feat(ui): canvas fullscreen toggle + keyboard tab navigation

The canvas header gains a fullscreen toggle (expands the panel to cover the
viewport; resize handle hidden while fullscreen). The artifact tab strip is now
a proper ARIA tablist with roving tabindex and Left/Right arrow-key navigation.

Assisted-by: Claude:claude-opus-4-8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

.agents/dllm-backend.md

@@ -1,138 +0,0 @@
-# Working on the dllm Backend
-
-`mudler/dllm.cpp` is a standalone C++/ggml engine for DiffusionGemma
-block-diffusion models. LocalAI wraps it with a **pure-Go** backend at
-`backend/go/dllm/` that dlopens `libdllm.so` via purego (ebitengine/purego) -
-NOT cgo, and NOT a C++ grpc-server fork. The Go side owns chat templating
-(gemma4 renderer) and output parsing (gemma4 streaming parser) and implements
-the rich gRPC interface (`PredictRich`/`PredictStreamRich`, ChatDelta replies).
-
-> NOTE: github.com/mudler/dllm.cpp is still **private** (publishing is
-> planned). Until then the Makefile's anonymous clone fails; use the local-dev
-> symlink shortcut documented at the top of `backend/go/dllm/Makefile`
-> (symlink an out-of-tree `build/libdllm.so` into the backend dir and skip the
-> clone), or a git credential helper with repo access.
-
-## Pin
-
-`backend/go/dllm/Makefile` pins `DLLM_VERSION?=<sha>` at the top
-(whisper / parakeet-cpp / ds4 convention). The bump-deps bot
-(`.github/workflows/bump_deps.yaml`) tracks `mudler/dllm.cpp` `main` and
-rewrites that variable. After a manual bump: `make -C backend/go/dllm purge &&
-make -C backend/go/dllm` (the clone is keyed on the directory existing, not
-the sha).
-
-## C-ABI and the serialization contract
-
-The binding covers the 9-symbol flat C-ABI from dllm.cpp's
-`include/dllm_capi.h` (ABI v1; `main.go` hard-fails on a version mismatch):
-`abi_version, load, free, last_error, free_string, tokenize_json, generate,
-generate_stream, cancel`. Contract points the Go wiring encodes (`capi.go`
-header comment has the full list):
-
-- **One ctx = one concurrent generate/tokenize.** A per-model worker
-  goroutine (`Dllm.jobs` in `dllm.go`) owns ALL C calls, making the
-  serialization structural instead of lock discipline.
-- **`dllm_capi_cancel` is the ONE exception**: it only flips an atomic and may
-  be called from any goroutine mid-generate, so `Dllm.Cancel` bypasses the
-  worker queue. The flag resets at the start of each generate, so a watchdog
-  racing a new generate must re-issue cancel.
-- **`last_error` is a borrowed pointer** and must only be read AFTER the
-  failing call returned (never while a generate is in flight on the same ctx).
-- **Free vs in-flight requests**: requests hold `genMu.RLock` for their full
-  duration; `Free` takes the write lock, so it only runs when nothing is in
-  flight, then drains and closes the worker. Post-Free requests get a clean
-  "model not loaded" error.
-- `tokenize_json`/`generate` return malloc'd `char*` (bound as `uintptr`,
-  copied, then `dllm_capi_free_string`d); opts/params JSON must be a FLAT
-  object of scalars (`buildOptsJSON` rejects anything else).
-
-## Wire shape
-
-| RPC | Implementation |
-|---|---|
-| LoadModel | `dllm_capi_load` (params: `n_gpu_layers`, `n_threads`, `ctx_len`); `Options[]` parsed into per-request gen opts (`eb_*`, `blocks`, `kv_cache`) by `parseModelGenOpts` |
-| PredictRich | render (if templated) → `dllm_capi_generate` → parse → ONE Reply with aggregated ChatDeltas + legacy `Message` bytes |
-| PredictStreamRich | `dllm_capi_generate_stream`; per committed diffusion block → UTF-8 holdback → parser.Feed → one Reply per non-empty delta batch (channel closed by the CALLER, per `pkg/grpc/interface.go`) |
-| Predict / PredictStream | Legacy paths, delegate to the rich pair (legacy stream INVERTS channel ownership: the impl closes) |
-| TokenizeString | `dllm_capi_tokenize_json` (C side prepends BOS per `vocab.add_bos`) |
-| Cancel | `dllm_capi_cancel`, exposed as the `grpc.Cancellable` capability (`pkg/grpc/interface.go`): the gRPC server arms it via `context.AfterFunc` on the Predict/PredictStream context, so client disconnects/timeouts abort the in-flight generate - llama.cpp `IsCancelled()` parity for Go backends |
-
-`n_threads` and `ctx_len` are accepted-but-ignored by the engine at the
-current pin (the context bound comes from GGUF `n_ctx_train`); they are sent
-for forward compatibility.
-
-## Renderer / parser (the templated chat path)
-
-With `use_tokenizer_template` + raw Messages, the backend owns templating and
-parsing (the ds4 precedent, but in Go):
-
-- `gemma4_renderer.go` - `RenderGemma4(msgs, toolsJSON, enableThinking,
-  addGenerationPrompt)`. The file embeds the FULL `tokenizer.chat_template`
