseline - empty content stays empty.
+    expect_norm_string("user_empty_stays_empty", "user", "", "");
+}
+
+// ---- normalize_template_message (BEFORE TEMPLATE sanitizer) ---------------
+
+static void test_template_sanitizer() {
+    // #7528 - a tool message with an ACTUAL array becomes a string.
+    {
+        ordered_json msg = {{"role", "tool"}, {"content", ordered_json::array({{{"type", "text"}, {"text", "r"}}})}};
+        normalize_template_message(msg);
+        check(msg["content"].is_string(), "before_template_tool_array_to_string", "got " + msg["content"].dump());
+    }
+    // #7324 - null content -> "" for any role.
+    {
+        ordered_json msg = {{"role", "assistant"}, {"content", nullptr}};
+        normalize_template_message(msg);
+        check(msg["content"].is_string() && msg["content"] == "", "before_template_null_to_empty");
+    }
+    // object content -> dumped string (would otherwise throw at the template).
+    {
+        ordered_json msg = {{"role", "assistant"}, {"content", {{"x", 1}}}};
+        normalize_template_message(msg);
+        check(msg["content"].is_string(), "before_template_object_to_string", "got " + msg["content"].dump());
+    }
+    // missing content field -> "".
+    {
+        ordered_json msg = {{"role", "user"}};
+        normalize_template_message(msg);
+        check(msg.contains("content") && msg["content"] == "", "before_template_missing_to_empty");
+    }
+    // multimodal: a well-typed user array must be left UNTOUCHED (role!=tool).
+    {
+        ordered_json parts = ordered_json::array();
+        parts.push_back({{"type", "text"}, {"text", "x"}});
+        ordered_json img; img["type"] = "image_url"; img["image_url"] = {{"url", "data:..."}};
+        parts.push_back(img);
+        ordered_json msg = {{"role", "user"}, {"content", parts}};
+        normalize_template_message(msg);
+        check(msg["content"].is_array() && msg["content"].size() == 2, "before_template_user_typed_array_preserved",
+              "got " + msg["content"].dump());
+    }
+    // a plain string is left untouched.
+    {
+        ordered_json msg = {{"role", "user"}, {"content", "hello"}};
+        normalize_template_message(msg);
+        check(msg["content"] == "hello", "before_template_string_untouched");
+    }
+}
+
+// ---- build_reconstructed_message ----------------------------------------
+
+static void test_reconstruction() {
+    const std::string ingredients = R"(["1/4 cup brown sugar","1 pound ground beef"])";
+
+    // #10524 end-state - user JSON-array text, no media -> string content.
+    {
+        ReconstructedMessageInput in;
+        in.role = "user"; in.content = ingredients;
+        auto m = build_reconstructed_message(in);
+        check(m["content"].is_string() && m["content"] == ingredients, "recon_user_json_array_string",
+              "got " + m["content"].dump());
+    }
+    // multimodal - user text + one image on last user msg -> typed array, image kept.
+    {
+        ReconstructedMessageInput in;
+        in.role = "user"; in.content = ingredients; in.is_last_user_msg = true;
+        in.images.push_back("BASE64IMG");
+        auto m = build_reconstructed_message(in);
+        check(m["content"].is_array() && m["content"].size() == 2, "recon_multimodal_text_plus_image",
+              "got " + m["content"].dump());
+        check(m["content"][0]["type"] == "text" && m["content"][0]["text"] == ingredients, "recon_multimodal_text_first");
+        check(m["content"][1]["type"] == "image_url", "recon_multimodal_image_kept");
+    }
+    // multimodal media-only - empty text + image on last user msg.
+    {
+        ReconstructedMessageInput in;
+        in.role = "user"; in.content = ""; in.is_last_user_msg = true;
+        in.images.push_back("BASE64IMG");
+        auto m = build_reconstructed_message(in);
+        check(m["content"].is_array() && m["content"].size() == 1 && m["content"][0]["type"] == "image_url",
+              "recon_media_only", "got " + m["content"].dump());
+    }
+    // #7528 - tool array-string content stays a string.
+    {
+        ReconstructedMessageInput in;
+        in.role = "tool"; in.content = R"(["a","b"])"; in.tool_call_id = "call_1";
+        auto m = build_reconstructed_message(in);
+        check(m["content"].is_string() && m["content"] == R"(["a","b"])", "recon_tool_array_string",
+              "got " + m["content"].dump());
+        check(m["tool_call_id"] == "call_1", "recon_tool_call_id_set");
+    }
+    // tool empty content -> "".
+    {
+        ReconstructedMessageInput in;
+        in.role = "tool"; in.content = "";
+        auto m = build_reconstructed_message(in);
+        check(m["content"].is_string() && m["content"] == "", "recon_tool_empty_to_string");
+    }
+    // #7324 - assistant + tool_calls + empty content -> " " (single space, not "").
+    {
+        ReconstructedMessageInput in;
+        in.role = "assistant"; in.content = "";
+        in.tool_calls = R"([{"id":"c1","type":"function","function":{"name":"f","arguments":"{}"}}])";
+        auto m = build_reconstructed_message(in);
+        check(m["content"].is_string() && m["content"] == " ", "recon_toolcalls_empty_content_space",
+              "got " + m["content"].dump());
+        check(m["tool_calls"].is_array() && m["tool_calls"].size() == 1, "recon_toolcalls_parsed");
+    }
+    // assistant + tool_calls + real content keeps the content.
+    {
+        ReconstructedMessageInput in;
+        in.role = "assistant"; in.content = "I'll call f";
+        in.tool_calls = R"([{"id":"c1","type":"function","function":{"name":"f","arguments":"{}"}}])";
+        auto m = build_reconstructed_message(in);
+        check(m["content"] == "I'll call f", "recon_toolcalls_with_content_kept");
+    }
+    // assistant null content -> "".
+    {
+        ReconstructedMessageInput in;
+        in.role = "assistant"; in.content = "null";
+        auto m = build_reconstructed_message(in);
+        check(m["content"] == "", "recon_assistant_null_to_empty");
+    }
+    // malformed tool_calls JSON must not throw; content preserved.
+    {
+        ReconstructedMessageInput in;
+        in.role = "assistant"; in.content = "hi"; in.tool_calls = "{not json";
+        auto m = build_reconstructed_message(in);
+        check(m["content"] == "hi" && !m.contains("tool_calls"), "recon_malformed_toolcalls_safe");
+    }
+    // optional fields: name + reasoning carried through.
+    {
+        ReconstructedMessageInput in;
+        in.role = "tool"; in.content = "result"; in.name = "get_weather"; in.reasoning_content = "thinking";
+        auto m = build_reconstructed_message(in);
+        check(m["name"] == "get_weather" && m["reasoning_content"] == "thinking", "recon_optional_fields");
+    }
+}
+
+int main() {
+    test_normalize();
+    test_template_sanitizer();
+    test_reconstruction();
+
+    if (failures == 0) {
+        std::cout << "OK: all message_content tests passed\n";
+        return 0;
+    }
+    std::cerr << failures << " test(s) failed\n";
+    return 1;
+}

backend/cpp/llama-cpp/package.sh

@@ -14,6 +14,22 @@ mkdir -p $CURDIR/package/lib
 cp -avrf $CURDIR/llama-cpp-* $CURDIR/package/
 cp -rfv $CURDIR/run.sh $CURDIR/package/
 
+# Bundle the ggml shared backends produced by the CPU_ALL_VARIANTS build (libggml-base.so,
+# libggml.so, libllama.so and the per-microarch libggml-cpu-*.so), all into package/lib.
+#
+# Two distinct resolution mechanisms both land here:
+#   - NEEDED deps (libggml-base/libggml/libllama): resolved by the dynamic linker via the
+#     LD_LIBRARY_PATH=$CURDIR/lib that run.sh exports.
+#   - The per-microarch libggml-cpu-*.so are NOT linked; ggml *discovers* them at runtime by
+#     scanning the executable's own directory (readlink /proc/self/exe). run.sh launches via
+#     the bundled $CURDIR/lib/ld.so, so /proc/self/exe -> .../lib/ld.so and ggml scans lib/.
+#     That is why the variants must sit in lib/ (next to ld.so), not just on the link path.
+# No-op on builds (arm64/darwin) that don't produce the all-variants set.
+if [ -d "$CURDIR/ggml-shared-libs" ]; then
+    echo "Bundling ggml shared backends (CPU_ALL_VARIANTS)..."
+    cp -avf $CURDIR/ggml-shared-libs/*.so* $CURDIR/package/lib/
+fi
+
 # Detect architecture and copy appropriate libraries
 if [ -f "/lib64/ld-linux-x86-64.so.2" ]; then
     # x86_64 architecture

backend/cpp/llama-cpp/prepare.sh

@@ -18,6 +18,10 @@ done
 
 cp -r CMakeLists.txt llama.cpp/tools/grpc-server/
 cp -r grpc-server.cpp llama.cpp/tools/grpc-server/
+# Shared message-reconstruction helpers (included by grpc-server.cpp) and their
+# unit test (compiled only when -DLLAMA_GRPC_BUILD_TESTS=ON).
+cp -r message_content.h llama.cpp/tools/grpc-server/
+cp -r message_content_test.cpp llama.cpp/tools/grpc-server/
 cp -rfv llama.cpp/vendor/nlohmann/json.hpp llama.cpp/tools/grpc-server/
 cp -rfv llama.cpp/vendor/cpp-httplib/httplib.h llama.cpp/tools/grpc-server/
 

backend/cpp/llama-cpp/run.sh

@@ -2,7 +2,7 @@
 set -ex
 
 # Get the absolute current dir where the script is located
-CURDIR=$(dirname "$(realpath $0)")
+CURDIR=$(dirname "$(realpath "$0")")
 
 cd /
 
@@ -12,55 +12,41 @@ grep -e "flags" /proc/cpuinfo | head -1
 
 BINARY=llama-cpp-fallback
 
-if grep -q -e "\savx\s" /proc/cpuinfo ; then
-	echo "CPU:    AVX    found OK"
-	if [ -e $CURDIR/llama-cpp-avx ]; then
-		BINARY=llama-cpp-avx
-	fi
-fi
-
-if grep -q -e "\savx2\s" /proc/cpuinfo ; then
-	echo "CPU:    AVX2   found OK"
-	if [ -e $CURDIR/llama-cpp-avx2 ]; then
-		BINARY=llama-cpp-avx2
-	fi
-fi
-
-# Check avx 512
-if grep -q -e "\savx512f\s" /proc/cpuinfo ; then
-	echo "CPU:    AVX512F found OK"
-	if [ -e $CURDIR/llama-cpp-avx512 ]; then
-		BINARY=llama-cpp-avx512
-	fi
+# CPU images (x86, arm64, darwin) ship a single llama-cpp-cpu-all built with ggml
+# CPU_ALL_VARIANTS: ggml's backend registry dlopens the best libggml-cpu-*.so for this
+# host, so no shell-side AVX probing. GPU images (cublas/sycl/vulkan/hipblas) ship only
+# llama-cpp-fallback (the accelerator does the compute), so fall back to it when absent.
+if [ -e "$CURDIR"/llama-cpp-cpu-all ]; then
+	BINARY=llama-cpp-cpu-all
 fi
 
 if [ -n "$LLAMACPP_GRPC_SERVERS" ]; then
-	if [ -e $CURDIR/llama-cpp-grpc ]; then
+	if [ -e "$CURDIR"/llama-cpp-grpc ]; then
 		BINARY=llama-cpp-grpc
 	fi
 fi
  
 # Extend ld library path with the dir where this script is located/lib
 if [ "$(uname)" == "Darwin" ]; then
-	export DYLD_LIBRARY_PATH=$CURDIR/lib:$DYLD_LIBRARY_PATH
-	#export DYLD_FALLBACK_LIBRARY_PATH=$CURDIR/lib:$DYLD_FALLBACK_LIBRARY_PATH
+	export DYLD_LIBRARY_PATH="$CURDIR"/lib:$DYLD_LIBRARY_PATH
+	#export DYLD_FALLBACK_LIBRARY_PATH="$CURDIR"/lib:$DYLD_FALLBACK_LIBRARY_PATH
 else
-	export LD_LIBRARY_PATH=$CURDIR/lib:$LD_LIBRARY_PATH
+	export LD_LIBRARY_PATH="$CURDIR"/lib:$LD_LIBRARY_PATH
 	# Tell rocBLAS where to find TensileLibrary data (GPU kernel tuning files)
 	if [ -d "$CURDIR/lib/rocblas/library" ]; then
-		export ROCBLAS_TENSILE_LIBPATH=$CURDIR/lib/rocblas/library
+		export ROCBLAS_TENSILE_LIBPATH="$CURDIR"/lib/rocblas/library
 	fi
 fi
 
 # If there is a lib/ld.so, use it
-if [ -f $CURDIR/lib/ld.so ]; then
+if [ -f "$CURDIR"/lib/ld.so ]; then
 	echo "Using lib/ld.so"
 	echo "Using binary: $BINARY"
-	exec $CURDIR/lib/ld.so $CURDIR/$BINARY "$@"
+	exec "$CURDIR"/lib/ld.so "$CURDIR"/$BINARY "$@"
 fi
 
 echo "Using binary: $BINARY"
-exec $CURDIR/$BINARY "$@"
+exec "$CURDIR"/$BINARY "$@"
 
 # We should never reach this point, however just in case we do, run fallback
-exec $CURDIR/llama-cpp-fallback "$@"
+exec "$CURDIR"/llama-cpp-fallback "$@"

backend/cpp/privacy-filter/CMakeLists.txt

@@ -51,6 +51,14 @@ add_library(hw_grpc_proto STATIC
     ${HW_GRPC_SRCS} ${HW_GRPC_HDRS}
     ${HW_PROTO_SRCS} ${HW_PROTO_HDRS})
 target_include_directories(hw_grpc_proto PUBLIC ${CMAKE_CURRENT_BINARY_DIR})
+# The generated proto/grpc sources include protobuf and grpc++ headers, so this
+# library must see their include dirs. Linking the imported targets propagates
+# them. On Linux the apt headers live in /usr/include (default search path) so
+# this was a no-op; on macOS the Homebrew headers are under /opt/homebrew and
+# would otherwise be missed (runtime_version.h not found).
+target_link_libraries(hw_grpc_proto PUBLIC
+    protobuf::libprotobuf
+    gRPC::grpc++)
 
 # Build only the pf static lib (+ ggml) from the engine tree — no CLI/bench/tests.
 # PF_VULKAN is honored when passed on the cmake command line (it lands in the

backend/cpp/privacy-filter/run.sh

@@ -2,7 +2,13 @@
 # Entry point for the privacy-filter backend image / BACKEND_BINARY mode.
 set -e
 CURDIR=$(dirname "$(realpath "$0")")
-export LD_LIBRARY_PATH="$CURDIR/lib:$LD_LIBRARY_PATH"
+# macOS has no bundled ld.so; the darwin package ships only dylibs under lib/,
+# resolved via DYLD_LIBRARY_PATH (the ld.so branch below is skipped there).
+if [ "$(uname)" = "Darwin" ]; then
+    export DYLD_LIBRARY_PATH="$CURDIR/lib:$DYLD_LIBRARY_PATH"
+else
+    export LD_LIBRARY_PATH="$CURDIR/lib:$LD_LIBRARY_PATH"
+fi
 if [ -f "$CURDIR/lib/ld.so" ]; then
     exec "$CURDIR/lib/ld.so" "$CURDIR/grpc-server" "$@"
 fi

backend/cpp/run-unit-tests.sh

@@ -0,0 +1,71 @@
+#!/bin/bash
+#
+# Discovers and runs every standalone C++ unit test under backend/cpp/.
+#
+# A "standalone" unit test is a *_test.cpp that depends only on the C++ standard
+# library and nlohmann/json (single header) - i.e. it exercises pure helpers and
+# does not need the full llama.cpp + gRPC backend build. Tests that DO need the
+# backend build use the CMake/ctest path (e.g. -DLLAMA_GRPC_BUILD_TESTS=ON)
+# instead and are skipped here.
+#
+# This keeps CI generic: adding a new pure-C++ unit test file named *_test.cpp in
+# an active backend source dir is picked up automatically, with no CI edits.
+#
+# Env:
+#   NLOHMANN_INCLUDE  include dir that contains nlohmann/json.hpp. If unset, the
+#                     nlohmann/json single header is fetched to a temp dir.
+#   CXX               compiler (default: g++).
+#   JSON_VERSION      nlohmann/json tag to fetch when NLOHMANN_INCLUDE is unset
+#                     (default: v3.11.3).
+set -uo pipefail
+
+ROOT="$(cd "$(dirname "$0")" && pwd)"
+CXX="${CXX:-g++}"
+JSON_VERSION="${JSON_VERSION:-v3.11.3}"
+
+JSON_INC="${NLOHMANN_INCLUDE:-}"
+if [ -z "$JSON_INC" ]; then
+    JSON_INC="$(mktemp -d)"
+    mkdir -p "$JSON_INC/nlohmann"
+    echo "Fetching nlohmann/json ${JSON_VERSION} single header..."
+    if ! curl -L -sf \
+        "https://raw.githubusercontent.com/nlohmann/json/${JSON_VERSION}/single_include/nlohmann/json.hpp" \
+        -o "$JSON_INC/nlohmann/json.hpp"; then
+        echo "ERROR: failed to fetch nlohmann/json header" >&2
+        exit 1
+    fi
+fi
+
+# Active source dirs only - exclude per-variant build copies, dev snapshots and
+# the vendored upstream llama.cpp tree.
+mapfile -t tests < <(find "$ROOT" -name '*_test.cpp' \
+    -not -path '*/llama.cpp/*' \
+    -not -path '*-build/*' \
+    -not -path '*-dev/*' \
+    -not -path '*fallback*' | sort)
+
+if [ "${#tests[@]}" -eq 0 ]; then
+    echo "No standalone C++ unit tests found under $ROOT"
+    exit 0
+fi
+
+fail=0
+for test_src in "${tests[@]}"; do
+    name="$(basename "$test_src" .cpp)"
+    bin="$(mktemp -d)/$name"
+    echo "==> $test_src"
+    if ! "$CXX" -std=c++17 -Wall -Wextra \
+        -I"$JSON_INC" -I"$(dirname "$test_src")" \
+        "$test_src" -o "$bin"; then
+        echo "COMPILE FAILED: $test_src" >&2
+        fail=1
+        continue
+    fi
+    if ! "$bin"; then
+        echo "TEST FAILED: $test_src" >&2
+        fail=1
+    fi
+done
+
+echo "Ran ${#tests[@]} standalone C++ unit test file(s)"
+exit "$fail"

backend/cpp/turboquant/Makefile

@@ -65,6 +65,29 @@ turboquant-avx:
 turboquant-fallback:
 	$(call turboquant-build,fallback,-DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off,--target grpc-server)
 