-  jinja (17466 bytes, md5 `8c34cf93c7a7815b3fdb300a009c4c17`) extracted
-  verbatim from `diffusiongemma-26B-A4B-it-BF16.gguf` via gguf-py - e.g.
-  `python scripts/dump_gguf.py model.gguf | grep -A400 chat_template` in the
-  dllm.cpp checkout - as a numbered comment block; every Go rule cites its
-  "tpl L<n>" line. Re-verify the md5 before blaming the renderer for a
-  mismatch with a new GGUF. **BOS exception**: the template emits
-  `{{- bos_token -}}` but the renderer deliberately does NOT - dllm.cpp's
-  `run_generate` tokenizes with `prepend_bos = vocab.add_bos` (true for
-  gemma4), so a literal `<bos>` would double it.
-- `gemma4_parser.go` - streaming state machine turning raw model text
-  (fragments can split anywhere, including mid-marker) into ChatDeltas:
-  thought channels → `reasoning_content`, `<|tool_call>call:name{...}` →
-  ToolCallDelta, `<turn|>` → done. Marker grammar cross-checked against vLLM
-  PR #45163's gemma4 tool/reasoning parsers. Malformed payloads are re-emitted
-  raw as content, never dropped.
-- Thinking is **opt-in** for this family (`Metadata["enable_thinking"]`,
-  default OFF - the inverse of ds4): the template gates every thinking branch
-  on `enable_thinking`, and the no-thinking render pre-closes an empty thought
-  channel, so the parser always starts in content state.
-- **UTF-8 boundary holdback** (`splitValidUTF8` in `dllm.go`): per-block
-  detokenization can split a multi-byte character across block boundaries, and
-  grpc-go refuses to marshal invalid UTF-8 in proto3 strings. An incomplete
-  trailing sequence (at most 3 bytes) is carried into the next block; genuinely
-  undecodable bytes become U+FFFD.
-
-Without `use_tokenizer_template`, the prompt passes through verbatim and the
-output is NOT gemma4-parsed (plain content, like any non-autoparsing backend).
-
-## Tests
-
-| Layer | Gate | What |
-|---|---|---|
-| `backend/go/dllm/*_test.go` (renderer/parser/wiring) | none - run in plain `go test ./backend/go/dllm/...` | Ginkgo specs over a fake `generator` seam; canonical renderer fixtures from transformers' `test_modeling_diffusion_gemma.py`, parser tables from the vLLM gemma4 parsers |
-| `backend/go/dllm/dllm_test.go` C-ABI smoke | `DLLM_TEST_LIBRARY` + `DLLM_TEST_TINY_MODEL` (dllm.cpp's `tests/fixtures/tiny_with_vocab.gguf`); Skips when unset | Drives the real `libdllm.so`: ABI check, load, tokenize `[2,18]`, deterministic generate, cancel (incl. mid-stream `Dllm.Cancel` aborting a deliberately slow `eb_max_steps:256` run in ~10ms) |
-| `tests/e2e-backends/dllm_test.go` | `BACKEND_TEST_DLLM=1` + `BACKEND_BINARY` (packaged run.sh) + `BACKEND_TEST_MODEL_FILE` (tiny fixture) | Templated chat round trip (Messages + UseTokenizerTemplate) over the real gRPC binary, non-streaming + streaming; plus client-context cancellation mid-stream (proves the `Cancellable` server plumbing end to end) |
-| Real-model e2e | `BACKEND_TEST_DLLM_REAL_MODEL_FILE` (26B BF16, ~50 GB) + `BACKEND_TEST_DLLM_REAL_GPU_LAYERS` | CUDA-13-class hardware only |
-
-Tool-call e2e is deliberately absent from the tiny-model spec: the fixture has
-random weights and cannot be coaxed into emitting tool markup; the unit tables
-carry that coverage.
-
-## Build matrix
-
-`cpu-dllm` (amd64 + arm64), `cuda13-dllm` (amd64), and
-`cuda13-nvidia-l4t-arm64-dllm` (arm64 CUDA: Jetson / DGX Spark GB10), via
-`.github/backend-matrix.yml`. No darwin/Metal. CUDA builds forward
-`-DDLLM_CUDA=ON` (dllm.cpp gates ggml's CUDA behind its own flag - a bare
-`-DGGML_CUDA=ON` is overridden by the cache FORCE). `libdllm.so` is
-self-contained (ggml statically absorbed, PIC), so `package.sh` only ships
-the binary, `run.sh` and that one .so (the parakeet-cpp-style stub layout;
-no ldd walk yet).
-
-## Known limitations
-
-- **Cancel granularity**: the C-ABI cancel flag is per-ctx and resets on
-  every generate entry, so a Cancel racing a NEW generate can be lost, and
-  with requests queued on the worker it aborts whichever generate is
-  currently running (acceptable: the server de-registers the hook on normal
-  completion, one process serves one model).
-- **Throughput**: ~0.15 tok/s on the 26B at default settings (GB10) - every
-  denoise step recomputes the full prompt+canvas. The upstream prefix-KV
-  cache (dllm.cpp P3) is the fix; `kv_cache:on` errors until it lands
-  (`auto`/`off` are accepted no-ops).
-- **Repo privacy**: see the note at the top - CI clone of dllm.cpp needs the
-  repo published (or credentials) before the backend images can build.
-- Engine spec/validation references: dllm.cpp `docs/validation.md` and
-  LocalAI `docs/superpowers/specs/2026-06-10-dllm-cpp-design.md`.

.agents/ds4-backend.md

@@ -44,6 +44,39 @@ maps to `DS4_THINK_HIGH`. We pass the chosen mode to `ds4_chat_append_assistant_
 via `ModelOptions.Options[] = "kv_cache_dir:/some/path"`. Format is **our own** -
 NOT bit-compatible with ds4-server's KVC files (interop is a follow-up plan).
 
+## Engine options (LoadModel)
+
+`LoadModel` maps `ModelOptions.Options[]` (`"key:value"`, from model-YAML
+`options:`) onto `ds4_engine_options` through a **declarative table**
+(`kEngineOptSpecs` + `apply_engine_option` in `grpc-server.cpp`). The struct is
+plain C with no reflection, so the field set is enumerated once in the table;
+adding a future engine knob is a one-line table row, not a new branch. Unknown
+keys are ignored (back-compat). A bare flag (`ssd_streaming` with no value)
+means `true`. Path-type values (`mtp_path`, `expert_profile_path`,
+`directional_steering_file`) resolve **relative to the model directory**, so a
+gallery entry can reference a companion file it downloaded by bare filename;
+absolute values pass through. `ds4_role` / `ds4_layers` / `ds4_listen` /
+`ds4_route_timeout` / `kv_cache_dir` keep their dedicated handling (validation
++ coordinator wiring) and are not in the table.
+
+Wired keys: `mtp_path`, `mtp_draft`, `mtp_margin`, `prefill_chunk`,
+`power_percent`, `warm_weights`, `quality`, `ssd_streaming`,
+`ssd_streaming_cold`, `ssd_streaming_preload_experts`,
+`ssd_streaming_cache_experts` (count or `NGB`, sets both experts+bytes via
+`ds4_parse_streaming_cache_experts_arg`), `simulate_used_memory` (`NGB` via
+`ds4_parse_gib_arg`), `expert_profile_path`, `directional_steering_file`,
+`directional_steering_attn`, `directional_steering_ffn`.
+
+## SSD streaming (running models larger than RAM)
+
+ds4's **SSD streaming** keeps non-routed weights resident and streams routed MoE
+experts from the GGUF on cache misses, turning "does it fit in RAM" into a speed
+spectrum. **Metal (Darwin) only** - it is a no-op on CUDA/CPU. Enable with
+`options: ["ssd_streaming"]`; size the routed-expert cache with
+`ssd_streaming_cache_experts:NGB` (omit for ds4's automatic 80%-of-working-set
+budget). Gallery entries built on this: `deepseek-v4-flash-q4-ssd` (153 GB Flash
+on a 128 GB Mac) and `deepseek-v4-pro-q2-ssd` (433 GB Pro, experimental).
+
 ## Build matrix
 
 | Build | Where | Notes |

.dockerignore

@@ -31,6 +31,15 @@ backend/python/**/source
 backend/cpp/llama-cpp/llama.cpp
 backend/cpp/llama-cpp-*-build
 
+# privacy-filter: same in-place pattern. The Makefile fetches privacy-filter.cpp
+# at the pinned commit (or symlinks a PRIVACY_FILTER_SRC checkout for local dev).
+# A stale dir/symlink COPY'd into the image makes the clone step fail (dangling
+# symlink) or compile against the wrong commit, so keep host build state out.
+backend/cpp/privacy-filter/privacy-filter.cpp
+backend/cpp/privacy-filter/build
+backend/cpp/privacy-filter/grpc-server
+backend/cpp/privacy-filter/package
+
 # Rust backend build output (sources are tracked; target/ is generated)
 backend/rust/*/target
 