+# Single-build CPU backend via ggml CPU_ALL_VARIANTS (mirrors llama-cpp-cpu-all).
+# turboquant reuses backend/cpp/llama-cpp's CMakeLists.txt (hw_grpc_proto STATIC) and
+# Makefile (SHARED_LIBS make-var + EXTRA_CMAKE_ARGS), so this passes the same overrides
+# through to the copied build: SHARED_LIBS=ON, the DL flags, and --target ggml (which
+# pulls in the per-microarch libggml-cpu-*.so via ggml's add_dependencies). The .so set
+# is collected for package.sh to bundle into package/lib.
+turboquant-cpu-all:
+	rm -rf $(CURRENT_MAKEFILE_DIR)/../turboquant-cpu-all-build
+	cp -rf $(LLAMA_CPP_DIR) $(CURRENT_MAKEFILE_DIR)/../turboquant-cpu-all-build
+	$(MAKE) -C $(CURRENT_MAKEFILE_DIR)/../turboquant-cpu-all-build purge
+	bash $(CURRENT_MAKEFILE_DIR)/patch-grpc-server.sh $(CURRENT_MAKEFILE_DIR)/../turboquant-cpu-all-build/grpc-server.cpp
+	$(info $(GREEN)I turboquant build info:cpu-all-variants$(RESET))
+	LLAMA_REPO=$(LLAMA_REPO) LLAMA_VERSION=$(TURBOQUANT_VERSION) \
+	$(MAKE) -C $(CURRENT_MAKEFILE_DIR)/../turboquant-cpu-all-build llama.cpp
+	bash $(CURRENT_MAKEFILE_DIR)/apply-patches.sh $(CURRENT_MAKEFILE_DIR)/../turboquant-cpu-all-build/llama.cpp $(PATCHES_DIR)
+	SHARED_LIBS=ON EXTRA_CMAKE_ARGS="-DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON" TARGET="--target grpc-server --target ggml" \
+	LLAMA_REPO=$(LLAMA_REPO) LLAMA_VERSION=$(TURBOQUANT_VERSION) \
+	$(MAKE) -C $(CURRENT_MAKEFILE_DIR)/../turboquant-cpu-all-build grpc-server
+	cp -rfv $(CURRENT_MAKEFILE_DIR)/../turboquant-cpu-all-build/grpc-server turboquant-cpu-all
+	rm -rf ggml-shared-libs && mkdir -p ggml-shared-libs
+	find $(CURRENT_MAKEFILE_DIR)/../turboquant-cpu-all-build/llama.cpp/build \( -name '*.so*' -o -name '*.dylib' \) -exec cp -av {} ggml-shared-libs/ \;
+	@echo "Collected ggml shared backends:" && ls -la ggml-shared-libs/
+
 turboquant-grpc:
 	$(call turboquant-build,grpc,-DGGML_RPC=ON -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off,--target grpc-server --target rpc-server)
 

backend/cpp/turboquant/package.sh

@@ -14,6 +14,15 @@ mkdir -p $CURDIR/package/lib
 cp -avrf $CURDIR/turboquant-* $CURDIR/package/
 cp -rfv $CURDIR/run.sh $CURDIR/package/
 
+# Bundle the ggml shared backends from the CPU_ALL_VARIANTS build into package/lib. ggml
+# discovers the per-microarch libggml-cpu-*.so by scanning the executable directory, which
+# (via the bundled lib/ld.so that run.sh launches through) resolves to lib/. See the
+# matching comment in backend/cpp/llama-cpp/package.sh. No-op on the fallback/ROCm builds.
+if [ -d "$CURDIR/ggml-shared-libs" ]; then
+    echo "Bundling ggml shared backends (CPU_ALL_VARIANTS)..."
+    cp -avf $CURDIR/ggml-shared-libs/*.so* $CURDIR/package/lib/
+fi
+
 # Detect architecture and copy appropriate libraries
 if [ -f "/lib64/ld-linux-x86-64.so.2" ]; then
     # x86_64 architecture

backend/cpp/turboquant/run.sh

@@ -2,7 +2,7 @@
 set -ex
 
 # Get the absolute current dir where the script is located
-CURDIR=$(dirname "$(realpath $0)")
+CURDIR=$(dirname "$(realpath "$0")")
 
 cd /
 
@@ -12,54 +12,39 @@ grep -e "flags" /proc/cpuinfo | head -1
 
 BINARY=turboquant-fallback
 
-if grep -q -e "\savx\s" /proc/cpuinfo ; then
-	echo "CPU:    AVX    found OK"
-	if [ -e $CURDIR/turboquant-avx ]; then
-		BINARY=turboquant-avx
-	fi
-fi
-
-if grep -q -e "\savx2\s" /proc/cpuinfo ; then
-	echo "CPU:    AVX2   found OK"
-	if [ -e $CURDIR/turboquant-avx2 ]; then
-		BINARY=turboquant-avx2
-	fi
-fi
-
-# Check avx 512
-if grep -q -e "\savx512f\s" /proc/cpuinfo ; then
-	echo "CPU:    AVX512F found OK"
-	if [ -e $CURDIR/turboquant-avx512 ]; then
-		BINARY=turboquant-avx512
-	fi
+# x86/arm64 ship a single turboquant-cpu-all built with ggml CPU_ALL_VARIANTS: ggml's
+# backend registry dlopens the best libggml-cpu-*.so for this host, so no shell-side
+# probing. ROCm ships only turboquant-fallback, so fall back to it when cpu-all is absent.
+if [ -e "$CURDIR"/turboquant-cpu-all ]; then
+	BINARY=turboquant-cpu-all
 fi
 
 if [ -n "$LLAMACPP_GRPC_SERVERS" ]; then
-	if [ -e $CURDIR/turboquant-grpc ]; then
+	if [ -e "$CURDIR"/turboquant-grpc ]; then
 		BINARY=turboquant-grpc
 	fi
 fi
 
 # Extend ld library path with the dir where this script is located/lib
 if [ "$(uname)" == "Darwin" ]; then
-	export DYLD_LIBRARY_PATH=$CURDIR/lib:$DYLD_LIBRARY_PATH
+	export DYLD_LIBRARY_PATH="$CURDIR"/lib:$DYLD_LIBRARY_PATH
 else
-	export LD_LIBRARY_PATH=$CURDIR/lib:$LD_LIBRARY_PATH
+	export LD_LIBRARY_PATH="$CURDIR"/lib:$LD_LIBRARY_PATH
 	# Tell rocBLAS where to find TensileLibrary data (GPU kernel tuning files)
 	if [ -d "$CURDIR/lib/rocblas/library" ]; then
-		export ROCBLAS_TENSILE_LIBPATH=$CURDIR/lib/rocblas/library
+		export ROCBLAS_TENSILE_LIBPATH="$CURDIR"/lib/rocblas/library
 	fi
 fi
 
 # If there is a lib/ld.so, use it
-if [ -f $CURDIR/lib/ld.so ]; then
+if [ -f "$CURDIR"/lib/ld.so ]; then
 	echo "Using lib/ld.so"
 	echo "Using binary: $BINARY"
-	exec $CURDIR/lib/ld.so $CURDIR/$BINARY "$@"
+	exec "$CURDIR"/lib/ld.so "$CURDIR"/$BINARY "$@"
 fi
 
 echo "Using binary: $BINARY"
-exec $CURDIR/$BINARY "$@"
+exec "$CURDIR"/$BINARY "$@"
 
 # We should never reach this point, however just in case we do, run fallback
-exec $CURDIR/turboquant-fallback "$@"
+exec "$CURDIR"/turboquant-fallback "$@"

backend/go/acestep-cpp/Makefile

@@ -117,7 +117,8 @@ libgoacestepcpp-custom: CMakeLists.txt cpp/goacestepcpp.cpp cpp/goacestepcpp.h
 	cmake .. $(CMAKE_ARGS) && \
 	cmake --build . --config Release -j$(JOBS) --target goacestepcpp && \
 	cd .. && \
-	mv build-$(SO_TARGET)/libgoacestepcpp.so ./$(SO_TARGET)
+	(mv build-$(SO_TARGET)/libgoacestepcpp.so ./$(SO_TARGET) 2>/dev/null || \
+	 mv build-$(SO_TARGET)/libgoacestepcpp.dylib ./$(SO_TARGET) 2>/dev/null)
 
 test: acestep-cpp
 	@echo "Running acestep-cpp tests..."

backend/go/acestep-cpp/main.go

@@ -4,6 +4,7 @@ package main
 import (
 	"flag"
 	"os"
+	"runtime"
 
 	"github.com/ebitengine/purego"
 	grpc "github.com/mudler/LocalAI/pkg/grpc"
@@ -22,7 +23,11 @@ func main() {
 	// Get library name from environment variable, default to fallback
 	libName := os.Getenv("ACESTEP_LIBRARY")
 	if libName == "" {
-		libName = "./libgoacestepcpp-fallback.so"
+		if runtime.GOOS == "darwin" {
+			libName = "./libgoacestepcpp-fallback.dylib"
+		} else {
+			libName = "./libgoacestepcpp-fallback.so"
+		}
 	}
 
 	gosd, err := purego.Dlopen(libName, purego.RTLD_NOW|purego.RTLD_GLOBAL)

backend/go/acestep-cpp/package.sh

@@ -13,6 +13,7 @@ mkdir -p $CURDIR/package/lib
 
 cp -avf $CURDIR/acestep-cpp $CURDIR/package/
 cp -fv $CURDIR/libgoacestepcpp-*.so $CURDIR/package/
+cp -fv $CURDIR/libgoacestepcpp-*.dylib $CURDIR/package/ 2>/dev/null || true
 cp -fv $CURDIR/run.sh $CURDIR/package/
 
 # Detect architecture and copy appropriate libraries

backend/go/acestep-cpp/run.sh

@@ -2,7 +2,7 @@
 set -ex
 
 # Get the absolute current dir where the script is located
-CURDIR=$(dirname "$(realpath $0)")
+CURDIR=$(dirname "$(realpath "$0")")
 
 cd /
 
@@ -12,19 +12,29 @@ if [ "$(uname)" != "Darwin" ]; then
 	grep -e "flags" /proc/cpuinfo | head -1
 fi
 
-LIBRARY="$CURDIR/libgoacestepcpp-fallback.so"
+if [ "$(uname)" = "Darwin" ]; then
+	# macOS: single library variant (Metal or Accelerate). The goacestepcpp
+	# target is built as a CMake MODULE, which emits a .dylib for a SHARED
+	# build but a .so for a MODULE build on Apple, so prefer .dylib and fall
+	# back to .so.
+	LIBRARY="$CURDIR/libgoacestepcpp-fallback.dylib"
+	if [ ! -e "$LIBRARY" ]; then
+		LIBRARY="$CURDIR/libgoacestepcpp-fallback.so"
+	fi
+	export DYLD_LIBRARY_PATH="$CURDIR"/lib:$DYLD_LIBRARY_PATH
+else
+	LIBRARY="$CURDIR/libgoacestepcpp-fallback.so"
 
-if [ "$(uname)" != "Darwin" ]; then
 	if grep -q -e "\savx\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX    found OK"
-		if [ -e $CURDIR/libgoacestepcpp-avx.so ]; then
+		if [ -e "$CURDIR"/libgoacestepcpp-avx.so ]; then
 			LIBRARY="$CURDIR/libgoacestepcpp-avx.so"
 		fi
 	fi
 
 	if grep -q -e "\savx2\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX2   found OK"
-		if [ -e $CURDIR/libgoacestepcpp-avx2.so ]; then
+		if [ -e "$CURDIR"/libgoacestepcpp-avx2.so ]; then
 			LIBRARY="$CURDIR/libgoacestepcpp-avx2.so"
 		fi
 	fi
@@ -32,21 +42,22 @@ if [ "$(uname)" != "Darwin" ]; then
 	# Check avx 512
 	if grep -q -e "\savx512f\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX512F found OK"
-		if [ -e $CURDIR/libgoacestepcpp-avx512.so ]; then
+		if [ -e "$CURDIR"/libgoacestepcpp-avx512.so ]; then
 			LIBRARY="$CURDIR/libgoacestepcpp-avx512.so"
 		fi
 	fi
+
+	export LD_LIBRARY_PATH="$CURDIR"/lib:$LD_LIBRARY_PATH
 fi
 
-export LD_LIBRARY_PATH=$CURDIR/lib:$LD_LIBRARY_PATH
 export ACESTEP_LIBRARY=$LIBRARY
 
 # If there is a lib/ld.so, use it
-if [ -f $CURDIR/lib/ld.so ]; then
+if [ -f "$CURDIR"/lib/ld.so ]; then
 	echo "Using lib/ld.so"
 	echo "Using library: $LIBRARY"
-	exec $CURDIR/lib/ld.so $CURDIR/acestep-cpp "$@"
+	exec "$CURDIR"/lib/ld.so "$CURDIR"/acestep-cpp "$@"
 fi
 
 echo "Using library: $LIBRARY"
-exec $CURDIR/acestep-cpp "$@"
+exec "$CURDIR"/acestep-cpp "$@"

backend/go/ced/Makefile

@@ -57,6 +57,7 @@ libced.so: sources/ced.cpp
 	cmake -B sources/ced.cpp/build-shared -S sources/ced.cpp $(CMAKE_ARGS)
 	cmake --build sources/ced.cpp/build-shared --config Release -j$(JOBS)
 	cp -fv sources/ced.cpp/build-shared/libced.so* ./ 2>/dev/null || true
+	cp -fv sources/ced.cpp/build-shared/libced.dylib ./ 2>/dev/null || true
 	cp -fv sources/ced.cpp/include/ced_capi.h ./
 
 ced-grpc: libced.so main.go goced.go

backend/go/ced/main.go

@@ -12,6 +12,7 @@ import (
 	"flag"
 	"fmt"
 	"os"
+	"runtime"
 
 	"github.com/ebitengine/purego"
 	grpc "github.com/mudler/LocalAI/pkg/grpc"
@@ -27,7 +28,11 @@ type libFunc struct {
 func main() {
 	libName := os.Getenv("CED_LIBRARY")
 	if libName == "" {
-		libName = "libced.so"
+		if runtime.GOOS == "darwin" {
+			libName = "libced.dylib"
+		} else {
+			libName = "libced.so"
+		}
 	}
 	lib, err := purego.Dlopen(libName, purego.RTLD_NOW|purego.RTLD_GLOBAL)
 	if err != nil {

backend/go/ced/package.sh

@@ -15,10 +15,12 @@ mkdir -p "$CURDIR/package/lib"
 cp -avf "$CURDIR/ced-grpc" "$CURDIR/package/"
 cp -avf "$CURDIR/run.sh" "$CURDIR/package/"
 
-cp -avf "$CURDIR"/libced.so* "$CURDIR/package/lib/" 2>/dev/null || {
-	echo "ERROR: libced.so not found in $CURDIR, run 'make' first" >&2
+cp -avf "$CURDIR"/libced.so* "$CURDIR/package/lib/" 2>/dev/null || true
+cp -avf "$CURDIR"/libced.dylib "$CURDIR/package/lib/" 2>/dev/null || true
+if ! ls "$CURDIR"/package/lib/libced.* >/dev/null 2>&1; then
+	echo "ERROR: libced shared library not found in $CURDIR, run 'make' first" >&2
 	exit 1
-}
+fi
 
 if [ -f "/lib64/ld-linux-x86-64.so.2" ]; then
     echo "Detected x86_64 architecture, copying x86_64 libraries..."

backend/go/ced/run.sh

@@ -3,7 +3,12 @@ set -e
 
 CURDIR=$(dirname "$(realpath "$0")")
 
-export LD_LIBRARY_PATH="$CURDIR/lib:$CURDIR:${LD_LIBRARY_PATH:-}"
+if [ "$(uname)" = "Darwin" ]; then
+	export DYLD_LIBRARY_PATH="$CURDIR/lib:"$CURDIR":${DYLD_LIBRARY_PATH:-}"
+	export CED_LIBRARY="$CURDIR/lib/libced.dylib"
+else
+	export LD_LIBRARY_PATH="$CURDIR/lib:"$CURDIR":${LD_LIBRARY_PATH:-}"
+fi
 
 # If a self-contained ld.so was packaged, route through it so the packaged
 # libc / libstdc++ are used instead of the host's (matches the sibling backends).

backend/go/cloud-proxy/run.sh

@@ -1,6 +1,6 @@
 #!/bin/bash
 set -ex
 
-CURDIR=$(dirname "$(realpath $0)")
+CURDIR=$(dirname "$(realpath "$0")")
 
-exec $CURDIR/cloud-proxy "$@"
+exec "$CURDIR"/cloud-proxy "$@"

backend/go/crispasr/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # CrispASR version (release tag)
 CRISPASR_REPO?=https://github.com/CrispStrobe/CrispASR
-CRISPASR_VERSION?=63b57289255267edf66e43e33bc3911e04a2e92d
+CRISPASR_VERSION?=6514c9da00b03a2f0f1b49a43fae4f3a01a41844
 SO_TARGET?=libgocrispasr.so
 
 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF
@@ -75,7 +75,8 @@ UNAME_S := $(shell uname -s)
 ifeq ($(UNAME_S),Linux)
 	VARIANT_TARGETS = libgocrispasr-avx.so libgocrispasr-avx2.so libgocrispasr-avx512.so libgocrispasr-fallback.so
 else
-	VARIANT_TARGETS = libgocrispasr-fallback.so
+	# On non-Linux (e.g., Darwin), build only fallback variant (as a dylib)
+	VARIANT_TARGETS = libgocrispasr-fallback.dylib
 endif
 
 crispasr: main.go gocrispasr.go $(VARIANT_TARGETS)
@@ -87,7 +88,7 @@ package: crispasr
 build: package
 
 clean: purge
-	rm -rf libgocrispasr*.so package sources/CrispASR crispasr
+	rm -rf libgocrispasr*.so libgocrispasr*.dylib package sources/CrispASR crispasr
 
 purge:
 	rm -rf build*
@@ -118,13 +119,21 @@ libgocrispasr-fallback.so: sources/CrispASR
 	SO_TARGET=libgocrispasr-fallback.so CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off" $(MAKE) libgocrispasr-custom
 	rm -rfv build*
 
+# Build fallback variant as a dylib (Darwin)
+libgocrispasr-fallback.dylib: sources/CrispASR
+	$(MAKE) purge
+	$(info ${GREEN}I crispasr build info:fallback (dylib)${RESET})
+	SO_TARGET=libgocrispasr-fallback.dylib CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off" $(MAKE) libgocrispasr-custom
+	rm -rfv build*
+
 libgocrispasr-custom: CMakeLists.txt cpp/crispasr_shim.cpp cpp/crispasr_shim.h
 	mkdir -p build-$(SO_TARGET) && \
 	cd build-$(SO_TARGET) && \
 	cmake .. $(CMAKE_ARGS) && \
 	cmake --build . --config Release -j$(JOBS) && \
 	cd .. && \
-	mv build-$(SO_TARGET)/libgocrispasr.so ./$(SO_TARGET)
+	(mv build-$(SO_TARGET)/libgocrispasr.so ./$(SO_TARGET) 2>/dev/null || \
+	 mv build-$(SO_TARGET)/libgocrispasr.dylib ./$(SO_TARGET) 2>/dev/null)
 
 test: crispasr
 	CGO_ENABLED=0 $(GOCMD) test -v ./...

backend/go/crispasr/main.go

@@ -4,6 +4,7 @@ package main
 import (
 	"flag"
 	"os"
+	"runtime"
 
 	"github.com/ebitengine/purego"
 	grpc "github.com/mudler/LocalAI/pkg/grpc"
@@ -21,7 +22,11 @@ type LibFuncs struct {
 func main() {
 	libName := os.Getenv("CRISPASR_LIBRARY")
 	if libName == "" {
-		libName = "./libgocrispasr-fallback.so"
+		if runtime.GOOS == "darwin" {
+			libName = "./libgocrispasr-fallback.dylib"
+		} else {
+			libName = "./libgocrispasr-fallback.so"
+		}
 	}
 
 	lib, err := purego.Dlopen(libName, purego.RTLD_NOW|purego.RTLD_GLOBAL)

backend/go/crispasr/package.sh

@@ -12,7 +12,8 @@ REPO_ROOT="${CURDIR}/../../.."
 mkdir -p $CURDIR/package/lib
 
 cp -avf $CURDIR/crispasr $CURDIR/package/
-cp -fv $CURDIR/libgocrispasr-*.so $CURDIR/package/
+cp -fv $CURDIR/libgocrispasr-*.so $CURDIR/package/ 2>/dev/null || true
+cp -fv $CURDIR/libgocrispasr-*.dylib $CURDIR/package/ 2>/dev/null || true
 cp -fv $CURDIR/run.sh $CURDIR/package/
 