.github/backend-matrix.yml

@@ -703,6 +703,32 @@ include:
     dockerfile: "./backend/Dockerfile.golang"
     context: "./"
     ubuntu-version: '2404'
+  - build-type: 'cublas'
+    cuda-major-version: "12"
+    cuda-minor-version: "8"
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-nvidia-cuda-12-locate-anything-cpp'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "locate-anything-cpp"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
+  - build-type: 'cublas'
+    cuda-major-version: "12"
+    cuda-minor-version: "8"
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-nvidia-cuda-12-depth-anything-cpp'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "depth-anything-cpp"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
   - build-type: 'cublas'
     cuda-major-version: "12"
     cuda-minor-version: "8"
@@ -768,6 +794,19 @@ include:
     dockerfile: "./backend/Dockerfile.golang"
     context: "./"
     ubuntu-version: '2404'
+  - build-type: 'cublas'
+    cuda-major-version: "12"
+    cuda-minor-version: "8"
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-nvidia-cuda-12-omnivoice-cpp'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "omnivoice-cpp"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
   - build-type: 'cublas'
     cuda-major-version: "12"
     cuda-minor-version: "8"
@@ -1543,6 +1582,32 @@ include:
     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-locate-anything-cpp'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "locate-anything-cpp"
+    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-depth-anything-cpp'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "depth-anything-cpp"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
   - build-type: 'cublas'
     cuda-major-version: "13"
     cuda-minor-version: "0"
@@ -1569,6 +1634,32 @@ include:
     backend: "rfdetr-cpp"
     dockerfile: "./backend/Dockerfile.golang"
     context: "./"
+  - 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-locate-anything-cpp'
+    base-image: "ubuntu:24.04"
+    ubuntu-version: '2404'
+    runs-on: 'ubuntu-24.04-arm'
+    backend: "locate-anything-cpp"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+  - 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-depth-anything-cpp'
+    base-image: "ubuntu:24.04"
+    ubuntu-version: '2404'
+    runs-on: 'ubuntu-24.04-arm'
+    backend: "depth-anything-cpp"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
   - build-type: 'cublas'
     cuda-major-version: "13"
     cuda-minor-version: "0"
@@ -1608,19 +1699,6 @@ include:
     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-dllm'
-    runs-on: 'ubuntu-latest'
-    base-image: "ubuntu:24.04"
-    skip-drivers: 'false'
-    backend: "dllm"
-    dockerfile: "./backend/Dockerfile.golang"
-    context: "./"
-    ubuntu-version: '2404'
   - build-type: 'cublas'
     cuda-major-version: "13"
     cuda-minor-version: "0"
@@ -1660,19 +1738,6 @@ include:
     backend: "parakeet-cpp"
     dockerfile: "./backend/Dockerfile.golang"
     context: "./"
-  - 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-dllm'
-    base-image: "ubuntu:24.04"
-    ubuntu-version: '2404'
-    runs-on: 'ubuntu-24.04-arm'
-    backend: "dllm"
-    dockerfile: "./backend/Dockerfile.golang"
-    context: "./"
   - build-type: 'cublas'
     cuda-major-version: "13"
     cuda-minor-version: "0"
@@ -1699,6 +1764,19 @@ include:
     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-omnivoice-cpp'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "omnivoice-cpp"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
   - build-type: 'cublas'
     cuda-major-version: "13"
     cuda-minor-version: "0"
@@ -1738,6 +1816,19 @@ include:
     backend: "qwen3-tts-cpp"
     dockerfile: "./backend/Dockerfile.golang"
     context: "./"
+  - 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-omnivoice-cpp'
+    base-image: "ubuntu:24.04"
+    ubuntu-version: '2404'
+    runs-on: 'ubuntu-24.04-arm'
+    backend: "omnivoice-cpp"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
   - build-type: 'cublas'
     cuda-major-version: "13"
     cuda-minor-version: "0"
@@ -2579,6 +2670,78 @@ include:
     dockerfile: "./backend/Dockerfile.ds4"
     context: "./"
     ubuntu-version: '2404'
+  # privacy-filter: PII/NER token classifier (per-arch native -> manifest merge).
+  # Every variant builds FROM a prebuilt quay.io/go-skynet/ci-cache:base-grpc-*
+  # image (gRPC + cmake + protoc + conditional CUDA/Vulkan already installed),
+  # exactly like llama-cpp — no toolchain is installed in Dockerfile.privacy-filter.
+  # builder-base-image makes the workflow use the Dockerfile's builder-prebuilt
+  # stage; without it (local builds) the builder-fromsource stage runs the same
+  # .docker/install-base-deps.sh.
+  - build-type: ''
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    platform-tag: 'amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-cpu-privacy-filter'
+    builder-base-image: 'quay.io/go-skynet/ci-cache:base-grpc-amd64'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'true'
+    backend: "privacy-filter"
+    dockerfile: "./backend/Dockerfile.privacy-filter"
+    context: "./"
+    ubuntu-version: '2404'
+  - build-type: ''
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/arm64'
+    platform-tag: 'arm64'
+    tag-latest: 'auto'
+    tag-suffix: '-cpu-privacy-filter'
+    builder-base-image: 'quay.io/go-skynet/ci-cache:base-grpc-arm64'
+    runs-on: 'ubuntu-24.04-arm'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'true'
+    backend: "privacy-filter"
+    dockerfile: "./backend/Dockerfile.privacy-filter"
+    context: "./"
+    ubuntu-version: '2404'
+  # Vulkan: base-grpc-vulkan-amd64 carries the SDK. arm64 vulkan is a one-line
+  # add once amd64 is proven in CI.
+  - build-type: 'vulkan'
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    platform-tag: 'amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-vulkan-privacy-filter'
+    builder-base-image: 'quay.io/go-skynet/ci-cache:base-grpc-vulkan-amd64'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "privacy-filter"
+    dockerfile: "./backend/Dockerfile.privacy-filter"
+    context: "./"
+    ubuntu-version: '2404'
+  # CUDA: base-grpc-cuda-13-amd64 carries the toolkit; BUILD_TYPE=cublas ->
+  # -DPF_CUDA=ON. cuda-12 and arm64/l4t are one-line adds once cuda-13 amd64 is
+  # proven in CI.
+  - build-type: 'cublas'
+    cuda-major-version: "13"
+    cuda-minor-version: "0"
+    platforms: 'linux/amd64'
+    platform-tag: 'amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-nvidia-cuda-13-privacy-filter'
+    builder-base-image: 'quay.io/go-skynet/ci-cache:base-grpc-cuda-13-amd64'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'true'
+    backend: "privacy-filter"
+    dockerfile: "./backend/Dockerfile.privacy-filter"
+    context: "./"
+    ubuntu-version: '2404'
   - build-type: ''
     cuda-major-version: ""
     cuda-minor-version: ""
@@ -2832,6 +2995,141 @@ include:
     dockerfile: "./backend/Dockerfile.golang"
     context: "./"
     ubuntu-version: '2404'
+  # locate-anything-cpp
+  - build-type: ''
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-cpu-locate-anything-cpp'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "locate-anything-cpp"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
+  - build-type: ''
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-cpu-depth-anything-cpp'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "depth-anything-cpp"
+    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-locate-anything-cpp'
+    runs-on: 'ubuntu-latest'
+    base-image: "intel/oneapi-basekit:2025.3.0-0-devel-ubuntu24.04"
+    skip-drivers: 'false'
+    backend: "locate-anything-cpp"
+    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-depth-anything-cpp'
+    runs-on: 'ubuntu-latest'
+    base-image: "intel/oneapi-basekit:2025.3.0-0-devel-ubuntu24.04"
+    skip-drivers: 'false'
+    backend: "depth-anything-cpp"
+    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-locate-anything-cpp'
+    runs-on: 'ubuntu-latest'
+    base-image: "intel/oneapi-basekit:2025.3.0-0-devel-ubuntu24.04"
+    skip-drivers: 'false'
+    backend: "locate-anything-cpp"
+    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-depth-anything-cpp'
+    runs-on: 'ubuntu-latest'
+    base-image: "intel/oneapi-basekit:2025.3.0-0-devel-ubuntu24.04"
+    skip-drivers: 'false'
+    backend: "depth-anything-cpp"
+    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-locate-anything-cpp'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "locate-anything-cpp"
+    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-depth-anything-cpp'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "depth-anything-cpp"
+    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-locate-anything-cpp'
+    runs-on: 'ubuntu-24.04-arm'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "locate-anything-cpp"
+    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-depth-anything-cpp'
+    runs-on: 'ubuntu-24.04-arm'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "depth-anything-cpp"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
   - build-type: 'sycl_f32'
     cuda-major-version: ""
     cuda-minor-version: ""
@@ -2925,6 +3223,32 @@ include:
     dockerfile: "./backend/Dockerfile.golang"
     context: "./"
     ubuntu-version: '2204'
+  - 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-locate-anything-cpp'
+    base-image: "nvcr.io/nvidia/l4t-jetpack:r36.4.0"
+    runs-on: 'ubuntu-24.04-arm'
+    backend: "locate-anything-cpp"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2204'
+  - 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-depth-anything-cpp'
+    base-image: "nvcr.io/nvidia/l4t-jetpack:r36.4.0"
+    runs-on: 'ubuntu-24.04-arm'
+    backend: "depth-anything-cpp"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2204'
   # whisper
   - build-type: ''
     cuda-major-version: ""
@@ -3171,35 +3495,6 @@ include:
     dockerfile: "./backend/Dockerfile.golang"
     context: "./"
     ubuntu-version: '2404'
-  # dllm
-  - build-type: ''
-    cuda-major-version: ""
-    cuda-minor-version: ""
-    platforms: 'linux/amd64'
-    platform-tag: 'amd64'
-    tag-latest: 'auto'
-    tag-suffix: '-cpu-dllm'
-    runs-on: 'ubuntu-latest'
-    base-image: "ubuntu:24.04"
-    skip-drivers: 'false'
-    backend: "dllm"
-    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-dllm'
-    runs-on: 'ubuntu-24.04-arm'
-    base-image: "ubuntu:24.04"
-    skip-drivers: 'false'
-    backend: "dllm"
-    dockerfile: "./backend/Dockerfile.golang"
-    context: "./"
-    ubuntu-version: '2404'
   - build-type: 'sycl_f32'
     cuda-major-version: ""
     cuda-minor-version: ""
@@ -3418,6 +3713,35 @@ include:
     dockerfile: "./backend/Dockerfile.golang"
     context: "./"
     ubuntu-version: '2404'
+  # omnivoice-cpp
+  - build-type: ''
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    platform-tag: 'amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-cpu-omnivoice-cpp'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "omnivoice-cpp"
+    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-omnivoice-cpp'
+    runs-on: 'ubuntu-24.04-arm'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "omnivoice-cpp"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
   - build-type: 'sycl_f32'
     cuda-major-version: ""
     cuda-minor-version: ""
@@ -3431,6 +3755,19 @@ include:
     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-omnivoice-cpp'
+    runs-on: 'ubuntu-latest'
+    base-image: "intel/oneapi-basekit:2025.3.0-0-devel-ubuntu24.04"
+    skip-drivers: 'false'
+    backend: "omnivoice-cpp"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
   - build-type: 'sycl_f16'
     cuda-major-version: ""
     cuda-minor-version: ""
@@ -3444,6 +3781,19 @@ include:
     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-omnivoice-cpp'
+    runs-on: 'ubuntu-latest'
+    base-image: "intel/oneapi-basekit:2025.3.0-0-devel-ubuntu24.04"
+    skip-drivers: 'false'
+    backend: "omnivoice-cpp"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
   - build-type: 'vulkan'
     cuda-major-version: ""
     cuda-minor-version: ""
@@ -3458,6 +3808,20 @@ include:
     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-omnivoice-cpp'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "omnivoice-cpp"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
   - build-type: 'vulkan'
     cuda-major-version: ""
     cuda-minor-version: ""
@@ -3472,6 +3836,20 @@ include:
     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-omnivoice-cpp'
+    runs-on: 'ubuntu-24.04-arm'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "omnivoice-cpp"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
   - build-type: 'cublas'
     cuda-major-version: "12"
     cuda-minor-version: "0"
@@ -3485,6 +3863,19 @@ include:
     dockerfile: "./backend/Dockerfile.golang"
     context: "./"
     ubuntu-version: '2204'
+  - 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-omnivoice-cpp'
+    base-image: "nvcr.io/nvidia/l4t-jetpack:r36.4.0"
+    runs-on: 'ubuntu-24.04-arm'
+    backend: "omnivoice-cpp"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2204'
   - build-type: 'hipblas'
     cuda-major-version: ""
     cuda-minor-version: ""
@@ -3498,6 +3889,19 @@ include:
     dockerfile: "./backend/Dockerfile.golang"
     context: "./"
     ubuntu-version: '2404'
+  - build-type: 'hipblas'
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-rocm-hipblas-omnivoice-cpp'
+    base-image: "rocm/dev-ubuntu-24.04:6.4.4"
+    runs-on: 'ubuntu-latest'
+    skip-drivers: 'false'
+    backend: "omnivoice-cpp"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
   # vibevoice-cpp
   - build-type: ''
     cuda-major-version: ""
@@ -4277,6 +4681,36 @@ include:
     dockerfile: "./backend/Dockerfile.golang"
     context: "./"
     ubuntu-version: '2404'
+  # supertonic CPU (amd64)
+  - build-type: ''
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    platform-tag: 'amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-cpu-supertonic'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "supertonic"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
+  # supertonic CPU (arm64)
+  - build-type: ''
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/arm64'
+    platform-tag: 'arm64'
+    tag-latest: 'auto'
+    tag-suffix: '-cpu-supertonic'
+    runs-on: 'ubuntu-24.04-arm'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "supertonic"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
 