 # Detect architecture and copy appropriate libraries

backend/go/crispasr/run.sh

@@ -2,7 +2,7 @@
 set -ex
 
 # Get the absolute current dir where the script is located
-CURDIR=$(dirname "$(realpath $0)")
+CURDIR=$(dirname "$(realpath "$0")")
 
 cd /
 
@@ -12,19 +12,23 @@ if [ "$(uname)" != "Darwin" ]; then
 	grep -e "flags" /proc/cpuinfo | head -1
 fi
 
-LIBRARY="$CURDIR/libgocrispasr-fallback.so"
+if [ "$(uname)" = "Darwin" ]; then
+	# macOS: single dylib variant (Metal or Accelerate)
+	LIBRARY="$CURDIR/libgocrispasr-fallback.dylib"
+	export DYLD_LIBRARY_PATH="$CURDIR"/lib:$DYLD_LIBRARY_PATH
+else
+	LIBRARY="$CURDIR/libgocrispasr-fallback.so"
 
-if [ "$(uname)" != "Darwin" ]; then
 	if grep -q -e "\savx\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX    found OK"
-		if [ -e $CURDIR/libgocrispasr-avx.so ]; then
+		if [ -e "$CURDIR"/libgocrispasr-avx.so ]; then
 			LIBRARY="$CURDIR/libgocrispasr-avx.so"
 		fi
 	fi
 
 	if grep -q -e "\savx2\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX2   found OK"
-		if [ -e $CURDIR/libgocrispasr-avx2.so ]; then
+		if [ -e "$CURDIR"/libgocrispasr-avx2.so ]; then
 			LIBRARY="$CURDIR/libgocrispasr-avx2.so"
 		fi
 	fi
@@ -32,26 +36,27 @@ if [ "$(uname)" != "Darwin" ]; then
 	# Check avx 512
 	if grep -q -e "\savx512f\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX512F found OK"
-		if [ -e $CURDIR/libgocrispasr-avx512.so ]; then
+		if [ -e "$CURDIR"/libgocrispasr-avx512.so ]; then
 			LIBRARY="$CURDIR/libgocrispasr-avx512.so"
 		fi
 	fi
+
+	export LD_LIBRARY_PATH="$CURDIR"/lib:$LD_LIBRARY_PATH
 fi
 
-export LD_LIBRARY_PATH=$CURDIR/lib:$LD_LIBRARY_PATH
 export CRISPASR_LIBRARY=$LIBRARY
 
 # Point piper's espeak-ng phonemizer at the bundled voice data. The variable
 # names the directory CONTAINING espeak-ng-data (package.sh drops it next to
 # this script). Harmless when espeak-ng wasn't bundled.
-export CRISPASR_ESPEAK_DATA_PATH=$CURDIR
+export CRISPASR_ESPEAK_DATA_PATH="$CURDIR"
 
 # If there is a lib/ld.so, use it
-if [ -f $CURDIR/lib/ld.so ]; then
+if [ -f "$CURDIR"/lib/ld.so ]; then
 	echo "Using lib/ld.so"
 	echo "Using library: $LIBRARY"
-	exec $CURDIR/lib/ld.so $CURDIR/crispasr "$@"
+	exec "$CURDIR"/lib/ld.so "$CURDIR"/crispasr "$@"
 fi
 
 echo "Using library: $LIBRARY"
-exec $CURDIR/crispasr "$@"
+exec "$CURDIR"/crispasr "$@"

backend/go/depth-anything-cpp/Makefile

@@ -40,6 +40,8 @@ else ifeq ($(BUILD_TYPE),hipblas)
 else ifeq ($(BUILD_TYPE),vulkan)
 	CMAKE_ARGS+=-DGGML_VULKAN=ON -DDA_GGML_VULKAN=ON
 else ifeq ($(OS),Darwin)
+	# macOS/Metal: built + published as an OCI image by CI (includeDarwin in
+	# .github/backend-matrix.yml) so Apple Silicon users can install this backend.
 	ifneq ($(BUILD_TYPE),metal)
 		CMAKE_ARGS+=-DGGML_METAL=OFF
 	else
@@ -77,7 +79,7 @@ ifeq ($(UNAME_S),Linux)
 	VARIANT_TARGETS = libdepthanythingcpp-avx.so libdepthanythingcpp-avx2.so libdepthanythingcpp-avx512.so libdepthanythingcpp-fallback.so
 else
 	# On non-Linux (e.g., Darwin), build only fallback variant
-	VARIANT_TARGETS = libdepthanythingcpp-fallback.so
+	VARIANT_TARGETS = libdepthanythingcpp-fallback.dylib
 endif
 
 depth-anything-cpp: main.go godepthanythingcpp.go $(VARIANT_TARGETS)
@@ -89,7 +91,7 @@ package: depth-anything-cpp
 build: package
 
 clean: purge
-	rm -rf libdepthanythingcpp*.so depth-anything-cpp package sources
+	rm -rf libdepthanythingcpp*.so libdepthanythingcpp*.dylib depth-anything-cpp package sources
 
 purge:
 	rm -rf build*
@@ -116,11 +118,19 @@ libdepthanythingcpp-avx512.so: sources/depth-anything.cpp
 endif
 
 # Build fallback variant (all platforms)
+ifeq ($(UNAME_S),Darwin)
+libdepthanythingcpp-fallback.dylib: sources/depth-anything.cpp
+	rm -rfv build-$@
+	$(info ${GREEN}I depth-anything-cpp build info:fallback${RESET})
+	SO_TARGET=$@ CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off" $(MAKE) libdepthanythingcpp-custom
+	rm -rfv build-$@
+else
 libdepthanythingcpp-fallback.so: sources/depth-anything.cpp
 	rm -rfv build-$@
 	$(info ${GREEN}I depth-anything-cpp build info:fallback${RESET})
 	SO_TARGET=$@ CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off" $(MAKE) libdepthanythingcpp-custom
 	rm -rfv build-$@
+endif
 
 libdepthanythingcpp-custom: CMakeLists.txt
 	mkdir -p build-$(SO_TARGET) && \
@@ -128,7 +138,8 @@ libdepthanythingcpp-custom: CMakeLists.txt
 	cmake .. $(CMAKE_ARGS) && \
 	cmake --build . --config Release -j$(JOBS) && \
 	cd .. && \
-	mv build-$(SO_TARGET)/libdepthanything.so ./$(SO_TARGET)
+	(mv build-$(SO_TARGET)/libdepthanything.so ./$(SO_TARGET) 2>/dev/null || \
+	 mv build-$(SO_TARGET)/libdepthanything.dylib ./$(SO_TARGET) 2>/dev/null)
 
 all: depth-anything-cpp package
 

backend/go/depth-anything-cpp/main.go

@@ -9,6 +9,7 @@ package main
 import (
 	"flag"
 	"os"
+	"runtime"
 
 	"github.com/ebitengine/purego"
 	grpc "github.com/mudler/LocalAI/pkg/grpc"
@@ -27,7 +28,11 @@ func main() {
 	// Get library name from environment variable, default to fallback
 	libName := os.Getenv("DEPTHANYTHING_LIBRARY")
 	if libName == "" {
-		libName = "./libdepthanythingcpp-fallback.so"
+		if runtime.GOOS == "darwin" {
+			libName = "./libdepthanythingcpp-fallback.dylib"
+		} else {
+			libName = "./libdepthanythingcpp-fallback.so"
+		}
 	}
 
 	lib, err := purego.Dlopen(libName, purego.RTLD_NOW|purego.RTLD_GLOBAL)

backend/go/depth-anything-cpp/package.sh

@@ -10,7 +10,8 @@ REPO_ROOT="${CURDIR}/../../.."
 # Create lib directory
 mkdir -p $CURDIR/package/lib
 
-cp -avf $CURDIR/libdepthanythingcpp-*.so $CURDIR/package/
+cp -fv $CURDIR/libdepthanythingcpp-*.so $CURDIR/package/ 2>/dev/null || true
+cp -fv $CURDIR/libdepthanythingcpp-*.dylib $CURDIR/package/ 2>/dev/null || true
 cp -avf $CURDIR/depth-anything-cpp $CURDIR/package/
 cp -fv $CURDIR/run.sh $CURDIR/package/
 

backend/go/depth-anything-cpp/run.sh

@@ -2,7 +2,7 @@
 set -ex
 
 # Get the absolute current dir where the script is located
-CURDIR=$(dirname "$(realpath $0)")
+CURDIR=$(dirname "$(realpath "$0")")
 
 cd /
 
@@ -12,19 +12,23 @@ if [ "$(uname)" != "Darwin" ]; then
 	grep -e "flags" /proc/cpuinfo | head -1
 fi
 
-LIBRARY="$CURDIR/libdepthanythingcpp-fallback.so"
+if [ "$(uname)" = "Darwin" ]; then
+	# macOS: single dylib variant (Metal or Accelerate)
+	LIBRARY="$CURDIR/libdepthanythingcpp-fallback.dylib"
+	export DYLD_LIBRARY_PATH="$CURDIR"/lib:$DYLD_LIBRARY_PATH
+else
+	LIBRARY="$CURDIR/libdepthanythingcpp-fallback.so"
 
-if [ "$(uname)" != "Darwin" ]; then
 	if grep -q -e "\savx\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX    found OK"
-		if [ -e $CURDIR/libdepthanythingcpp-avx.so ]; then
+		if [ -e "$CURDIR"/libdepthanythingcpp-avx.so ]; then
 			LIBRARY="$CURDIR/libdepthanythingcpp-avx.so"
 		fi
 	fi
 
 	if grep -q -e "\savx2\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX2   found OK"
-		if [ -e $CURDIR/libdepthanythingcpp-avx2.so ]; then
+		if [ -e "$CURDIR"/libdepthanythingcpp-avx2.so ]; then
 			LIBRARY="$CURDIR/libdepthanythingcpp-avx2.so"
 		fi
 	fi
@@ -32,21 +36,22 @@ if [ "$(uname)" != "Darwin" ]; then
 	# Check avx 512
 	if grep -q -e "\savx512f\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX512F found OK"
-		if [ -e $CURDIR/libdepthanythingcpp-avx512.so ]; then
+		if [ -e "$CURDIR"/libdepthanythingcpp-avx512.so ]; then
 			LIBRARY="$CURDIR/libdepthanythingcpp-avx512.so"
 		fi
 	fi
+
+	export LD_LIBRARY_PATH="$CURDIR"/lib:$LD_LIBRARY_PATH
 fi
 
-export LD_LIBRARY_PATH=$CURDIR/lib:$LD_LIBRARY_PATH
 export DEPTHANYTHING_LIBRARY=$LIBRARY
 
 # If there is a lib/ld.so, use it
-if [ -f $CURDIR/lib/ld.so ]; then
+if [ -f "$CURDIR"/lib/ld.so ]; then
 	echo "Using lib/ld.so"
 	echo "Using library: $LIBRARY"
-	exec $CURDIR/lib/ld.so $CURDIR/depth-anything-cpp "$@"
+	exec "$CURDIR"/lib/ld.so "$CURDIR"/depth-anything-cpp "$@"
 fi
 
 echo "Using library: $LIBRARY"
-exec $CURDIR/depth-anything-cpp "$@"
+exec "$CURDIR"/depth-anything-cpp "$@"

backend/go/face-detect/.gitignore

@@ -0,0 +1,18 @@
+# Fetched upstream sources
+sources/
+
+# CMake build directories
+build*/
+
+# build artifacts staged in-tree by the Makefile (cp from sources/) or
+# symlinked for local dev; the real sources live in face-detect.cpp upstream.
+*.so
+*.so.*
+facedetect_capi.h
+compile_commands.json
+
+# Compiled backend binary
+face-detect-grpc
+
+# Packaging output
+package/

backend/go/face-detect/Makefile

@@ -0,0 +1,110 @@
+# face-detect backend Makefile.
+#
+# Upstream pin lives below as FACEDETECT_VERSION?=06914b0... (.github/bump_deps.sh
+# can find and update it - matches the voice-detect / parakeet.cpp / whisper.cpp
+# convention).
+#
+# Local dev shortcut: if you already have an out-of-tree face-detect.cpp build,
+# symlink the .so + header into this directory and skip the clone/cmake steps:
+#
+#   ln -sf /path/to/face-detect.cpp/build-shared/libfacedetect.so .
+#   ln -sf /path/to/face-detect.cpp/include/facedetect_capi.h .
+#   go build -o face-detect-grpc .
+#
+# The default target below does the proper clone-at-pin + cmake build so CI does
+# not need a side-checkout.
+
+FACEDETECT_VERSION?=06914b077d52f90d5421299138e7be6bdd06b5e8
+FACEDETECT_REPO?=https://github.com/mudler/face-detect.cpp
+
+GOCMD?=go
+GO_TAGS?=
+JOBS?=$(shell nproc 2>/dev/null || sysctl -n hw.ncpu 2>/dev/null || echo 4)
+
+BUILD_TYPE?=
+NATIVE?=false
+
+# Resolve the target arch. The backend matrix / Docker build pass TARGETARCH
+# (amd64|arm64); fall back to uname -m (aarch64|x86_64) for a local build.
+RECON_ARCH?=$(or $(TARGETARCH),$(shell uname -m))
+
+# Build ggml + the vendored libjpeg-turbo statically into libfacedetect.so (PIC)
+# so the shared lib is self-contained: dlopen needs no libggml*.so alongside it,
+# only system libs (libstdc++/libgomp/libc) the runtime image already provides.
+# The vendored jpeg symbols are hidden via -Wl,--exclude-libs,ALL on the C++
+# side, so only the facedetect_capi_* surface is exported.
+CMAKE_ARGS?=-DCMAKE_BUILD_TYPE=Release -DFACEDETECT_SHARED=ON -DFACEDETECT_BUILD_CLI=OFF -DFACEDETECT_BUILD_TESTS=OFF -DBUILD_SHARED_LIBS=OFF -DCMAKE_POSITION_INDEPENDENT_CODE=ON
+
+ifeq ($(NATIVE),false)
+	CMAKE_ARGS+=-DGGML_NATIVE=OFF
+endif
+
+# face-detect.cpp gates its GGML backends behind FACEDETECT_GGML_* options and
+# does set(GGML_CUDA ${FACEDETECT_GGML_CUDA} CACHE BOOL "" FORCE), so a bare
+# -DGGML_CUDA=ON is overwritten back to OFF. Forward the FACEDETECT_GGML_*
+# options instead. (openblas is not gated, so -DGGML_BLAS passes through.)
+ifeq ($(BUILD_TYPE),cublas)
+	CMAKE_ARGS+=-DFACEDETECT_GGML_CUDA=ON
+	# Opt-in cuDNN implicit-GEMM conv path (kills im2col on GPU, SCRFD 2.3x
+	# vs torch-cuDNN parity). Only the arm64 + CUDA 13 image (GB10/Jetson/L4T)
+	# ships libcudnn9 + the -dev headers, so gate cuDNN to that variant.
+	# x86 CUDA images carry no cuDNN -> enabling it there is a link failure.
+	ifeq ($(CUDA_MAJOR_VERSION),13)
+	ifneq (,$(filter arm64 aarch64,$(RECON_ARCH)))
+		CMAKE_ARGS+=-DFACEDETECT_GGML_CUDNN=ON
+	endif
+	endif
+else ifeq ($(BUILD_TYPE),openblas)
+	CMAKE_ARGS+=-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
+else ifeq ($(BUILD_TYPE),hipblas)
+	CMAKE_ARGS+=-DFACEDETECT_GGML_HIP=ON
+else ifeq ($(BUILD_TYPE),vulkan)
+	CMAKE_ARGS+=-DFACEDETECT_GGML_VULKAN=ON
+else ifeq ($(BUILD_TYPE),metal)
+	CMAKE_ARGS+=-DFACEDETECT_GGML_METAL=ON
+endif
+
+.PHONY: face-detect-grpc package build clean purge test all
+
+all: face-detect-grpc
+
+# Clone the upstream face-detect.cpp source at the pinned commit. Directory acts
+# as the target so make only re-clones when missing. After a FACEDETECT_VERSION
+# bump, run 'make purge && make' to refetch.
+sources/face-detect.cpp:
+	mkdir -p sources/face-detect.cpp
+	cd sources/face-detect.cpp && \
+	git init -q && \
+	git remote add origin $(FACEDETECT_REPO) && \
+	git fetch --depth 1 origin $(FACEDETECT_VERSION) && \
+	git checkout FETCH_HEAD && \
+	git submodule update --init --recursive --depth 1 --single-branch
+
+# Build the shared lib + header out-of-tree, then stage them next to the Go
+# sources so purego.Dlopen("libfacedetect.so") and the cgo-less build both pick
+# them up.
+libfacedetect.so: sources/face-detect.cpp
+	cmake -B sources/face-detect.cpp/build-shared -S sources/face-detect.cpp $(CMAKE_ARGS)
+	cmake --build sources/face-detect.cpp/build-shared --config Release -j$(JOBS) --target facedetect
+	cp -fv sources/face-detect.cpp/build-shared/libfacedetect.so* ./ 2>/dev/null || true
+	cp -fv sources/face-detect.cpp/include/facedetect_capi.h ./
+
+face-detect-grpc: libfacedetect.so main.go gofacedetect.go options.go
+	CGO_ENABLED=0 $(GOCMD) build -tags "$(GO_TAGS)" -o face-detect-grpc .
+
+package: face-detect-grpc
+	bash package.sh
+
+build: package
+
+# Test target. The embed/detect/verify/analyze smoke specs are gated on
+# FACEDETECT_BACKEND_TEST_MODEL + FACEDETECT_BACKEND_TEST_IMAGE; without them the
+# heavy specs auto-skip and only the pure-Go parsing specs run.
+test:
+	LD_LIBRARY_PATH=$(CURDIR):$$LD_LIBRARY_PATH $(GOCMD) test ./... -count=1
+
+clean: purge
+	rm -rf libfacedetect.so* facedetect_capi.h package face-detect-grpc
+
+purge:
+	rm -rf sources/face-detect.cpp

backend/go/face-detect/gofacedetect.go

@@ -0,0 +1,431 @@
+package main
+
+import (
+	"encoding/base64"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"math"
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"time"
+	"unsafe"
+
+	"github.com/mudler/LocalAI/pkg/grpc/base"
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+	"github.com/mudler/xlog"
+)
+
+// purego-bound entry points from libfacedetect.so. Names match
+// facedetect_capi.h exactly so a `nm libfacedetect.so | grep facedetect_capi`
+// is enough to spot drift.
+//
+// The opaque ctx and the malloc'd char*/float* return values are declared as
+// uintptr so we get the raw pointer back and can release it via the matching
+// capi free function. purego's native string/[]float32 returns would copy and
+// forget the original pointer, leaking the C-owned buffer on every call.
+var (
+	CppAbiVersion  func() int32
+	CppLoad        func(ggufPath string) uintptr
+	CppFree        func(ctx uintptr)
+	CppLastError   func(ctx uintptr) string
+	CppFreeString  func(s uintptr)
+	CppFreeVec     func(v uintptr)
+	CppEmbedPath   func(ctx uintptr, imagePath string, outVec, outDim unsafe.Pointer) int32
+	CppEmbedRGB    func(ctx uintptr, rgb []byte, width, height int32, outVec, outDim unsafe.Pointer) int32
+	CppDetectJSON  func(ctx uintptr, imagePath string) uintptr
+	CppVerifyPaths func(ctx uintptr, a, b string, threshold float32, antiSpoof int32, outDistance, outVerified unsafe.Pointer) int32
+	CppAnalyzeJSON func(ctx uintptr, imagePath string) uintptr
+)
+
+// FaceDetect implements the face-recognition (biometric) subset of the Backend
+// gRPC service over libfacedetect.so. The C side keeps a single loaded model
+// pack plus a per-ctx last-error buffer and is not reentrant, so
+// base.SingleThread serializes every call.
+type FaceDetect struct {
+	base.SingleThread
+	opts   loadOptions
+	ctxPtr uintptr
+}
+
+func (f *FaceDetect) Load(opts *pb.ModelOptions) error {
+	model := opts.ModelFile
+	if model == "" {
+		model = opts.ModelPath
+	}
+	if !filepath.IsAbs(model) && opts.ModelPath != "" {
+		model = filepath.Join(opts.ModelPath, model)
+	}
+	if model == "" {
+		return errors.New("face-detect: ModelFile is required")
+	}
+
+	f.opts = parseOptions(opts.Options)
+	if f.opts.modelName == "" {
+		f.opts.modelName = filepath.Base(model)
+	}
+
+	// Propagate LocalAI's per-model thread budget to the engine. LocalAI spawns
+	// one backend process per model and serves requests concurrently, so the
+	// engine's own min(hardware_concurrency, 8) default can oversubscribe cores.
+	// FACEDETECT_THREADS is read by the engine at backend construction, so it
+	// must be set before the capi load. A non-positive Threads means "unset":
+	// leave the env alone so the engine keeps its sane default.
+	threads := opts.Threads
+	if threads > 0 {
+		if err := os.Setenv("FACEDETECT_THREADS", strconv.Itoa(int(threads))); err != nil {
+			return fmt.Errorf("face-detect: set FACEDETECT_THREADS: %w", err)
+		}
+		xlog.Info("face-detect: applying LocalAI thread budget", "threads", threads)
+	}
+
+	xlog.Info("face-detect: loading model", "model", model,
+		"verify_threshold", f.opts.verifyThreshold, "abi", CppAbiVersion())
+
+	ctx := CppLoad(model)
+	if ctx == 0 {
+		// The last-error buffer lives on the ctx that was never returned, so
+		// surface the path the operator tried to load instead.
+		return fmt.Errorf("face-detect: facedetect_capi_load failed for %q", model)
+	}
+	f.ctxPtr = ctx
+	return nil
+}
+
+// Embeddings returns the L2-normalized ArcFace embedding of the primary face in
+// the supplied image. Mirroring the Python face backend, the image is read from
+// Images[0] as a base64 payload; materializeImage decodes it to a temp file so
+// the path-based C-API can run its own decode (cv2.imread parity). The gRPC
+// server wraps the returned slice in an EmbeddingResult.
+func (f *FaceDetect) Embeddings(req *pb.PredictOptions) ([]float32, error) {
+	if f.ctxPtr == 0 {
+		return nil, errors.New("face-detect: model not loaded")
+	}
+	if len(req.Images) == 0 || req.Images[0] == "" {
+		return nil, errors.New("face-detect: Embedding requires Images[0] to be a base64 image")
+	}
+
+	path, cleanup, err := materializeImage(req.Images[0])
+	if err != nil {
+		return nil, err
+	}
+	defer cleanup()
+
+	return f.embedPath(path)
+}
+
+func (f *FaceDetect) embedPath(path string) ([]float32, error) {
+	var vec uintptr
+	var dim int32
+	rc := CppEmbedPath(f.ctxPtr, path, unsafe.Pointer(&vec), unsafe.Pointer(&dim))
+	if rc != 0 || vec == 0 || dim <= 0 {
+		return nil, f.lastErr("embed", path)
+	}
+	defer CppFreeVec(vec)
+	// Copy out of the C-owned malloc'd buffer before freeing it. The
+	// uintptr->Pointer conversion trips vet's unsafeptr check, which can't tell
+	// a C heap pointer from Go-managed memory; safe here, the GC neither tracks
+	// nor moves this buffer and we copy immediately.
+	src := unsafe.Slice((*float32)(unsafe.Pointer(vec)), int(dim)) //nolint:govet // C-owned malloc'd vector, copied out before free
+	out := make([]float32, int(dim))
+	copy(out, src)
+	return out, nil
+}
+
+// Detect runs SCRFD over the image and returns one Detection per face. The
+// C-API emits a box as [x1,y1,x2,y2] in pixels; the proto carries x/y plus
+// width/height, so the corners are converted. The 5 facial landmarks the engine
+// also returns are dropped: the Detection message has no field for them.
+func (f *FaceDetect) Detect(req *pb.DetectOptions) (pb.DetectResponse, error) {
+	if f.ctxPtr == 0 {
+		return pb.DetectResponse{}, errors.New("face-detect: model not loaded")
+	}
+	if req.Src == "" {
+		return pb.DetectResponse{}, errors.New("face-detect: src image is required")
+	}
+
+	path, cleanup, err := materializeImage(req.Src)
+	if err != nil {
+		return pb.DetectResponse{}, err
+	}
+	defer cleanup()
+
+	faces, err := f.detectFaces(path)
+	if err != nil {
+		return pb.DetectResponse{}, err
+	}
+
+	dets := make([]*pb.Detection, 0, len(faces))
+	for _, fc := range faces {
+		if req.Threshold > 0 && fc.Score < req.Threshold {
+			continue
+		}
+		x, y, w, h := fc.xywh()
+		dets = append(dets, &pb.Detection{
+			X:          x,
+			Y:          y,
+			Width:      w,
+			Height:     h,
+			Confidence: fc.Score,
+			ClassName:  "face",
+		})
+	}
+	return pb.DetectResponse{Detections: dets}, nil
+}
+
+// FaceVerify embeds the primary face in each image and reports whether they are
+// the same identity by cosine distance against a threshold. A request threshold
+// <= 0 falls back to the model-configured default (verify_threshold option,
+// 0.35 if unset). When anti_spoofing is set, the C-API applies a MiniFASNet
+// veto internally (verified forced false on a spoof); the per-image liveness
+// scores are not exposed by the verify entry point, so img*_is_real /
+// img*_antispoof_score stay at their zero values.
+func (f *FaceDetect) FaceVerify(req *pb.FaceVerifyRequest) (pb.FaceVerifyResponse, error) {
+	if f.ctxPtr == 0 {
+		return pb.FaceVerifyResponse{}, errors.New("face-detect: model not loaded")
+	}
+	if req.Img1 == "" || req.Img2 == "" {
+		return pb.FaceVerifyResponse{}, errors.New("face-detect: img1 and img2 are required")
+	}
+
+	path1, cleanup1, err := materializeImage(req.Img1)
+	if err != nil {
+		return pb.FaceVerifyResponse{}, err
+	}
+	defer cleanup1()
+	path2, cleanup2, err := materializeImage(req.Img2)
+	if err != nil {
+		return pb.FaceVerifyResponse{}, err
+	}
+	defer cleanup2()
+
+	threshold := req.Threshold
+	if threshold <= 0 {
+		threshold = f.opts.verifyThreshold
+	}
+
+	antiSpoof := int32(0)
+	if req.AntiSpoofing {
+		antiSpoof = 1
+	}
+
+	started := time.Now()
+	var distance float32
+	var verified int32
+	rc := CppVerifyPaths(f.ctxPtr, path1, path2, threshold, antiSpoof,
+		unsafe.Pointer(&distance), unsafe.Pointer(&verified))
+	if rc != 0 {
+		return pb.FaceVerifyResponse{}, f.lastErr("verify", req.Img1[:min(8, len(req.Img1))]+"...")
+	}
+	elapsedMs := float32(time.Since(started).Seconds() * 1000.0)
+
+	// Confidence decays linearly from 100 at distance 0 to 0 at the threshold,
+	// matching the Python face backend's reporting.
+	confidence := float32(0)
+	if threshold > 0 {
+		confidence = float32(math.Max(0, math.Min(100, (1.0-float64(distance)/float64(threshold))*100.0)))
+	}
+
+	return pb.FaceVerifyResponse{
+		Verified:         verified != 0,
+		Distance:         distance,
+		Threshold:        threshold,
+		Confidence:       confidence,
+		Model:            f.opts.modelName,
+		Img1Area:         f.bestArea(path1),
+		Img2Area:         f.bestArea(path2),
+		ProcessingTimeMs: elapsedMs,
+	}, nil
+}
+
+// FaceAnalyze runs the genderage head on every detected face. The C-API returns
+// "M"/"F" gender labels and a rounded age; the labels are normalized to the
+// "Man"/"Woman" values the proto documents.
+func (f *FaceDetect) FaceAnalyze(req *pb.FaceAnalyzeRequest) (pb.FaceAnalyzeResponse, error) {
+	if f.ctxPtr == 0 {
+		return pb.FaceAnalyzeResponse{}, errors.New("face-detect: model not loaded")
+	}
+	if req.Img == "" {
+		return pb.FaceAnalyzeResponse{}, errors.New("face-detect: img is required")
+	}
+
+	path, cleanup, err := materializeImage(req.Img)
+	if err != nil {
+		return pb.FaceAnalyzeResponse{}, err
+	}
+	defer cleanup()
+
+	ptr := CppAnalyzeJSON(f.ctxPtr, path)
+	if ptr == 0 {
+		return pb.FaceAnalyzeResponse{}, f.lastErr("analyze", path)
+	}
+	defer CppFreeString(ptr)
+
+	faces, err := parseAnalyzeJSON(goStringFromCPtr(ptr))
+	if err != nil {
+		return pb.FaceAnalyzeResponse{}, fmt.Errorf("face-detect: analyze JSON: %w", err)
+	}
+	return pb.FaceAnalyzeResponse{Faces: faces}, nil
+}
+
+// faceBox is one entry of the detect/analyze JSON documents the engine emits.
+type faceBox struct {
+	Score  float32   `json:"score"`
+	Box    []float32 `json:"box"`
+	Age    float32   `json:"age"`
+	Gender string    `json:"gender"`
+}
+
+// xywh converts the engine's [x1,y1,x2,y2] box into the x/y/width/height the
+// proto carries. A short or missing box yields zeros.
+func (b faceBox) xywh() (x, y, w, h float32) {
+	if len(b.Box) < 4 {
+		return 0, 0, 0, 0
+	}
+	return b.Box[0], b.Box[1], b.Box[2] - b.Box[0], b.Box[3] - b.Box[1]
+}
+
+type facesJSON struct {
+	Faces []faceBox `json:"faces"`
+}
+
+func (f *FaceDetect) detectFaces(path string) ([]faceBox, error) {
+	ptr := CppDetectJSON(f.ctxPtr, path)
+	if ptr == 0 {
+		return nil, f.lastErr("detect", path)
+	}
+	defer CppFreeString(ptr)
+
+	var doc facesJSON
+	if err := json.Unmarshal([]byte(goStringFromCPtr(ptr)), &doc); err != nil {
+		return nil, fmt.Errorf("face-detect: detect JSON: %w", err)
+	}
+	return doc.Faces, nil
+}
+
+// bestArea returns the FacialArea of the highest-scoring face in an image, or an
+// empty area when detection fails or finds nothing. Best-effort: verify already
+// succeeded, so a missing region must not turn a valid match into an error.
+func (f *FaceDetect) bestArea(path string) *pb.FacialArea {
+	faces, err := f.detectFaces(path)
+	if err != nil || len(faces) == 0 {
+		return &pb.FacialArea{}
+	}
+	best := faces[0]
+	for _, fc := range faces[1:] {
+		if fc.Score > best.Score {
+			best = fc
+		}
+	}
+	x, y, w, h := best.xywh()
+	return &pb.FacialArea{X: x, Y: y, W: w, H: h}
+}
+
+// parseAnalyzeJSON maps the engine's analyze document onto FaceAnalysis entries.
+// The engine reports gender as "M"/"F"; both the dominant label and the score
+// map are filled with the "Man"/"Woman" form the proto documents.
+func parseAnalyzeJSON(doc string) ([]*pb.FaceAnalysis, error) {
+	var parsed facesJSON
+	if err := json.Unmarshal([]byte(doc), &parsed); err != nil {
+		return nil, err
+	}
+
+	out := make([]*pb.FaceAnalysis, 0, len(parsed.Faces))
+	for _, fc := range parsed.Faces {
+		x, y, w, h := fc.xywh()
+		fa := &pb.FaceAnalysis{
+			Region:         &pb.FacialArea{X: x, Y: y, W: w, H: h},
+			FaceConfidence: fc.Score,
+			Age:            fc.Age,
+		}
+		if label := normalizeGender(fc.Gender); label != "" {
+			fa.DominantGender = label
+			fa.Gender = map[string]float32{label: 1.0}
+		}
+		out = append(out, fa)
+	}
+	return out, nil
+}
+
+// normalizeGender maps the engine's "M"/"F" code to the "Man"/"Woman" labels the
+// proto documents. Unknown codes pass through unchanged.
+func normalizeGender(g string) string {
+	switch strings.ToUpper(strings.TrimSpace(g)) {
+	case "M":
+		return "Man"
+	case "F":
+		return "Woman"
+	case "":
+		return ""
+	default:
+		return g
+	}
+}
+
+// materializeImage decodes a base64 image payload into a temp file and returns
+// its path plus a cleanup func. As a convenience for callers that already pass a
+// filesystem path (e.g. a test fixture), an existing path is used as-is with a
+// no-op cleanup. data: URI prefixes are stripped before decoding.
+func materializeImage(src string) (path string, cleanup func(), err error) {
+	noop := func() {}
+	if src == "" {
+		return "", noop, errors.New("face-detect: empty image input")
+	}
+	if _, statErr := os.Stat(src); statErr == nil {
+		return src, noop, nil
+	}
+
+	payload := src
+	if i := strings.Index(payload, ","); strings.HasPrefix(payload, "data:") && i >= 0 {
+		payload = payload[i+1:]
+	}
+	data, decErr := base64.StdEncoding.DecodeString(strings.TrimSpace(payload))
+	if decErr != nil || len(data) == 0 {
+		return "", noop, errors.New("face-detect: image is neither an existing path nor valid base64")
+	}
+
+	tmp, createErr := os.CreateTemp("", "face-detect-*.img")
+	if createErr != nil {
+		return "", noop, fmt.Errorf("face-detect: create temp image: %w", createErr)
+	}
+	cleanup = func() { _ = os.Remove(tmp.Name()) }
+	if _, wErr := tmp.Write(data); wErr != nil {
+		_ = tmp.Close()
+		cleanup()
+		return "", noop, fmt.Errorf("face-detect: write temp image: %w", wErr)
+	}
+	if cErr := tmp.Close(); cErr != nil {
+		cleanup()
+		return "", noop, fmt.Errorf("face-detect: close temp image: %w", cErr)
+	}
+	return tmp.Name(), cleanup, nil
+}
+
+// lastErr wraps the C-API's per-ctx last-error buffer into a Go error.
+func (f *FaceDetect) lastErr(op, subject string) error {
+	msg := strings.TrimSpace(CppLastError(f.ctxPtr))
+	if msg == "" {
+		msg = "no error detail"
+	}
+	return fmt.Errorf("face-detect: %s failed for %q: %s", op, subject, msg)
+}
+
+// goStringFromCPtr copies a NUL-terminated C string into Go memory. cptr is a
+// malloc'd buffer the caller owns; release it via CppFreeString after the copy.
+//
+// The uintptr->Pointer conversion trips vet's unsafeptr check, which can't tell
+// a C heap pointer from Go-managed memory. Safe here: the GC neither tracks nor
+// moves the buffer and we dereference it immediately to copy the bytes out.
+func goStringFromCPtr(cptr uintptr) string {
+	if cptr == 0 {
+		return ""
+	}
+	p := unsafe.Pointer(cptr) //nolint:govet // C-owned malloc'd buffer, not Go-GC memory (see doc above)
+	n := 0
+	for *(*byte)(unsafe.Add(p, n)) != 0 {
+		n++
+	}
+	return string(unsafe.Slice((*byte)(p), n))
+}