 # Darwin matrix (consumed by backend-jobs-darwin).
 includeDarwin:
@@ -4328,6 +4762,10 @@ includeDarwin:
     tag-suffix: "-metal-darwin-arm64-qwen3-tts-cpp"
     build-type: "metal"
     lang: "go"
+  - backend: "omnivoice-cpp"
+    tag-suffix: "-metal-darwin-arm64-omnivoice-cpp"
+    build-type: "metal"
+    lang: "go"
   - backend: "vibevoice-cpp"
     tag-suffix: "-metal-darwin-arm64-vibevoice-cpp"
     build-type: "metal"
@@ -4396,6 +4834,10 @@ includeDarwin:
     tag-suffix: "-metal-darwin-arm64-silero-vad"
     build-type: "metal"
     lang: "go"
+  - backend: "sherpa-onnx"
+    tag-suffix: "-metal-darwin-arm64-sherpa-onnx"
+    build-type: "metal"
+    lang: "go"
   - backend: "local-store"
     tag-suffix: "-metal-darwin-arm64-local-store"
     build-type: "metal"
@@ -4403,3 +4845,9 @@ includeDarwin:
   - backend: "llama-cpp-quantization"
     tag-suffix: "-metal-darwin-arm64-llama-cpp-quantization"
     build-type: "mps"
+  - backend: "speaker-recognition"
+    tag-suffix: "-metal-darwin-arm64-speaker-recognition"
+    build-type: "mps"
+  - backend: "ds4"
+    tag-suffix: "-metal-darwin-arm64-ds4"
+    lang: "go"

.github/workflows/backend_build_darwin.yml

@@ -98,6 +98,7 @@ jobs:
             /opt/homebrew/Cellar/hiredis
             /opt/homebrew/Cellar/xxhash
             /opt/homebrew/Cellar/zstd
+            /opt/homebrew/Cellar/nlohmann-json
           key: brew-${{ runner.os }}-${{ runner.arch }}-v1-${{ hashFiles('.github/workflows/backend_build_darwin.yml') }}
 
       - name: Dependencies
@@ -109,7 +110,10 @@ jobs:
           # Without explicitly installing them, a brew cache-hit run restores
           # ccache's Cellar dir but skips installing those transitive deps,
           # and ccache fails at runtime with `dyld: Library not loaded`.
-          brew install protobuf grpc make protoc-gen-go protoc-gen-go-grpc libomp llvm ccache blake3 fmt hiredis xxhash zstd
+          # 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
           # 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
@@ -128,7 +132,7 @@ 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 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 2>/dev/null || true
 
       - name: Save Homebrew cache
         if: github.event_name != 'pull_request' && steps.brew-cache.outputs.cache-hit != 'true'
@@ -148,6 +152,7 @@ jobs:
             /opt/homebrew/Cellar/hiredis
             /opt/homebrew/Cellar/xxhash
             /opt/homebrew/Cellar/zstd
+            /opt/homebrew/Cellar/nlohmann-json
           key: brew-${{ runner.os }}-${{ runner.arch }}-v1-${{ hashFiles('.github/workflows/backend_build_darwin.yml') }}
 
       # ---- ccache for llama.cpp CMake builds ----

.github/workflows/bump_deps.yaml

@@ -26,6 +26,10 @@ jobs:
             variable: "DS4_VERSION"
             branch: "main"
             file: "backend/cpp/ds4/Makefile"
+          - repository: "localai-org/privacy-filter.cpp"
+            variable: "PRIVACY_FILTER_VERSION"
+            branch: "master"
+            file: "backend/cpp/privacy-filter/Makefile"
           - repository: "ggml-org/whisper.cpp"
             variable: "WHISPER_CPP_VERSION"
             branch: "master"
@@ -38,10 +42,10 @@ jobs:
             variable: "PARAKEET_VERSION"
             branch: "master"
             file: "backend/go/parakeet-cpp/Makefile"
-          - repository: "mudler/dllm.cpp"
-            variable: "DLLM_VERSION"
-            branch: "main"
-            file: "backend/go/dllm/Makefile"
+          - repository: "mudler/depth-anything.cpp"
+            variable: "DEPTHANYTHING_VERSION"
+            branch: "master"
+            file: "backend/go/depth-anything-cpp/Makefile"
           - repository: "leejet/stable-diffusion.cpp"
             variable: "STABLEDIFFUSION_GGML_VERSION"
             branch: "master"
@@ -66,10 +70,18 @@ jobs:
             variable: "RFDETR_VERSION"
             branch: "main"
             file: "backend/go/rfdetr-cpp/Makefile"
-          - repository: "predict-woo/qwen3-tts.cpp"
+          - repository: "mudler/locate-anything.cpp"
+            variable: "LOCATEANYTHING_VERSION"
+            branch: "master"
+            file: "backend/go/locate-anything-cpp/Makefile"
+          - repository: "ServeurpersoCom/qwentts.cpp"
             variable: "QWEN3TTS_CPP_VERSION"
-            branch: "main"
+            branch: "master"
             file: "backend/go/qwen3-tts-cpp/Makefile"
+          - repository: "ServeurpersoCom/omnivoice.cpp"
+            variable: "OMNIVOICE_VERSION"
+            branch: "master"
+            file: "backend/go/omnivoice-cpp/Makefile"
           - repository: "localai-org/vibevoice.cpp"
             variable: "VIBEVOICE_CPP_VERSION"
             branch: "master"

.github/workflows/secscan.yaml

@@ -21,7 +21,10 @@ jobs:
         uses: securego/gosec@v2.27.1
         with:
           # we let the report trigger content trigger a failure using the GitHub Security features.
-          args: '-no-fail -fmt sarif -out results.sarif ./...'
+          # backend/go/supertonic is excluded: it vendors upstream supertone-inc/supertonic
+          # (helper.go), whose findings (G304 model-file loads, G404 math/rand for flow-matching
+          # noise, G104 unhandled errors) are inherent to that upstream code, not ours to rewrite.
+          args: '-no-fail -exclude-dir=backend/go/supertonic -fmt sarif -out results.sarif ./...'
       - name: Upload SARIF file
         if: ${{ github.actor != 'dependabot[bot]' }}
         uses: github/codeql-action/upload-sarif@v4