backend/go/face-detect/gofacedetect_test.go

@@ -0,0 +1,230 @@
+package main
+
+import (
+	"encoding/base64"
+	"os"
+	"sync"
+	"testing"
+
+	"github.com/ebitengine/purego"
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+func TestFaceDetect(t *testing.T) {
+	RegisterFailHandler(Fail)
+	RunSpecs(t, "face-detect Backend Suite")
+}
+
+var (
+	libLoadOnce sync.Once
+	libLoadErr  error
+)
+
+// ensureLibLoaded mirrors main.go's bootstrap so a Go test can drive the C-API
+// bridge without spinning up the gRPC server. Records the error (the smoke
+// specs skip themselves) when libfacedetect.so is not loadable from cwd
+// (LD_LIBRARY_PATH or a symlink in ./).
+func ensureLibLoaded() error {
+	libLoadOnce.Do(func() {
+		libName := os.Getenv("FACEDETECT_LIBRARY")
+		if libName == "" {
+			libName = "libfacedetect.so"
+		}
+		lib, err := purego.Dlopen(libName, purego.RTLD_NOW|purego.RTLD_GLOBAL)
+		if err != nil {
+			libLoadErr = err
+			return
+		}
+		purego.RegisterLibFunc(&CppAbiVersion, lib, "facedetect_capi_abi_version")
+		purego.RegisterLibFunc(&CppLoad, lib, "facedetect_capi_load")
+		purego.RegisterLibFunc(&CppFree, lib, "facedetect_capi_free")
+		purego.RegisterLibFunc(&CppLastError, lib, "facedetect_capi_last_error")
+		purego.RegisterLibFunc(&CppFreeString, lib, "facedetect_capi_free_string")
+		purego.RegisterLibFunc(&CppFreeVec, lib, "facedetect_capi_free_vec")
+		purego.RegisterLibFunc(&CppEmbedPath, lib, "facedetect_capi_embed_path")
+		purego.RegisterLibFunc(&CppEmbedRGB, lib, "facedetect_capi_embed_rgb")
+		purego.RegisterLibFunc(&CppDetectJSON, lib, "facedetect_capi_detect_path_json")
+		purego.RegisterLibFunc(&CppVerifyPaths, lib, "facedetect_capi_verify_paths")
+		purego.RegisterLibFunc(&CppAnalyzeJSON, lib, "facedetect_capi_analyze_path_json")
+	})
+	return libLoadErr
+}
+
+var _ = Describe("parseOptions", func() {
+	It("defaults verify_threshold to 0.35", func() {
+		o := parseOptions(nil)
+		Expect(o.verifyThreshold).To(Equal(float32(0.35)))
+		Expect(o.modelName).To(Equal(""))
+	})
+
+	It("parses verify_threshold, threshold alias and model_name", func() {
+		o := parseOptions([]string{"verify_threshold:0.4", "model_name:buffalo_l", "unknown:x"})
+		Expect(o.verifyThreshold).To(Equal(float32(0.4)))
+		Expect(o.modelName).To(Equal("buffalo_l"))
+
+		o2 := parseOptions([]string{"threshold:0.3"})
+		Expect(o2.verifyThreshold).To(Equal(float32(0.3)))
+	})
+
+	It("ignores non-positive thresholds and keeps the default", func() {
+		o := parseOptions([]string{"verify_threshold:0", "threshold:-1"})
+		Expect(o.verifyThreshold).To(Equal(float32(0.35)))
+	})
+})
+
+var _ = Describe("normalizeGender", func() {
+	It("maps M/F codes to Man/Woman", func() {
+		Expect(normalizeGender("M")).To(Equal("Man"))
+		Expect(normalizeGender("f")).To(Equal("Woman"))
+		Expect(normalizeGender(" m ")).To(Equal("Man"))
+	})
+
+	It("passes empty and unknown codes through", func() {
+		Expect(normalizeGender("")).To(Equal(""))
+		Expect(normalizeGender("nonbinary")).To(Equal("nonbinary"))
+	})
+})
+
+var _ = Describe("faceBox.xywh", func() {
+	It("converts an [x1,y1,x2,y2] box to x/y/width/height", func() {
+		b := faceBox{Box: []float32{10, 20, 50, 80}}
+		x, y, w, h := b.xywh()
+		Expect(x).To(Equal(float32(10)))
+		Expect(y).To(Equal(float32(20)))
+		Expect(w).To(Equal(float32(40)))
+		Expect(h).To(Equal(float32(60)))
+	})
+
+	It("returns zeros for a short box", func() {
+		x, y, w, h := faceBox{Box: []float32{1, 2}}.xywh()
+		Expect([]float32{x, y, w, h}).To(Equal([]float32{0, 0, 0, 0}))
+	})
+})
+
+var _ = Describe("parseAnalyzeJSON", func() {
+	It("maps region, age and gender for each face", func() {
+		doc := `{"faces":[
+			{"score":0.997,"box":[10,20,50,80],"age":31,"gender":"M"},
+			{"score":0.81,"box":[0,0,40,40],"age":24,"gender":"F"}]}`
+		faces, err := parseAnalyzeJSON(doc)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(faces).To(HaveLen(2))
+
+		Expect(faces[0].FaceConfidence).To(BeNumerically("~", 0.997, 1e-4))
+		Expect(faces[0].Age).To(BeNumerically("~", 31, 1e-4))
+		Expect(faces[0].DominantGender).To(Equal("Man"))
+		Expect(faces[0].Gender).To(HaveKeyWithValue("Man", float32(1.0)))
+		Expect(faces[0].Region.W).To(Equal(float32(40)))
+		Expect(faces[0].Region.H).To(Equal(float32(60)))
+
+		Expect(faces[1].DominantGender).To(Equal("Woman"))
+	})
+
+	It("tolerates a missing gender field", func() {
+		faces, err := parseAnalyzeJSON(`{"faces":[{"score":0.5,"box":[0,0,10,10],"age":40}]}`)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(faces).To(HaveLen(1))
+		Expect(faces[0].DominantGender).To(Equal(""))
+		Expect(faces[0].Gender).To(BeEmpty())
+	})
+
+	It("returns no faces for an empty document", func() {
+		faces, err := parseAnalyzeJSON(`{"faces":[]}`)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(faces).To(BeEmpty())
+	})
+
+	It("returns an error on malformed JSON", func() {
+		_, err := parseAnalyzeJSON(`{not-json`)
+		Expect(err).To(HaveOccurred())
+	})
+})
+
+var _ = Describe("materializeImage", func() {
+	It("decodes a base64 payload to a temp file", func() {
+		payload := base64.StdEncoding.EncodeToString([]byte("\xff\xd8\xff\xe0fake-jpeg"))
+		path, cleanup, err := materializeImage(payload)
+		Expect(err).ToNot(HaveOccurred())
+		defer cleanup()
+		data, rerr := os.ReadFile(path)
+		Expect(rerr).ToNot(HaveOccurred())
+		Expect(data).To(Equal([]byte("\xff\xd8\xff\xe0fake-jpeg")))
+	})
+
+	It("strips a data: URI prefix before decoding", func() {
+		payload := "data:image/png;base64," + base64.StdEncoding.EncodeToString([]byte("hello"))
+		path, cleanup, err := materializeImage(payload)
+		Expect(err).ToNot(HaveOccurred())
+		defer cleanup()
+		data, rerr := os.ReadFile(path)
+		Expect(rerr).ToNot(HaveOccurred())
+		Expect(data).To(Equal([]byte("hello")))
+	})
+
+	It("uses an existing path as-is", func() {
+		tmp, err := os.CreateTemp("", "face-detect-fixture-*.bin")
+		Expect(err).ToNot(HaveOccurred())
+		defer func() { _ = os.Remove(tmp.Name()) }()
+		Expect(tmp.Close()).To(Succeed())
+
+		path, cleanup, err := materializeImage(tmp.Name())
+		Expect(err).ToNot(HaveOccurred())
+		defer cleanup()
+		Expect(path).To(Equal(tmp.Name()))
+	})
+
+	It("errors on input that is neither a path nor base64", func() {
+		_, _, err := materializeImage("not base64!!!")
+		Expect(err).To(HaveOccurred())
+	})
+})
+
+// The specs below exercise the real C-API end to end. They run only when both a
+// model GGUF and a test image are provided, and skip cleanly otherwise so the
+// suite stays green without large assets.
+var _ = Describe("FaceDetect end-to-end", Ordered, func() {
+	var (
+		f         *FaceDetect
+		modelPath = os.Getenv("FACEDETECT_BACKEND_TEST_MODEL")
+		imagePath = os.Getenv("FACEDETECT_BACKEND_TEST_IMAGE")
+	)
+
+	BeforeAll(func() {
+		if modelPath == "" || imagePath == "" {
+			Skip("set FACEDETECT_BACKEND_TEST_MODEL and FACEDETECT_BACKEND_TEST_IMAGE to run the e2e specs")
+		}
+		if err := ensureLibLoaded(); err != nil {
+			Skip("libfacedetect.so not loadable: " + err.Error())
+		}
+		f = &FaceDetect{}
+		Expect(f.Load(&pb.ModelOptions{ModelFile: modelPath})).To(Succeed())
+	})
+
+	It("embeds the primary face in an image", func() {
+		emb, err := f.Embeddings(&pb.PredictOptions{Images: []string{imagePath}})
+		Expect(err).ToNot(HaveOccurred())
+		Expect(emb).ToNot(BeEmpty())
+	})
+
+	It("detects at least one face", func() {
+		resp, err := f.Detect(&pb.DetectOptions{Src: imagePath})
+		Expect(err).ToNot(HaveOccurred())
+		Expect(resp.Detections).ToNot(BeEmpty())
+		Expect(resp.Detections[0].ClassName).To(Equal("face"))
+	})
+
+	It("verifies an image against itself as the same identity", func() {
+		resp, err := f.FaceVerify(&pb.FaceVerifyRequest{Img1: imagePath, Img2: imagePath})
+		Expect(err).ToNot(HaveOccurred())
+		Expect(resp.Verified).To(BeTrue())
+		Expect(resp.Distance).To(BeNumerically("<=", resp.Threshold))
+	})
+
+	It("analyzes age/gender for each face", func() {
+		resp, err := f.FaceAnalyze(&pb.FaceAnalyzeRequest{Img: imagePath})
+		Expect(err).ToNot(HaveOccurred())
+		Expect(resp.Faces).ToNot(BeEmpty())
+	})
+})

backend/go/face-detect/main.go

@@ -0,0 +1,65 @@
+package main
+
+// Started internally by LocalAI - one gRPC server per loaded model.
+//
+// Loads libfacedetect.so via purego and registers the flat C-API entry points
+// declared in facedetect_capi.h. The library name can be overridden with
+// FACEDETECT_LIBRARY (mirrors the VOICEDETECT_LIBRARY / PARAKEET_LIBRARY
+// convention in the sibling backends); the default looks for the .so next to
+// this binary (resolved via LD_LIBRARY_PATH by run.sh).
+import (
+	"flag"
+	"fmt"
+	"os"
+
+	"github.com/ebitengine/purego"
+	grpc "github.com/mudler/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+type LibFuncs struct {
+	FuncPtr any
+	Name    string
+}
+
+func main() {
+	libName := os.Getenv("FACEDETECT_LIBRARY")
+	if libName == "" {
+		libName = "libfacedetect.so"
+	}
+
+	lib, err := purego.Dlopen(libName, purego.RTLD_NOW|purego.RTLD_GLOBAL)
+	if err != nil {
+		panic(fmt.Errorf("face-detect: dlopen %q: %w", libName, err))
+	}
+
+	// Bound 1:1 to facedetect_capi.h. char*/float* returns are registered as
+	// uintptr so the raw pointer can be freed via the matching capi free fn.
+	libFuncs := []LibFuncs{
+		{&CppAbiVersion, "facedetect_capi_abi_version"},
+		{&CppLoad, "facedetect_capi_load"},
+		{&CppFree, "facedetect_capi_free"},
+		{&CppLastError, "facedetect_capi_last_error"},
+		{&CppFreeString, "facedetect_capi_free_string"},
+		{&CppFreeVec, "facedetect_capi_free_vec"},
+		{&CppEmbedPath, "facedetect_capi_embed_path"},
+		{&CppEmbedRGB, "facedetect_capi_embed_rgb"},
+		{&CppDetectJSON, "facedetect_capi_detect_path_json"},
+		{&CppVerifyPaths, "facedetect_capi_verify_paths"},
+		{&CppAnalyzeJSON, "facedetect_capi_analyze_path_json"},
+	}
+	for _, lf := range libFuncs {
+		purego.RegisterLibFunc(lf.FuncPtr, lib, lf.Name)
+	}
+
+	fmt.Fprintf(os.Stderr, "[face-detect] ABI=%d\n", CppAbiVersion())
+
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &FaceDetect{}); err != nil {
+		panic(err)
+	}
+}

backend/go/face-detect/options.go

@@ -0,0 +1,47 @@
+package main
+
+import (
+	"strconv"
+	"strings"
+)
+
+// defaultVerifyThreshold is the cosine-distance cutoff used when a request does
+// not set one. Matches the insightface buffalo_l ArcFace R50 default the Python
+// face backend ships with so the two implementations agree on verdicts out of
+// the box.
+const defaultVerifyThreshold float32 = 0.35
+
+// loadOptions holds the parsed model-level options for face-detect.
+type loadOptions struct {
+	verifyThreshold float32
+	modelName       string
+}
+
+func splitOption(o string) (key, value string, ok bool) {
+	i := strings.Index(o, ":")
+	if i < 0 {
+		return "", "", false
+	}
+	return strings.TrimSpace(o[:i]), strings.TrimSpace(o[i+1:]), true
+}
+
+// parseOptions reads the backend "key:value" option slice. Unknown keys are
+// ignored. Defaults: verify_threshold 0.35, model_name derived from the file.
+func parseOptions(opts []string) loadOptions {
+	o := loadOptions{verifyThreshold: defaultVerifyThreshold}
+	for _, oo := range opts {
+		key, value, ok := splitOption(oo)
+		if !ok {
+			continue
+		}
+		switch key {
+		case "verify_threshold", "threshold":
+			if f, err := strconv.ParseFloat(value, 32); err == nil && f > 0 {
+				o.verifyThreshold = float32(f)
+			}
+		case "model_name":
+			o.modelName = value
+		}
+	}
+	return o
+}

backend/go/face-detect/package.sh

@@ -0,0 +1,68 @@
+#!/bin/bash
+#
+# Bundle the face-detect-grpc binary, libfacedetect.so, the core runtime libs
+# (libc/libstdc++/libgomp + ld.so) and the GPU runtime for the active BUILD_TYPE
+# so the package is self-contained. Mirrors backend/go/voice-detect/package.sh;
+# run.sh routes the (CGO_ENABLED=0) binary through lib/ld.so so the packaged libc
+# is used instead of the host's.
+
+set -e
+
+CURDIR=$(dirname "$(realpath "$0")")
+REPO_ROOT="${CURDIR}/../../.."
+
+mkdir -p "$CURDIR/package/lib"
+
+cp -avf "$CURDIR/face-detect-grpc" "$CURDIR/package/"
+cp -avf "$CURDIR/run.sh" "$CURDIR/package/"
+
+# libfacedetect.so + any soname symlinks. purego.Dlopen resolves it via
+# LD_LIBRARY_PATH, which run.sh points at lib/.
+cp -avf "$CURDIR"/libfacedetect.so* "$CURDIR/package/lib/" 2>/dev/null || {
+	echo "ERROR: libfacedetect.so not found in $CURDIR, run 'make' first" >&2
+	exit 1
+}
+
+# Detect architecture and copy the core runtime libs libfacedetect.so links
+# against, plus the matching dynamic loader as lib/ld.so.
+if [ -f "/lib64/ld-linux-x86-64.so.2" ]; then
+    echo "Detected x86_64 architecture, copying x86_64 libraries..."
+    cp -arfLv /lib64/ld-linux-x86-64.so.2 "$CURDIR/package/lib/ld.so"
+    cp -arfLv /lib/x86_64-linux-gnu/libc.so.6 "$CURDIR/package/lib/libc.so.6"
+    cp -arfLv /lib/x86_64-linux-gnu/libgcc_s.so.1 "$CURDIR/package/lib/libgcc_s.so.1"
+    cp -arfLv /lib/x86_64-linux-gnu/libstdc++.so.6 "$CURDIR/package/lib/libstdc++.so.6"
+    cp -arfLv /lib/x86_64-linux-gnu/libm.so.6 "$CURDIR/package/lib/libm.so.6"
+    cp -arfLv /lib/x86_64-linux-gnu/libgomp.so.1 "$CURDIR/package/lib/libgomp.so.1"
+    cp -arfLv /lib/x86_64-linux-gnu/libdl.so.2 "$CURDIR/package/lib/libdl.so.2"
+    cp -arfLv /lib/x86_64-linux-gnu/librt.so.1 "$CURDIR/package/lib/librt.so.1"
+    cp -arfLv /lib/x86_64-linux-gnu/libpthread.so.0 "$CURDIR/package/lib/libpthread.so.0"
+elif [ -f "/lib/ld-linux-aarch64.so.1" ]; then
+    echo "Detected ARM64 architecture, copying ARM64 libraries..."
+    cp -arfLv /lib/ld-linux-aarch64.so.1 "$CURDIR/package/lib/ld.so"
+    cp -arfLv /lib/aarch64-linux-gnu/libc.so.6 "$CURDIR/package/lib/libc.so.6"
+    cp -arfLv /lib/aarch64-linux-gnu/libgcc_s.so.1 "$CURDIR/package/lib/libgcc_s.so.1"
+    cp -arfLv /lib/aarch64-linux-gnu/libstdc++.so.6 "$CURDIR/package/lib/libstdc++.so.6"
+    cp -arfLv /lib/aarch64-linux-gnu/libm.so.6 "$CURDIR/package/lib/libm.so.6"
+    cp -arfLv /lib/aarch64-linux-gnu/libgomp.so.1 "$CURDIR/package/lib/libgomp.so.1"
+    cp -arfLv /lib/aarch64-linux-gnu/libdl.so.2 "$CURDIR/package/lib/libdl.so.2"
+    cp -arfLv /lib/aarch64-linux-gnu/librt.so.1 "$CURDIR/package/lib/librt.so.1"
+    cp -arfLv /lib/aarch64-linux-gnu/libpthread.so.0 "$CURDIR/package/lib/libpthread.so.0"
+elif [ "$(uname -s)" = "Darwin" ]; then
+    echo "Detected Darwin"
+else
+    echo "Error: Could not detect architecture"
+    exit 1
+fi
+
+# Package GPU libraries (CUDA/ROCm/Intel/Vulkan loader + ICDs + drivers) based on
+# BUILD_TYPE so the backend can reach the GPU without the runtime base image
+# shipping those drivers.
+GPU_LIB_SCRIPT="${REPO_ROOT}/scripts/build/package-gpu-libs.sh"
+if [ -f "$GPU_LIB_SCRIPT" ]; then
+    echo "Packaging GPU libraries for BUILD_TYPE=${BUILD_TYPE:-cpu}..."
+    source "$GPU_LIB_SCRIPT" "$CURDIR/package/lib"
+    package_gpu_libs
+fi
+
+echo "Packaging completed successfully"
+ls -liah "$CURDIR/package/" "$CURDIR/package/lib/"

backend/go/face-detect/run.sh

@@ -0,0 +1,16 @@
+#!/bin/bash
+set -e
+
+CURDIR=$(dirname "$(realpath "$0")")
+
+export LD_LIBRARY_PATH="$CURDIR/lib:$CURDIR:${LD_LIBRARY_PATH:-}"
+
+# If a self-contained ld.so was packaged, route through it so the packaged
+# libc / libstdc++ are used instead of the host's (matches the voice-detect /
+# whisper / parakeet backends' runtime layout).
+if [ -f "$CURDIR/lib/ld.so" ]; then
+	echo "Using lib/ld.so"
+	exec "$CURDIR/lib/ld.so" "$CURDIR/face-detect-grpc" "$@"
+fi
+
+exec "$CURDIR/face-detect-grpc" "$@"

backend/go/face-detect/test.sh

@@ -0,0 +1,15 @@
+#!/bin/bash
+set -e
+
+CURDIR=$(dirname "$(realpath "$0")")
+cd "$CURDIR"
+
+echo "Running face-detect backend tests..."
+
+# The pure-Go parsing specs always run. The embed/detect/verify/analyze smoke
+# specs run only when a model + image are provided via
+# FACEDETECT_BACKEND_TEST_MODEL and FACEDETECT_BACKEND_TEST_IMAGE; otherwise they
+# auto-skip.
+LD_LIBRARY_PATH="$CURDIR:${LD_LIBRARY_PATH:-}" go test -v -timeout 1200s .
+
+echo "face-detect tests completed."

backend/go/local-store/run.sh

@@ -1,6 +1,6 @@
 #!/bin/bash
 set -ex
 
-CURDIR=$(dirname "$(realpath $0)")
+CURDIR=$(dirname "$(realpath "$0")")
 
-exec $CURDIR/local-store "$@"
+exec "$CURDIR"/local-store "$@"

backend/go/localvqe/Makefile

@@ -32,6 +32,8 @@ endif
 ifeq ($(BUILD_TYPE),vulkan)
 	CMAKE_ARGS+=-DGGML_VULKAN=ON -DLOCALVQE_VULKAN=ON
 else ifeq ($(OS),Darwin)
+	# Apple Silicon: CPU-only (no Metal upstream); built + published as an arm64
+	# image by CI (includeDarwin in .github/backend-matrix.yml) for macOS install.
 	CMAKE_ARGS+=-DGGML_METAL=OFF
 endif
 
@@ -67,8 +69,9 @@ $(LIB_SENTINEL): sources/LocalVQE
 	# that the loader picks at runtime. We must build every target — the
 	# default `--target localvqe_shared` drops these. CMAKE_LIBRARY_OUTPUT_DIRECTORY
 	# routes all of them into build/bin; copy them out next to the binary.
-	cp -P build/bin/liblocalvqe.so* . 2>/dev/null || cp -P build/liblocalvqe.so* .
+	cp -P build/bin/liblocalvqe.so* . 2>/dev/null || cp -P build/bin/liblocalvqe.dylib . 2>/dev/null || cp -P build/liblocalvqe.so* . 2>/dev/null || cp -P build/liblocalvqe.dylib .
 	cp -P build/bin/libggml*.so* . 2>/dev/null || true
+	cp -P build/bin/libggml*.dylib . 2>/dev/null || true
 	touch $(LIB_SENTINEL)
 
 liblocalvqe.so: $(LIB_SENTINEL)

backend/go/localvqe/main.go

@@ -4,6 +4,7 @@ package main
 import (
 	"flag"
 	"os"
+	"runtime"
 
 	"github.com/ebitengine/purego"
 	grpc "github.com/mudler/LocalAI/pkg/grpc"
@@ -21,7 +22,11 @@ type LibFuncs struct {
 func main() {
 	libName := os.Getenv("LOCALVQE_LIBRARY")
 	if libName == "" {
-		libName = "./liblocalvqe.so"
+		if runtime.GOOS == "darwin" {
+			libName = "./liblocalvqe.dylib"
+		} else {
+			libName = "./liblocalvqe.so"
+		}
 	}
 
 	lib, err := purego.Dlopen(libName, purego.RTLD_NOW|purego.RTLD_GLOBAL)

backend/go/localvqe/package.sh

@@ -15,7 +15,9 @@ cp -avf $CURDIR/localvqe $CURDIR/package/
 # liblocalvqe.so* (with SOVERSION symlinks) and the libggml-*.so runtime
 # variants — LocalVQE picks the matching CPU variant at load time.
 cp -P $CURDIR/liblocalvqe.so* $CURDIR/package/ 2>/dev/null || true
+cp -P $CURDIR/liblocalvqe.dylib $CURDIR/package/ 2>/dev/null || true
 cp -P $CURDIR/libggml*.so* $CURDIR/package/ 2>/dev/null || true
+cp -P $CURDIR/libggml*.dylib $CURDIR/package/ 2>/dev/null || true
 cp -fv $CURDIR/run.sh $CURDIR/package/
 
 # Detect architecture and copy appropriate libraries

backend/go/localvqe/run.sh

@@ -1,23 +1,34 @@
 #!/bin/bash
 set -ex
 
-CURDIR=$(dirname "$(realpath $0)")
+CURDIR=$(dirname "$(realpath "$0")")
 
 # LocalVQE's runtime CPU-variant loader (ggml_backend_load_all) searches
 # get_executable_path() and current_path() — the second one is what saves us
 # when /proc/self/exe resolves to lib/ld.so under the bundled-loader path.
-# So we cd into $CURDIR (where all the libggml-cpu-*.so files live) before
+# So we cd into "$CURDIR" (where all the libggml-cpu-*.so files live) before
 # exec'ing the binary.
 cd "$CURDIR"
 
-export LD_LIBRARY_PATH=$CURDIR:$CURDIR/lib:$LD_LIBRARY_PATH
-export LOCALVQE_LIBRARY=$CURDIR/liblocalvqe.so
+if [ "$(uname)" = "Darwin" ]; then
+	# macOS: LocalVQE is built as a SHARED library, so dyld needs the .dylib +
+	# DYLD_LIBRARY_PATH. Prefer .dylib and fall back to .so just in case.
+	export DYLD_LIBRARY_PATH="$CURDIR":"$CURDIR"/lib:$DYLD_LIBRARY_PATH
+	LOCALVQE_LIBRARY="$CURDIR"/liblocalvqe.dylib
+	if [ ! -e "$LOCALVQE_LIBRARY" ]; then
+		LOCALVQE_LIBRARY="$CURDIR"/liblocalvqe.so
+	fi
+	export LOCALVQE_LIBRARY
+else
+	export LD_LIBRARY_PATH="$CURDIR":"$CURDIR"/lib:$LD_LIBRARY_PATH
+	export LOCALVQE_LIBRARY="$CURDIR"/liblocalvqe.so
+fi
 
-if [ -f $CURDIR/lib/ld.so ]; then
+if [ -f "$CURDIR"/lib/ld.so ]; then
 	echo "Using lib/ld.so"
 	echo "Using library: $LOCALVQE_LIBRARY"
-	exec $CURDIR/lib/ld.so $CURDIR/localvqe "$@"
+	exec "$CURDIR"/lib/ld.so "$CURDIR"/localvqe "$@"
 fi
 
 echo "Using library: $LOCALVQE_LIBRARY"
-exec $CURDIR/localvqe "$@"
+exec "$CURDIR"/localvqe "$@"

backend/go/locate-anything-cpp/Makefile

@@ -33,6 +33,8 @@ else ifeq ($(BUILD_TYPE),hipblas)
 else ifeq ($(BUILD_TYPE),vulkan)
 	CMAKE_ARGS+=-DGGML_VULKAN=ON -DLA_GGML_VULKAN=ON
 else ifeq ($(OS),Darwin)
+	# macOS/Metal: built + published as an OCI image by CI (includeDarwin in
+	# .github/backend-matrix.yml) so Apple Silicon users can install this backend.
 	ifneq ($(BUILD_TYPE),metal)
 		CMAKE_ARGS+=-DGGML_METAL=OFF
 	else
@@ -70,7 +72,7 @@ ifeq ($(UNAME_S),Linux)
 	VARIANT_TARGETS = liblocateanythingcpp-avx.so liblocateanythingcpp-avx2.so liblocateanythingcpp-avx512.so liblocateanythingcpp-fallback.so
 else
 	# On non-Linux (e.g., Darwin), build only fallback variant
-	VARIANT_TARGETS = liblocateanythingcpp-fallback.so
+	VARIANT_TARGETS = liblocateanythingcpp-fallback.dylib
 endif
 
 locate-anything-cpp: main.go golocateanythingcpp.go $(VARIANT_TARGETS)
@@ -82,7 +84,7 @@ package: locate-anything-cpp
 build: package
 
 clean: purge
-	rm -rf liblocateanythingcpp*.so locate-anything-cpp package sources
+	rm -rf liblocateanythingcpp*.so liblocateanythingcpp*.dylib locate-anything-cpp package sources
 
 purge:
 	rm -rf build*
@@ -109,11 +111,19 @@ liblocateanythingcpp-avx512.so: sources/locate-anything.cpp
 endif
 
 # Build fallback variant (all platforms)
+ifeq ($(UNAME_S),Darwin)
+liblocateanythingcpp-fallback.dylib: sources/locate-anything.cpp
+	rm -rfv build-$@
+	$(info ${GREEN}I locate-anything-cpp build info:fallback${RESET})
+	SO_TARGET=$@ CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off" $(MAKE) liblocateanythingcpp-custom
+	rm -rfv build-$@
+else
 liblocateanythingcpp-fallback.so: sources/locate-anything.cpp
 	rm -rfv build-$@
 	$(info ${GREEN}I locate-anything-cpp build info:fallback${RESET})
 	SO_TARGET=$@ CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off" $(MAKE) liblocateanythingcpp-custom
 	rm -rfv build-$@
+endif
 
 liblocateanythingcpp-custom: CMakeLists.txt
 	mkdir -p build-$(SO_TARGET) && \
@@ -121,7 +131,8 @@ liblocateanythingcpp-custom: CMakeLists.txt
 	cmake .. $(CMAKE_ARGS) && \
 	cmake --build . --config Release -j$(JOBS) && \
 	cd .. && \
-	mv build-$(SO_TARGET)/liblocateanythingcpp.so ./$(SO_TARGET)
+	(mv build-$(SO_TARGET)/liblocateanythingcpp.so ./$(SO_TARGET) 2>/dev/null || \
+	 mv build-$(SO_TARGET)/liblocateanythingcpp.dylib ./$(SO_TARGET) 2>/dev/null)
 
 all: locate-anything-cpp package
 

backend/go/locate-anything-cpp/main.go

@@ -9,6 +9,7 @@ package main
 import (
 	"flag"
 	"os"
+	"runtime"
 
 	"github.com/ebitengine/purego"
 	grpc "github.com/mudler/LocalAI/pkg/grpc"
@@ -27,7 +28,11 @@ func main() {
 	// Get library name from environment variable, default to fallback
 	libName := os.Getenv("LOCATEANYTHING_LIBRARY")
 	if libName == "" {
-		libName = "./liblocateanythingcpp-fallback.so"
+		if runtime.GOOS == "darwin" {
+			libName = "./liblocateanythingcpp-fallback.dylib"
+		} else {
+			libName = "./liblocateanythingcpp-fallback.so"
+		}
 	}
 
 	lib, err := purego.Dlopen(libName, purego.RTLD_NOW|purego.RTLD_GLOBAL)

backend/go/locate-anything-cpp/package.sh

@@ -10,7 +10,8 @@ REPO_ROOT="${CURDIR}/../../.."
 # Create lib directory
 mkdir -p $CURDIR/package/lib
 
-cp -avf $CURDIR/liblocateanythingcpp-*.so $CURDIR/package/
+cp -fv $CURDIR/liblocateanythingcpp-*.so $CURDIR/package/ 2>/dev/null || true
+cp -fv $CURDIR/liblocateanythingcpp-*.dylib $CURDIR/package/ 2>/dev/null || true
 cp -avf $CURDIR/locate-anything-cpp $CURDIR/package/
 cp -fv $CURDIR/run.sh $CURDIR/package/
 

backend/go/locate-anything-cpp/run.sh

@@ -2,7 +2,7 @@
 set -ex
 
 # Get the absolute current dir where the script is located
-CURDIR=$(dirname "$(realpath $0)")
+CURDIR=$(dirname "$(realpath "$0")")
 
 cd /
 
@@ -12,19 +12,23 @@ if [ "$(uname)" != "Darwin" ]; then
 	grep -e "flags" /proc/cpuinfo | head -1
 fi
 
-LIBRARY="$CURDIR/liblocateanythingcpp-fallback.so"
+if [ "$(uname)" = "Darwin" ]; then
+	# macOS: single dylib variant (Metal or Accelerate)
+	LIBRARY="$CURDIR/liblocateanythingcpp-fallback.dylib"
+	export DYLD_LIBRARY_PATH="$CURDIR"/lib:$DYLD_LIBRARY_PATH
+else
+	LIBRARY="$CURDIR/liblocateanythingcpp-fallback.so"
 
-if [ "$(uname)" != "Darwin" ]; then
 	if grep -q -e "\savx\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX    found OK"
-		if [ -e $CURDIR/liblocateanythingcpp-avx.so ]; then
+		if [ -e "$CURDIR"/liblocateanythingcpp-avx.so ]; then
 			LIBRARY="$CURDIR/liblocateanythingcpp-avx.so"
 		fi
 	fi
 
 	if grep -q -e "\savx2\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX2   found OK"
-		if [ -e $CURDIR/liblocateanythingcpp-avx2.so ]; then
+		if [ -e "$CURDIR"/liblocateanythingcpp-avx2.so ]; then
 			LIBRARY="$CURDIR/liblocateanythingcpp-avx2.so"
 		fi
 	fi
@@ -32,21 +36,22 @@ if [ "$(uname)" != "Darwin" ]; then
 	# Check avx 512
 	if grep -q -e "\savx512f\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX512F found OK"
-		if [ -e $CURDIR/liblocateanythingcpp-avx512.so ]; then
+		if [ -e "$CURDIR"/liblocateanythingcpp-avx512.so ]; then
 			LIBRARY="$CURDIR/liblocateanythingcpp-avx512.so"
 		fi
 	fi
+
+	export LD_LIBRARY_PATH="$CURDIR"/lib:$LD_LIBRARY_PATH
 fi
 
-export LD_LIBRARY_PATH=$CURDIR/lib:$LD_LIBRARY_PATH
 export LOCATEANYTHING_LIBRARY=$LIBRARY
 
 # If there is a lib/ld.so, use it
-if [ -f $CURDIR/lib/ld.so ]; then
+if [ -f "$CURDIR"/lib/ld.so ]; then
 	echo "Using lib/ld.so"
 	echo "Using library: $LIBRARY"
-	exec $CURDIR/lib/ld.so $CURDIR/locate-anything-cpp "$@"
+	exec "$CURDIR"/lib/ld.so "$CURDIR"/locate-anything-cpp "$@"
 fi
 
 echo "Using library: $LIBRARY"
-exec $CURDIR/locate-anything-cpp "$@"
+exec "$CURDIR"/locate-anything-cpp "$@"

backend/go/omnivoice-cpp/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # omnivoice.cpp version
 OMNIVOICE_REPO?=https://github.com/ServeurpersoCom/omnivoice.cpp
-OMNIVOICE_VERSION?=96d30169afd5e6bb3fd6a0e9be0eb505bfe81fcd
+OMNIVOICE_VERSION?=0f37401bebe9b20c0160a888e592108fc1d17607
 SO_TARGET?=libgomnivoicecpp.so
 
 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF
@@ -65,7 +65,8 @@ UNAME_S := $(shell uname -s)
 ifeq ($(UNAME_S),Linux)
 	VARIANT_TARGETS = libgomnivoicecpp-avx.so libgomnivoicecpp-avx2.so libgomnivoicecpp-avx512.so libgomnivoicecpp-fallback.so
 else
-	VARIANT_TARGETS = libgomnivoicecpp-fallback.so
+	# On non-Linux (e.g., Darwin), build only fallback variant (as a dylib)
+	VARIANT_TARGETS = libgomnivoicecpp-fallback.dylib
 endif
 
 omnivoice-cpp: main.go gomnivoicecpp.go $(VARIANT_TARGETS)
@@ -77,7 +78,7 @@ package: omnivoice-cpp
 build: package
 
 clean: purge
-	rm -rf libgomnivoicecpp*.so package sources/omnivoice.cpp omnivoice-cpp
+	rm -rf libgomnivoicecpp*.so libgomnivoicecpp*.dylib package sources/omnivoice.cpp omnivoice-cpp
 
 purge:
 	rm -rf build*
@@ -106,13 +107,20 @@ libgomnivoicecpp-fallback.so: sources/omnivoice.cpp
 	SO_TARGET=libgomnivoicecpp-fallback.so CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off" $(MAKE) libgomnivoicecpp-custom
 	rm -rf build-libgomnivoicecpp-fallback.so
 
+# Build fallback variant as a dylib (Darwin)
+libgomnivoicecpp-fallback.dylib: sources/omnivoice.cpp
+	$(info ${GREEN}I omnivoice-cpp build info:fallback (dylib)${RESET})
+	SO_TARGET=libgomnivoicecpp-fallback.dylib CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off" $(MAKE) libgomnivoicecpp-custom
+	rm -rf build-libgomnivoicecpp-fallback.dylib
+
 libgomnivoicecpp-custom: CMakeLists.txt cpp/gomnivoicecpp.cpp cpp/gomnivoicecpp.h
 	mkdir -p build-$(SO_TARGET) && \
 	cd build-$(SO_TARGET) && \
 	cmake .. $(CMAKE_ARGS) && \
 	cmake --build . --config Release -j$(JOBS) --target gomnivoicecpp && \
 	cd .. && \
-	mv build-$(SO_TARGET)/libgomnivoicecpp.so ./$(SO_TARGET)
+	(mv build-$(SO_TARGET)/libgomnivoicecpp.so ./$(SO_TARGET) 2>/dev/null || \
+	 mv build-$(SO_TARGET)/libgomnivoicecpp.dylib ./$(SO_TARGET) 2>/dev/null)
 
 test: omnivoice-cpp
 	@echo "Running omnivoice-cpp tests..."

backend/go/omnivoice-cpp/main.go

@@ -4,6 +4,7 @@ package main
 import (
 	"flag"
 	"os"
+	"runtime"
 
 	"github.com/ebitengine/purego"
 	grpc "github.com/mudler/LocalAI/pkg/grpc"
@@ -21,7 +22,11 @@ type LibFuncs struct {
 func main() {
 	libName := os.Getenv("OMNIVOICE_LIBRARY")
 	if libName == "" {
-		libName = "./libgomnivoicecpp-fallback.so"
+		if runtime.GOOS == "darwin" {
+			libName = "./libgomnivoicecpp-fallback.dylib"
+		} else {
+			libName = "./libgomnivoicecpp-fallback.so"
+		}
 	}
 
 	lib, err := purego.Dlopen(libName, purego.RTLD_NOW|purego.RTLD_GLOBAL)

backend/go/omnivoice-cpp/package.sh

@@ -12,7 +12,8 @@ REPO_ROOT="${CURDIR}/../../.."
 mkdir -p $CURDIR/package/lib
 
 cp -avf $CURDIR/omnivoice-cpp $CURDIR/package/
-cp -fv $CURDIR/libgomnivoicecpp-*.so $CURDIR/package/
+cp -fv $CURDIR/libgomnivoicecpp-*.so $CURDIR/package/ 2>/dev/null || true
+cp -fv $CURDIR/libgomnivoicecpp-*.dylib $CURDIR/package/ 2>/dev/null || true
 cp -fv $CURDIR/run.sh $CURDIR/package/
 
 # Detect architecture and copy appropriate libraries

backend/go/omnivoice-cpp/run.sh

@@ -2,7 +2,7 @@
 set -ex
 
 # Get the absolute current dir where the script is located
-CURDIR=$(dirname "$(realpath $0)")
+CURDIR=$(dirname "$(realpath "$0")")
 
 cd /
 
@@ -12,19 +12,23 @@ if [ "$(uname)" != "Darwin" ]; then
 	grep -e "flags" /proc/cpuinfo | head -1
 fi
 
-LIBRARY="$CURDIR/libgomnivoicecpp-fallback.so"
+if [ "$(uname)" = "Darwin" ]; then
+	# macOS: single dylib variant (Metal or Accelerate)
+	LIBRARY="$CURDIR/libgomnivoicecpp-fallback.dylib"
+	export DYLD_LIBRARY_PATH="$CURDIR"/lib:$DYLD_LIBRARY_PATH
+else
+	LIBRARY="$CURDIR/libgomnivoicecpp-fallback.so"
 
-if [ "$(uname)" != "Darwin" ]; then
 	if grep -q -e "\savx\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX    found OK"
-		if [ -e $CURDIR/libgomnivoicecpp-avx.so ]; then
+		if [ -e "$CURDIR"/libgomnivoicecpp-avx.so ]; then
 			LIBRARY="$CURDIR/libgomnivoicecpp-avx.so"
 		fi
 	fi
 
 	if grep -q -e "\savx2\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX2   found OK"
-		if [ -e $CURDIR/libgomnivoicecpp-avx2.so ]; then
+		if [ -e "$CURDIR"/libgomnivoicecpp-avx2.so ]; then
 			LIBRARY="$CURDIR/libgomnivoicecpp-avx2.so"
 		fi
 	fi
@@ -32,21 +36,22 @@ if [ "$(uname)" != "Darwin" ]; then
 	# Check avx 512
 	if grep -q -e "\savx512f\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX512F found OK"
-		if [ -e $CURDIR/libgomnivoicecpp-avx512.so ]; then
+		if [ -e "$CURDIR"/libgomnivoicecpp-avx512.so ]; then
 			LIBRARY="$CURDIR/libgomnivoicecpp-avx512.so"
 		fi
 	fi
+
+	export LD_LIBRARY_PATH="$CURDIR"/lib:$LD_LIBRARY_PATH
 fi
 
-export LD_LIBRARY_PATH=$CURDIR/lib:$LD_LIBRARY_PATH
 export OMNIVOICE_LIBRARY=$LIBRARY
 
 # If there is a lib/ld.so, use it
-if [ -f $CURDIR/lib/ld.so ]; then
+if [ -f "$CURDIR"/lib/ld.so ]; then
 	echo "Using lib/ld.so"
 	echo "Using library: $LIBRARY"
-	exec $CURDIR/lib/ld.so $CURDIR/omnivoice-cpp "$@"
+	exec "$CURDIR"/lib/ld.so "$CURDIR"/omnivoice-cpp "$@"
 fi
 
 echo "Using library: $LIBRARY"
-exec $CURDIR/omnivoice-cpp "$@"
+exec "$CURDIR"/omnivoice-cpp "$@"

backend/go/opus/Makefile

@@ -1,13 +1,30 @@
 GOCMD?=go
 GO_TAGS?=
 
+# The opus shim is a small C wrapper around libopus' variadic
+# opus_encoder_ctl (see csrc/opus_shim.c). It is built as a shared library
+# and dlopen'd at runtime by the Go backend (codec.go). The extension is
+# OS-specific: Linux uses .so, macOS uses .dylib. OS is exported by the root
+# Makefile (`export OS := $(shell uname -s)`).
+SHIM_EXT=so
+
 OPUS_CFLAGS := $(shell pkg-config --cflags opus)
 OPUS_LIBS := $(shell pkg-config --libs opus)
+SHIM_LDFLAGS := $(OPUS_LIBS)
 
-libopusshim.so: csrc/opus_shim.c
-	$(CC) -shared -fPIC -o $@ $< $(OPUS_CFLAGS) $(OPUS_LIBS)
+ifeq ($(OS),Darwin)
+	SHIM_EXT=dylib
+	# Resolve libopus symbols lazily from the already globally-loaded
+	# libopus (codec.go dlopens it RTLD_GLOBAL before the shim) rather than
+	# recording an absolute Homebrew path in the dylib. This keeps the
+	# packaged shim relocatable on machines that have no Homebrew.
+	SHIM_LDFLAGS := -undefined dynamic_lookup
+endif
 
-opus: libopusshim.so
+libopusshim.$(SHIM_EXT): csrc/opus_shim.c
+	$(CC) -shared -fPIC -o $@ $< $(OPUS_CFLAGS) $(SHIM_LDFLAGS)
+
+opus: libopusshim.$(SHIM_EXT)
 	$(GOCMD) build -tags "$(GO_TAGS)" -o opus ./
 
 package: opus
@@ -16,4 +33,7 @@ package: opus
 build: package
 
 clean:
-	rm -f opus libopusshim.so
+	rm -f opus libopusshim.$(SHIM_EXT)
+	rm -rf package
+
+.PHONY: build package clean

backend/go/opus/package.sh

@@ -8,13 +8,23 @@ mkdir -p $CURDIR/package/lib
 cp -avf $CURDIR/opus $CURDIR/package/
 cp -avf $CURDIR/run.sh $CURDIR/package/
 
-# Copy the opus shim library
-cp -avf $CURDIR/libopusshim.so $CURDIR/package/lib/
+# The shim extension is OS-specific (.so on Linux, .dylib on macOS).
+SHIM_EXT=so
+if [ "$(uname)" = "Darwin" ]; then
+    SHIM_EXT=dylib
+fi
 
-# Copy system libopus
+# Copy the opus shim library
+cp -avf $CURDIR/libopusshim.$SHIM_EXT $CURDIR/package/lib/
+
+# Copy system libopus so the backend is self-contained: the runtime base
+# image has neither libopus-dev (Linux) nor Homebrew (macOS), so codec.go's
+# dlopen would otherwise fail. Both name patterns are attempted; only the
+# host's matching one exists.
 if command -v pkg-config >/dev/null 2>&1 && pkg-config --exists opus; then
     LIBOPUS_DIR=$(pkg-config --variable=libdir opus)
-    cp -avfL $LIBOPUS_DIR/libopus.so* $CURDIR/package/lib/ 2>/dev/null || true
+    cp -avf $LIBOPUS_DIR/libopus.so* $CURDIR/package/lib/ 2>/dev/null || true
+    cp -avf $LIBOPUS_DIR/libopus*.dylib $CURDIR/package/lib/ 2>/dev/null || true
 fi
 
 # Detect architecture and copy appropriate libraries
@@ -38,6 +48,8 @@ elif [ -f "/lib/ld-linux-aarch64.so.1" ]; then
     cp -arfLv /lib/aarch64-linux-gnu/libdl.so.2 $CURDIR/package/lib/libdl.so.2
     cp -arfLv /lib/aarch64-linux-gnu/librt.so.1 $CURDIR/package/lib/librt.so.1
     cp -arfLv /lib/aarch64-linux-gnu/libpthread.so.0 $CURDIR/package/lib/libpthread.so.0
+elif [ "$(uname -s)" = "Darwin" ]; then
+    echo "Detected Darwin — system libraries linked dynamically, no bundled loader needed"
 else
     echo "Warning: Could not detect architecture for system library bundling"
 fi

backend/go/opus/run.sh

@@ -1,15 +1,20 @@
 #!/bin/bash
 set -ex
 
-CURDIR=$(dirname "$(realpath $0)")
+CURDIR=$(dirname "$(realpath "$0")")
 
-export LD_LIBRARY_PATH=$CURDIR/lib:$LD_LIBRARY_PATH
-export OPUS_SHIM_LIBRARY=$CURDIR/lib/libopusshim.so
-
-# If there is a lib/ld.so, use it
-if [ -f $CURDIR/lib/ld.so ]; then
-	echo "Using lib/ld.so"
-	exec $CURDIR/lib/ld.so $CURDIR/opus "$@"
+if [ "$(uname)" = "Darwin" ]; then
+	export DYLD_LIBRARY_PATH="$CURDIR"/lib:$DYLD_LIBRARY_PATH
+	export OPUS_SHIM_LIBRARY="$CURDIR"/lib/libopusshim.dylib
+else
+	export LD_LIBRARY_PATH="$CURDIR"/lib:$LD_LIBRARY_PATH
+	export OPUS_SHIM_LIBRARY="$CURDIR"/lib/libopusshim.so
 fi
 
-exec $CURDIR/opus "$@"
+# If there is a lib/ld.so, use it
+if [ -f "$CURDIR"/lib/ld.so ]; then
+	echo "Using lib/ld.so"
+	exec "$CURDIR"/lib/ld.so "$CURDIR"/opus "$@"
+fi
+
+exec "$CURDIR"/opus "$@"

backend/go/parakeet-cpp/Makefile

@@ -1,6 +1,6 @@
 # parakeet-cpp backend Makefile.
 #
-# Upstream pin lives below as PARAKEET_VERSION?=db755a78d39f789bb7d4e3935158a9e8105dbe36
+# Upstream pin lives below as PARAKEET_VERSION?=f469a57270a1cc4554acb15febf60e56619673b9
 # (.github/bump_deps.sh) can find and update it - matches the
 # whisper.cpp / ds4 / vibevoice-cpp convention.
 #
@@ -15,7 +15,7 @@
 # That's what the L0 smoke test uses. The default target below does the
 # proper clone-at-pin + cmake build so CI doesn't need a side-checkout.
 
-PARAKEET_VERSION?=db755a78d39f789bb7d4e3935158a9e8105dbe36
+PARAKEET_VERSION?=f469a57270a1cc4554acb15febf60e56619673b9
 PARAKEET_REPO?=https://github.com/mudler/parakeet.cpp
 
 GOCMD?=go
@@ -74,6 +74,7 @@ libparakeet.so: sources/parakeet.cpp
 	cmake -B sources/parakeet.cpp/build-shared -S sources/parakeet.cpp $(CMAKE_ARGS)
 	cmake --build sources/parakeet.cpp/build-shared --config Release -j$(JOBS)
 	cp -fv sources/parakeet.cpp/build-shared/libparakeet.so* ./ 2>/dev/null || true
+	cp -fv sources/parakeet.cpp/build-shared/libparakeet.dylib ./ 2>/dev/null || true
 	cp -fv sources/parakeet.cpp/include/parakeet_capi.h ./
 
 parakeet-cpp-grpc: libparakeet.so main.go goparakeetcpp.go

backend/go/parakeet-cpp/main.go

@@ -2,15 +2,17 @@ package main
 
 // Started internally by LocalAI - one gRPC server per loaded model.
 //
-// Loads libparakeet.so via purego and registers the flat C-API entry
-// points declared in parakeet_capi.h. The library name can be overridden
-// with PARAKEET_LIBRARY (mirrors the WHISPER_LIBRARY / VIBEVOICECPP_LIBRARY
-// convention in the sibling backends); the default looks for the .so next
-// to this binary.
+// Loads the parakeet shared library via purego and registers the flat
+// C-API entry points declared in parakeet_capi.h. The library name can be
+// overridden with PARAKEET_LIBRARY (mirrors the WHISPER_LIBRARY /
+// VIBEVOICECPP_LIBRARY convention in the sibling backends); the default
+// looks next to this binary for libparakeet.so on Linux and
+// libparakeet.dylib on macOS.
 import (
 	"flag"
 	"fmt"
 	"os"
+	"runtime"
 
 	"github.com/ebitengine/purego"
 	grpc "github.com/mudler/LocalAI/pkg/grpc"
@@ -28,7 +30,11 @@ type LibFuncs struct {
 func main() {
 	libName := os.Getenv("PARAKEET_LIBRARY")
 	if libName == "" {
-		libName = "libparakeet.so"
+		if runtime.GOOS == "darwin" {
+			libName = "libparakeet.dylib"
+		} else {
+			libName = "libparakeet.so"
+		}
 	}
 
 	lib, err := purego.Dlopen(libName, purego.RTLD_NOW|purego.RTLD_GLOBAL)

backend/go/parakeet-cpp/package.sh

@@ -16,12 +16,15 @@ mkdir -p "$CURDIR/package/lib"
 cp -avf "$CURDIR/parakeet-cpp-grpc" "$CURDIR/package/"
 cp -avf "$CURDIR/run.sh" "$CURDIR/package/"
 
-# libparakeet.so + any soname symlinks (libparakeet.so.X[.Y]). purego.Dlopen
-# resolves it via LD_LIBRARY_PATH, which run.sh points at lib/.
-cp -avf "$CURDIR"/libparakeet.so* "$CURDIR/package/lib/" 2>/dev/null || {
-	echo "ERROR: libparakeet.so not found in $CURDIR, run 'make' first" >&2
+# libparakeet shared lib + any soname symlinks. On Linux this is
+# libparakeet.so[.X.Y]; on macOS it is libparakeet.dylib. purego.Dlopen
+# resolves it via the *_LIBRARY_PATH that run.sh points at lib/.
+cp -avf "$CURDIR"/libparakeet.so* "$CURDIR/package/lib/" 2>/dev/null || true
+cp -avf "$CURDIR"/libparakeet.dylib "$CURDIR/package/lib/" 2>/dev/null || true
+if ! ls "$CURDIR"/package/lib/libparakeet.* >/dev/null 2>&1; then
+	echo "ERROR: libparakeet shared library not found in $CURDIR, run 'make' first" >&2
 	exit 1
-}
+fi
 
 # Detect architecture and copy the core runtime libs libparakeet.so links
 # against, plus the matching dynamic loader as lib/ld.so.
@@ -48,7 +51,7 @@ elif [ -f "/lib/ld-linux-aarch64.so.1" ]; then
     cp -arfLv /lib/aarch64-linux-gnu/librt.so.1 "$CURDIR/package/lib/librt.so.1"
     cp -arfLv /lib/aarch64-linux-gnu/libpthread.so.0 "$CURDIR/package/lib/libpthread.so.0"
 elif [ "$(uname -s)" = "Darwin" ]; then
-    echo "Detected Darwin"
+    echo "Detected Darwin — system frameworks linked dynamically, no bundled libs needed"
 else
     echo "Error: Could not detect architecture"
     exit 1

backend/go/parakeet-cpp/run.sh

@@ -3,11 +3,17 @@ set -e
 
 CURDIR=$(dirname "$(realpath "$0")")
 
-export LD_LIBRARY_PATH="$CURDIR/lib:$CURDIR:${LD_LIBRARY_PATH:-}"
+if [ "$(uname)" = "Darwin" ]; then
+	export DYLD_LIBRARY_PATH="$CURDIR/lib:"$CURDIR":${DYLD_LIBRARY_PATH:-}"
+	export PARAKEET_LIBRARY="$CURDIR/lib/libparakeet.dylib"
+else
+	export LD_LIBRARY_PATH="$CURDIR/lib:"$CURDIR":${LD_LIBRARY_PATH:-}"
+	export PARAKEET_LIBRARY="$CURDIR/lib/libparakeet.so"
+fi
 
 # If a self-contained ld.so was packaged, route through it so the
 # packaged libc / libstdc++ are used instead of the host's (matches the
-# whisper backend's runtime layout).
+# whisper backend's runtime layout). Linux only.
 if [ -f "$CURDIR/lib/ld.so" ]; then
 	echo "Using lib/ld.so"
 	exec "$CURDIR/lib/ld.so" "$CURDIR/parakeet-cpp-grpc" "$@"

backend/go/piper/package.sh

@@ -16,7 +16,15 @@ cp -rfv $CURDIR/run.sh $CURDIR/package/
 cp -rfLv $CURDIR/sources/go-piper/piper-phonemize/pi/lib/* $CURDIR/package/lib/
 
 # Detect architecture and copy appropriate libraries
-if [ -f "/lib64/ld-linux-x86-64.so.2" ]; then
+if [ "$(uname)" = "Darwin" ]; then
+    # macOS has no glibc loader to bundle. The piper binary links its bundled
+    # libs (libucd, libespeak-ng, libpiper_phonemize, libonnxruntime) via
+    # @rpath but ships with no LC_RPATH, so dyld aborts at launch with
+    # "Library not loaded: @rpath/libucd.dylib ... no LC_RPATH's found".
+    # Add an @loader_path/lib rpath so @rpath resolves to package/lib/.
+    echo "Detected macOS; adding @loader_path/lib rpath so bundled libs resolve via @rpath..."
+    install_name_tool -add_rpath @loader_path/lib "$CURDIR/package/piper"
+elif [ -f "/lib64/ld-linux-x86-64.so.2" ]; then
     # x86_64 architecture
     echo "Detected x86_64 architecture, copying x86_64 libraries..."
     cp -arfLv /lib64/ld-linux-x86-64.so.2 $CURDIR/package/lib/ld.so

backend/go/piper/run.sh

@@ -1,15 +1,20 @@
 #!/bin/bash
 set -ex
 
-CURDIR=$(dirname "$(realpath $0)")
+CURDIR=$(dirname "$(realpath "$0")")
 
-export ESPEAK_NG_DATA=$CURDIR/espeak-ng-data
-export LD_LIBRARY_PATH=$CURDIR/lib:$LD_LIBRARY_PATH
+export ESPEAK_NG_DATA="$CURDIR"/espeak-ng-data
 
-# If there is a lib/ld.so, use it
-if [ -f $CURDIR/lib/ld.so ]; then
-	echo "Using lib/ld.so"
-	exec $CURDIR/lib/ld.so $CURDIR/piper "$@"
+if [ "$(uname)" = "Darwin" ]; then
+	export DYLD_LIBRARY_PATH="$CURDIR"/lib:$DYLD_LIBRARY_PATH
+else
+	export LD_LIBRARY_PATH="$CURDIR"/lib:$LD_LIBRARY_PATH
 fi
 
-exec $CURDIR/piper "$@"
+# If there is a lib/ld.so, use it
+if [ -f "$CURDIR"/lib/ld.so ]; then
+	echo "Using lib/ld.so"
+	exec "$CURDIR"/lib/ld.so "$CURDIR"/piper "$@"
+fi
+
+exec "$CURDIR"/piper "$@"

backend/go/qwen3-tts-cpp/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # qwentts.cpp version
 QWEN3TTS_REPO?=https://github.com/ServeurpersoCom/qwentts.cpp
-QWEN3TTS_CPP_VERSION?=4536dcdce27c3764a93a06d6bf64026b124962f5
+QWEN3TTS_CPP_VERSION?=9dbe7ea26a01b30fccb117ae5e86807c1dc23d42
 SO_TARGET?=libgoqwen3ttscpp.so
 
 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF
@@ -65,8 +65,8 @@ UNAME_S := $(shell uname -s)
 ifeq ($(UNAME_S),Linux)
 	VARIANT_TARGETS = libgoqwen3ttscpp-avx.so libgoqwen3ttscpp-avx2.so libgoqwen3ttscpp-avx512.so libgoqwen3ttscpp-fallback.so
 else
-	# On non-Linux (e.g., Darwin), build only fallback variant
-	VARIANT_TARGETS = libgoqwen3ttscpp-fallback.so
+	# On non-Linux (e.g., Darwin), build only fallback variant (as a dylib)
+	VARIANT_TARGETS = libgoqwen3ttscpp-fallback.dylib
 endif
 
 qwen3-tts-cpp: main.go goqwen3ttscpp.go $(VARIANT_TARGETS)
@@ -78,7 +78,7 @@ package: qwen3-tts-cpp
 build: package
 
 clean: purge
-	rm -rf libgoqwen3ttscpp*.so package sources/qwentts.cpp qwen3-tts-cpp
+	rm -rf libgoqwen3ttscpp*.so libgoqwen3ttscpp*.dylib package sources/qwentts.cpp qwen3-tts-cpp
 
 purge:
 	rm -rf build*
@@ -110,13 +110,20 @@ libgoqwen3ttscpp-fallback.so: sources/qwentts.cpp
 	SO_TARGET=libgoqwen3ttscpp-fallback.so CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off" $(MAKE) libgoqwen3ttscpp-custom
 	rm -rf build-libgoqwen3ttscpp-fallback.so
 
+# Build fallback variant as a dylib (Darwin)
+libgoqwen3ttscpp-fallback.dylib: sources/qwentts.cpp
+	$(info ${GREEN}I qwen3-tts-cpp build info:fallback (dylib)${RESET})
+	SO_TARGET=libgoqwen3ttscpp-fallback.dylib CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off" $(MAKE) libgoqwen3ttscpp-custom
+	rm -rf build-libgoqwen3ttscpp-fallback.dylib
+
 libgoqwen3ttscpp-custom: CMakeLists.txt cpp/goqwen3ttscpp.cpp cpp/goqwen3ttscpp.h
 	mkdir -p build-$(SO_TARGET) && \
 	cd build-$(SO_TARGET) && \
 	cmake .. $(CMAKE_ARGS) && \
 	cmake --build . --config Release -j$(JOBS) --target goqwen3ttscpp && \
 	cd .. && \
-	mv build-$(SO_TARGET)/libgoqwen3ttscpp.so ./$(SO_TARGET)
+	(mv build-$(SO_TARGET)/libgoqwen3ttscpp.so ./$(SO_TARGET) 2>/dev/null || \
+	 mv build-$(SO_TARGET)/libgoqwen3ttscpp.dylib ./$(SO_TARGET) 2>/dev/null)
 
 test: qwen3-tts-cpp
 	@echo "Running qwen3-tts-cpp tests..."

backend/go/qwen3-tts-cpp/main.go

@@ -4,6 +4,7 @@ package main
 import (
 	"flag"
 	"os"
+	"runtime"
 
 	"github.com/ebitengine/purego"
 	grpc "github.com/mudler/LocalAI/pkg/grpc"
@@ -21,7 +22,11 @@ type LibFuncs struct {
 func main() {
 	libName := os.Getenv("QWEN3TTS_LIBRARY")
 	if libName == "" {
-		libName = "./libgoqwen3ttscpp-fallback.so"
+		if runtime.GOOS == "darwin" {
+			libName = "./libgoqwen3ttscpp-fallback.dylib"
+		} else {
+			libName = "./libgoqwen3ttscpp-fallback.so"
+		}
 	}
 
 	lib, err := purego.Dlopen(libName, purego.RTLD_NOW|purego.RTLD_GLOBAL)

backend/go/qwen3-tts-cpp/package.sh

@@ -12,7 +12,8 @@ REPO_ROOT="${CURDIR}/../../.."
 mkdir -p $CURDIR/package/lib
 
 cp -avf $CURDIR/qwen3-tts-cpp $CURDIR/package/
-cp -fv $CURDIR/libgoqwen3ttscpp-*.so $CURDIR/package/
+cp -fv $CURDIR/libgoqwen3ttscpp-*.so $CURDIR/package/ 2>/dev/null || true
+cp -fv $CURDIR/libgoqwen3ttscpp-*.dylib $CURDIR/package/ 2>/dev/null || true
 cp -fv $CURDIR/run.sh $CURDIR/package/
 
 # Detect architecture and copy appropriate libraries

backend/go/qwen3-tts-cpp/run.sh

@@ -2,7 +2,7 @@
 set -ex
 
 # Get the absolute current dir where the script is located
-CURDIR=$(dirname "$(realpath $0)")
+CURDIR=$(dirname "$(realpath "$0")")
 
 cd /
 
@@ -12,19 +12,23 @@ if [ "$(uname)" != "Darwin" ]; then
 	grep -e "flags" /proc/cpuinfo | head -1
 fi
 
-LIBRARY="$CURDIR/libgoqwen3ttscpp-fallback.so"
+if [ "$(uname)" = "Darwin" ]; then
+	# macOS: single dylib variant (Metal or Accelerate)
+	LIBRARY="$CURDIR/libgoqwen3ttscpp-fallback.dylib"
+	export DYLD_LIBRARY_PATH="$CURDIR"/lib:$DYLD_LIBRARY_PATH
+else
+	LIBRARY="$CURDIR/libgoqwen3ttscpp-fallback.so"
 
-if [ "$(uname)" != "Darwin" ]; then
 	if grep -q -e "\savx\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX    found OK"
-		if [ -e $CURDIR/libgoqwen3ttscpp-avx.so ]; then
+		if [ -e "$CURDIR"/libgoqwen3ttscpp-avx.so ]; then
 			LIBRARY="$CURDIR/libgoqwen3ttscpp-avx.so"
 		fi
 	fi
 
 	if grep -q -e "\savx2\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX2   found OK"
-		if [ -e $CURDIR/libgoqwen3ttscpp-avx2.so ]; then
+		if [ -e "$CURDIR"/libgoqwen3ttscpp-avx2.so ]; then
 			LIBRARY="$CURDIR/libgoqwen3ttscpp-avx2.so"
 		fi
 	fi
@@ -32,21 +36,22 @@ if [ "$(uname)" != "Darwin" ]; then
 	# Check avx 512
 	if grep -q -e "\savx512f\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX512F found OK"
-		if [ -e $CURDIR/libgoqwen3ttscpp-avx512.so ]; then
+		if [ -e "$CURDIR"/libgoqwen3ttscpp-avx512.so ]; then
 			LIBRARY="$CURDIR/libgoqwen3ttscpp-avx512.so"
 		fi
 	fi
+
+	export LD_LIBRARY_PATH="$CURDIR"/lib:$LD_LIBRARY_PATH
 fi
 
-export LD_LIBRARY_PATH=$CURDIR/lib:$LD_LIBRARY_PATH
 export QWEN3TTS_LIBRARY=$LIBRARY
 
 # If there is a lib/ld.so, use it
-if [ -f $CURDIR/lib/ld.so ]; then
+if [ -f "$CURDIR"/lib/ld.so ]; then
 	echo "Using lib/ld.so"
 	echo "Using library: $LIBRARY"
-	exec $CURDIR/lib/ld.so $CURDIR/qwen3-tts-cpp "$@"
+	exec "$CURDIR"/lib/ld.so "$CURDIR"/qwen3-tts-cpp "$@"
 fi
 
 echo "Using library: $LIBRARY"
-exec $CURDIR/qwen3-tts-cpp "$@"
+exec "$CURDIR"/qwen3-tts-cpp "$@"

backend/go/rfdetr-cpp/Makefile

@@ -34,6 +34,8 @@ else ifeq ($(BUILD_TYPE),hipblas)
 else ifeq ($(BUILD_TYPE),vulkan)
 	CMAKE_ARGS+=-DGGML_VULKAN=ON -DRFDETR_GGML_VULKAN=ON
 else ifeq ($(OS),Darwin)
+	# macOS/Metal: built + published as an OCI image by CI (includeDarwin in
+	# .github/backend-matrix.yml) so Apple Silicon users can install this backend.
 	ifneq ($(BUILD_TYPE),metal)
 		CMAKE_ARGS+=-DGGML_METAL=OFF
 	else
@@ -71,7 +73,7 @@ ifeq ($(UNAME_S),Linux)
 	VARIANT_TARGETS = librfdetrcpp-avx.so librfdetrcpp-avx2.so librfdetrcpp-avx512.so librfdetrcpp-fallback.so
 else
 	# On non-Linux (e.g., Darwin), build only fallback variant
-	VARIANT_TARGETS = librfdetrcpp-fallback.so
+	VARIANT_TARGETS = librfdetrcpp-fallback.dylib
 endif
 
 rfdetr-cpp: main.go gorfdetrcpp.go $(VARIANT_TARGETS)
@@ -83,7 +85,7 @@ package: rfdetr-cpp
 build: package
 
 clean: purge
-	rm -rf librfdetrcpp*.so rfdetr-cpp package sources
+	rm -rf librfdetrcpp*.so librfdetrcpp*.dylib rfdetr-cpp package sources
 
 purge:
 	rm -rf build*
@@ -110,11 +112,19 @@ librfdetrcpp-avx512.so: sources/rt-detr.cpp
 endif
 
 # Build fallback variant (all platforms)
+ifeq ($(UNAME_S),Darwin)
+librfdetrcpp-fallback.dylib: sources/rt-detr.cpp
+	rm -rfv build-$@
+	$(info ${GREEN}I rfdetr-cpp build info:fallback${RESET})
+	SO_TARGET=$@ CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off" $(MAKE) librfdetrcpp-custom
+	rm -rfv build-$@
+else
 librfdetrcpp-fallback.so: sources/rt-detr.cpp
 	rm -rfv build-$@
 	$(info ${GREEN}I rfdetr-cpp build info:fallback${RESET})
 	SO_TARGET=$@ CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off" $(MAKE) librfdetrcpp-custom
 	rm -rfv build-$@
+endif
 
 librfdetrcpp-custom: CMakeLists.txt
 	mkdir -p build-$(SO_TARGET) && \
@@ -122,7 +132,8 @@ librfdetrcpp-custom: CMakeLists.txt
 	cmake .. $(CMAKE_ARGS) && \
 	cmake --build . --config Release -j$(JOBS) && \
 	cd .. && \
-	mv build-$(SO_TARGET)/librfdetrcpp.so ./$(SO_TARGET)
+	(mv build-$(SO_TARGET)/librfdetrcpp.so ./$(SO_TARGET) 2>/dev/null || \
+	 mv build-$(SO_TARGET)/librfdetrcpp.dylib ./$(SO_TARGET) 2>/dev/null)
 
 all: rfdetr-cpp package
 

backend/go/rfdetr-cpp/main.go

@@ -9,6 +9,7 @@ package main
 import (
 	"flag"
 	"os"
+	"runtime"
 
 	"github.com/ebitengine/purego"
 	grpc "github.com/mudler/LocalAI/pkg/grpc"
@@ -27,7 +28,11 @@ func main() {
 	// Get library name from environment variable, default to fallback
 	libName := os.Getenv("RFDETR_LIBRARY")
 	if libName == "" {
-		libName = "./librfdetrcpp-fallback.so"
+		if runtime.GOOS == "darwin" {
+			libName = "./librfdetrcpp-fallback.dylib"
+		} else {
+			libName = "./librfdetrcpp-fallback.so"
+		}
 	}
 
 	rfdetrLib, err := purego.Dlopen(libName, purego.RTLD_NOW|purego.RTLD_GLOBAL)

backend/go/rfdetr-cpp/package.sh

@@ -10,7 +10,8 @@ REPO_ROOT="${CURDIR}/../../.."
 # Create lib directory
 mkdir -p $CURDIR/package/lib
 
-cp -avf $CURDIR/librfdetrcpp-*.so $CURDIR/package/
+cp -fv $CURDIR/librfdetrcpp-*.so $CURDIR/package/ 2>/dev/null || true
+cp -fv $CURDIR/librfdetrcpp-*.dylib $CURDIR/package/ 2>/dev/null || true
 cp -avf $CURDIR/rfdetr-cpp $CURDIR/package/
 cp -fv $CURDIR/run.sh $CURDIR/package/
 

backend/go/rfdetr-cpp/run.sh

@@ -2,7 +2,7 @@
 set -ex
 
 # Get the absolute current dir where the script is located
-CURDIR=$(dirname "$(realpath $0)")
+CURDIR=$(dirname "$(realpath "$0")")
 
 cd /
 
@@ -12,19 +12,23 @@ if [ "$(uname)" != "Darwin" ]; then
 	grep -e "flags" /proc/cpuinfo | head -1
 fi
 
-LIBRARY="$CURDIR/librfdetrcpp-fallback.so"
+if [ "$(uname)" = "Darwin" ]; then
+	# macOS: single dylib variant (Metal or Accelerate)
+	LIBRARY="$CURDIR/librfdetrcpp-fallback.dylib"
+	export DYLD_LIBRARY_PATH="$CURDIR"/lib:$DYLD_LIBRARY_PATH
+else
+	LIBRARY="$CURDIR/librfdetrcpp-fallback.so"
 
-if [ "$(uname)" != "Darwin" ]; then
 	if grep -q -e "\savx\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX    found OK"
-		if [ -e $CURDIR/librfdetrcpp-avx.so ]; then
+		if [ -e "$CURDIR"/librfdetrcpp-avx.so ]; then
 			LIBRARY="$CURDIR/librfdetrcpp-avx.so"
 		fi
 	fi
 
 	if grep -q -e "\savx2\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX2   found OK"
-		if [ -e $CURDIR/librfdetrcpp-avx2.so ]; then
+		if [ -e "$CURDIR"/librfdetrcpp-avx2.so ]; then
 			LIBRARY="$CURDIR/librfdetrcpp-avx2.so"
 		fi
 	fi
@@ -32,21 +36,22 @@ if [ "$(uname)" != "Darwin" ]; then
 	# Check avx 512
 	if grep -q -e "\savx512f\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX512F found OK"
-		if [ -e $CURDIR/librfdetrcpp-avx512.so ]; then
+		if [ -e "$CURDIR"/librfdetrcpp-avx512.so ]; then
 			LIBRARY="$CURDIR/librfdetrcpp-avx512.so"
 		fi
 	fi
+
+	export LD_LIBRARY_PATH="$CURDIR"/lib:$LD_LIBRARY_PATH
 fi
 
-export LD_LIBRARY_PATH=$CURDIR/lib:$LD_LIBRARY_PATH
 export RFDETR_LIBRARY=$LIBRARY
 
 # If there is a lib/ld.so, use it
-if [ -f $CURDIR/lib/ld.so ]; then
+if [ -f "$CURDIR"/lib/ld.so ]; then
 	echo "Using lib/ld.so"
 	echo "Using library: $LIBRARY"
-	exec $CURDIR/lib/ld.so $CURDIR/rfdetr-cpp "$@"
+	exec "$CURDIR"/lib/ld.so "$CURDIR"/rfdetr-cpp "$@"
 fi
 
 echo "Using library: $LIBRARY"
-exec $CURDIR/rfdetr-cpp "$@"
+exec "$CURDIR"/rfdetr-cpp "$@"

backend/go/sam3-cpp/Makefile

@@ -31,6 +31,8 @@ else ifeq ($(BUILD_TYPE),hipblas)
 else ifeq ($(BUILD_TYPE),vulkan)
 	CMAKE_ARGS+=-DGGML_VULKAN=ON
 else ifeq ($(OS),Darwin)
+	# macOS/Metal: built + published as an OCI image by CI (includeDarwin in
+	# .github/backend-matrix.yml) so Apple Silicon users can install this backend.
 	ifneq ($(BUILD_TYPE),metal)
 		CMAKE_ARGS+=-DGGML_METAL=OFF
 	else
@@ -66,7 +68,7 @@ ifeq ($(UNAME_S),Linux)
 	VARIANT_TARGETS = libgosam3-avx.so libgosam3-avx2.so libgosam3-avx512.so libgosam3-fallback.so
 else
 	# On non-Linux (e.g., Darwin), build only fallback variant
-	VARIANT_TARGETS = libgosam3-fallback.so
+	VARIANT_TARGETS = libgosam3-fallback.dylib
 endif
 
 sam3-cpp: main.go gosam3.go $(VARIANT_TARGETS)
@@ -78,7 +80,7 @@ package: sam3-cpp
 build: package
 
 clean: purge
-	rm -rf libgosam3*.so sam3-cpp package sources
+	rm -rf libgosam3*.so libgosam3*.dylib sam3-cpp package sources
 
 purge:
 	rm -rf build*
@@ -105,11 +107,19 @@ libgosam3-avx512.so: sources/sam3.cpp
 endif
 
 # Build fallback variant (all platforms)
+ifeq ($(UNAME_S),Darwin)
+libgosam3-fallback.dylib: sources/sam3.cpp
+	$(MAKE) purge
+	$(info ${GREEN}I sam3-cpp build info:fallback${RESET})
+	SO_TARGET=libgosam3-fallback.dylib CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off" $(MAKE) libgosam3-custom
+	rm -rfv build*
+else
 libgosam3-fallback.so: sources/sam3.cpp
 	$(MAKE) purge
 	$(info ${GREEN}I sam3-cpp build info:fallback${RESET})
 	SO_TARGET=libgosam3-fallback.so CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off" $(MAKE) libgosam3-custom
 	rm -rfv build*
+endif
 
 libgosam3-custom: CMakeLists.txt cpp/gosam3.cpp cpp/gosam3.h
 	mkdir -p build-$(SO_TARGET) && \
@@ -117,6 +127,7 @@ libgosam3-custom: CMakeLists.txt cpp/gosam3.cpp cpp/gosam3.h
 	cmake .. $(CMAKE_ARGS) && \
 	cmake --build . --config Release -j$(JOBS) && \
 	cd .. && \
-	mv build-$(SO_TARGET)/libgosam3.so ./$(SO_TARGET)
+	(mv build-$(SO_TARGET)/libgosam3.so ./$(SO_TARGET) 2>/dev/null || \
+	 mv build-$(SO_TARGET)/libgosam3.dylib ./$(SO_TARGET) 2>/dev/null)
 
 all: sam3-cpp package

backend/go/sam3-cpp/main.go

@@ -3,6 +3,7 @@ package main
 import (
 	"flag"
 	"os"
+	"runtime"
 
 	"github.com/ebitengine/purego"
 	grpc "github.com/mudler/LocalAI/pkg/grpc"
@@ -21,7 +22,11 @@ func main() {
 	// Get library name from environment variable, default to fallback
 	libName := os.Getenv("SAM3_LIBRARY")
 	if libName == "" {
-		libName = "./libgosam3-fallback.so"
+		if runtime.GOOS == "darwin" {
+			libName = "./libgosam3-fallback.dylib"
+		} else {
+			libName = "./libgosam3-fallback.so"
+		}
 	}
 
 	gosamLib, err := purego.Dlopen(libName, purego.RTLD_NOW|purego.RTLD_GLOBAL)

backend/go/sam3-cpp/package.sh

@@ -10,7 +10,8 @@ REPO_ROOT="${CURDIR}/../../.."
 # Create lib directory
 mkdir -p $CURDIR/package/lib
 
-cp -avf $CURDIR/libgosam3-*.so $CURDIR/package/
+cp -fv $CURDIR/libgosam3-*.so $CURDIR/package/ 2>/dev/null || true
+cp -fv $CURDIR/libgosam3-*.dylib $CURDIR/package/ 2>/dev/null || true
 cp -avf $CURDIR/sam3-cpp $CURDIR/package/
 cp -fv $CURDIR/run.sh $CURDIR/package/
 

backend/go/sam3-cpp/run.sh

@@ -2,7 +2,7 @@
 set -ex
 
 # Get the absolute current dir where the script is located
-CURDIR=$(dirname "$(realpath $0)")
+CURDIR=$(dirname "$(realpath "$0")")
 
 cd /
 
@@ -12,19 +12,23 @@ if [ "$(uname)" != "Darwin" ]; then
 	grep -e "flags" /proc/cpuinfo | head -1
 fi
 
-LIBRARY="$CURDIR/libgosam3-fallback.so"
+if [ "$(uname)" = "Darwin" ]; then
+	# macOS: single dylib variant (Metal or Accelerate)
+	LIBRARY="$CURDIR/libgosam3-fallback.dylib"
+	export DYLD_LIBRARY_PATH="$CURDIR"/lib:$DYLD_LIBRARY_PATH
+else
+	LIBRARY="$CURDIR/libgosam3-fallback.so"
 
-if [ "$(uname)" != "Darwin" ]; then
 	if grep -q -e "\savx\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX    found OK"
-		if [ -e $CURDIR/libgosam3-avx.so ]; then
+		if [ -e "$CURDIR"/libgosam3-avx.so ]; then
 			LIBRARY="$CURDIR/libgosam3-avx.so"
 		fi
 	fi
 
 	if grep -q -e "\savx2\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX2   found OK"
-		if [ -e $CURDIR/libgosam3-avx2.so ]; then
+		if [ -e "$CURDIR"/libgosam3-avx2.so ]; then
 			LIBRARY="$CURDIR/libgosam3-avx2.so"
 		fi
 	fi
@@ -32,21 +36,22 @@ if [ "$(uname)" != "Darwin" ]; then
 	# Check avx 512
 	if grep -q -e "\savx512f\s" /proc/cpuinfo ; then
 		echo "CPU:    AVX512F found OK"
-		if [ -e $CURDIR/libgosam3-avx512.so ]; then
+		if [ -e "$CURDIR"/libgosam3-avx512.so ]; then
 			LIBRARY="$CURDIR/libgosam3-avx512.so"
 		fi
 	fi
+
+	export LD_LIBRARY_PATH="$CURDIR"/lib:$LD_LIBRARY_PATH
 fi
 
-export LD_LIBRARY_PATH=$CURDIR/lib:$LD_LIBRARY_PATH
 export SAM3_LIBRARY=$LIBRARY
 
 # If there is a lib/ld.so, use it
-if [ -f $CURDIR/lib/ld.so ]; then
+if [ -f "$CURDIR"/lib/ld.so ]; then
 	echo "Using lib/ld.so"
 	echo "Using library: $LIBRARY"
-	exec $CURDIR/lib/ld.so $CURDIR/sam3-cpp "$@"
+	exec "$CURDIR"/lib/ld.so "$CURDIR"/sam3-cpp "$@"
 fi
 
 echo "Using library: $LIBRARY"
-exec $CURDIR/sam3-cpp "$@"
+exec "$CURDIR"/sam3-cpp "$@"

backend/go/sherpa-onnx/backend.go

@@ -7,6 +7,7 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
+	"runtime"
 	"strconv"
 	"strings"
 	"sync"
@@ -238,11 +239,19 @@ func loadSherpaLibs() error {
 func loadSherpaLibsOnce() error {
 	shimLib := os.Getenv("SHERPA_SHIM_LIBRARY")
 	if shimLib == "" {
-		shimLib = "libsherpa-shim.so"
+		if runtime.GOOS == "darwin" {
+			shimLib = "libsherpa-shim.dylib"
+		} else {
+			shimLib = "libsherpa-shim.so"
+		}
 	}
 	capiLib := os.Getenv("SHERPA_ONNX_LIBRARY")
 	if capiLib == "" {
-		capiLib = "libsherpa-onnx-c-api.so"
+		if runtime.GOOS == "darwin" {
+			capiLib = "libsherpa-onnx-c-api.dylib"
+		} else {
+			capiLib = "libsherpa-onnx-c-api.so"
+		}
 	}
 
 	shim, err := purego.Dlopen(shimLib, purego.RTLD_NOW|purego.RTLD_GLOBAL)

backend/go/sherpa-onnx/run.sh

@@ -1,13 +1,19 @@
 #!/bin/bash
 set -ex
 
-CURDIR=$(dirname "$(realpath $0)")
+CURDIR=$(dirname "$(realpath "$0")")
 
-export LD_LIBRARY_PATH=$CURDIR/lib:$LD_LIBRARY_PATH
-
-if [ -f $CURDIR/lib/ld.so ]; then
-	echo "Using lib/ld.so"
-	exec $CURDIR/lib/ld.so $CURDIR/sherpa-onnx "$@"
+if [ "$(uname)" = "Darwin" ]; then
+	export DYLD_LIBRARY_PATH="$CURDIR"/lib:$DYLD_LIBRARY_PATH
+	export SHERPA_SHIM_LIBRARY="$CURDIR"/lib/libsherpa-shim.dylib
+	export SHERPA_ONNX_LIBRARY="$CURDIR"/lib/libsherpa-onnx-c-api.dylib
+else
+	export LD_LIBRARY_PATH="$CURDIR"/lib:$LD_LIBRARY_PATH
 fi
 
-exec $CURDIR/sherpa-onnx "$@"
+if [ -f "$CURDIR"/lib/ld.so ]; then
+	echo "Using lib/ld.so"
+	exec "$CURDIR"/lib/ld.so "$CURDIR"/sherpa-onnx "$@"
+fi
+
+exec "$CURDIR"/sherpa-onnx "$@"

backend/go/silero-vad/package.sh

@@ -15,7 +15,14 @@ cp -avf $CURDIR/run.sh $CURDIR/package/
 cp -rfLv $CURDIR/backend-assets/lib/* $CURDIR/package/lib/
 
 # Detect architecture and copy appropriate libraries
-if [ -f "/lib64/ld-linux-x86-64.so.2" ]; then
+if [ "$(uname)" = "Darwin" ]; then
+    # macOS has no glibc loader to bundle. silero-vad links its bundled
+    # libonnxruntime via @rpath but ships with no LC_RPATH, so dyld can't find
+    # it at runtime. Add an @loader_path/lib rpath so @rpath resolves to
+    # package/lib/ (matching the piper darwin fix, #10525).
+    echo "Detected macOS; adding @loader_path/lib rpath so bundled libs resolve via @rpath..."
+    install_name_tool -add_rpath @loader_path/lib "$CURDIR/package/silero-vad"
+elif [ -f "/lib64/ld-linux-x86-64.so.2" ]; then
     # x86_64 architecture
     echo "Detected x86_64 architecture, copying x86_64 libraries..."
     cp -arfLv /lib64/ld-linux-x86-64.so.2 $CURDIR/package/lib/ld.so