.github/workflows/test-extra.yml

@@ -38,6 +38,7 @@ jobs:
       acestep-cpp: ${{ steps.detect.outputs.acestep-cpp }}
       qwen3-tts-cpp: ${{ steps.detect.outputs.qwen3-tts-cpp }}
       rfdetr-cpp: ${{ steps.detect.outputs.rfdetr-cpp }}
+      locate-anything-cpp: ${{ steps.detect.outputs.locate-anything-cpp }}
       vibevoice-cpp: ${{ steps.detect.outputs.vibevoice-cpp }}
       localvqe: ${{ steps.detect.outputs.localvqe }}
       voxtral: ${{ steps.detect.outputs.voxtral }}
@@ -563,7 +564,7 @@ jobs:
       - name: Run e2e-backends smoke
         env:
           BACKEND_IMAGE: quay.io/go-skynet/local-ai-backends:master-cpu-llama-cpp
-          BACKEND_TEST_CAPS: health,load,predict,stream,logprobs,logit_bias
+          BACKEND_TEST_CAPS: health,load,predict,stream,logprobs,logit_bias,tokenize
         run: |
           make test-extra-backend
   # Realtime e2e with sherpa-onnx driving VAD + STT + TTS against a mocked LLM.
@@ -901,6 +902,45 @@ jobs:
       - name: Test rfdetr-cpp
         run: |
           make --jobs=5 --output-sync=target -C backend/go/rfdetr-cpp test
+  # Per-backend e2e for locate-anything-cpp: builds the .so + Go binary and
+  # runs `make -C backend/go/locate-anything-cpp test`. test.sh fetches the
+  # locate-anything-q8_0 GGUF (~6.3 GB, NVIDIA LocateAnything-3B) from the
+  # published mudler/locate-anything.cpp-gguf HF repo + a COCO image, then the
+  # Go wire test loads the model and runs an open-vocabulary Detect, asserting
+  # at least one labeled box. Heavier than the other Go backends (it is a 3B),
+  # so it is gated to changes under backend/go/locate-anything-cpp/.
+  tests-locate-anything-cpp:
+    needs: detect-changes
+    if: needs.detect-changes.outputs.locate-anything-cpp == 'true' || needs.detect-changes.outputs.run-all == 'true'
+    runs-on: ubuntu-latest
+    steps:
+      - name: Clone
+        uses: actions/checkout@v6
+        with:
+          submodules: true
+      - name: Dependencies
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y build-essential cmake curl libopenblas-dev
+      - name: Setup Go
+        uses: actions/setup-go@v5
+      - name: Display Go version
+        run: go version
+      - name: Proto Dependencies
+        run: |
+          # Install protoc
+          curl -L -s https://github.com/protocolbuffers/protobuf/releases/download/v26.1/protoc-26.1-linux-x86_64.zip -o protoc.zip && \
+          unzip -j -d /usr/local/bin protoc.zip bin/protoc && \
+          rm protoc.zip
+          go install google.golang.org/protobuf/cmd/protoc-gen-go@v1.34.2
+          go install google.golang.org/grpc/cmd/protoc-gen-go-grpc@1958fcbe2ca8bd93af633f11e97d44e567e945af
+          PATH="$PATH:$HOME/go/bin" make protogen-go
+      - name: Build locate-anything-cpp
+        run: |
+          make --jobs=5 --output-sync=target -C backend/go/locate-anything-cpp
+      - name: Test locate-anything-cpp
+        run: |
+          make --jobs=5 --output-sync=target -C backend/go/locate-anything-cpp test
   # Per-backend smoke for vibevoice-cpp: builds the .so + Go binary and
   # runs `make -C backend/go/vibevoice-cpp test`. test.sh auto-downloads
   # the published mudler/vibevoice.cpp-models bundle (TTS Q8_0 + ASR Q4_K

.golangci.yml

@@ -74,6 +74,8 @@ linters:
     paths:
       # Upstream whisper.cpp source tree fetched by the whisper backend Makefile.
       - 'backend/go/whisper/sources'
+      # Vendored upstream supertonic pipeline (supertone-inc/supertonic go/helper.go).
+      - 'backend/go/supertonic/helper.go'
       - 'docs/'
     rules:
       # CLI entry points: kong's `env:"..."` tag is the legitimate env→struct

AGENTS.md

@@ -26,7 +26,6 @@ LocalAI follows the Linux kernel project's [guidelines for AI coding assistants]
 | [.agents/vllm-backend.md](.agents/vllm-backend.md) | Working on the vLLM / vLLM-omni backends — native parsers, ChatDelta, CPU build, libnuma packaging, backend hooks |
 | [.agents/sglang-backend.md](.agents/sglang-backend.md) | Working on the SGLang backend — `engine_args` validation against ServerArgs, speculative-decoding (EAGLE/EAGLE3/DFLASH/MTP) recipes, parser handling |
 | [.agents/ds4-backend.md](.agents/ds4-backend.md) | Working on the ds4 backend - DSML state machine, thinking modes, KV cache, Metal+CUDA matrix |
-| [.agents/dllm-backend.md](.agents/dllm-backend.md) | Working on the dllm backend (DiffusionGemma block-diffusion) - purego C-ABI binding, per-ctx serialization contract, gemma4 renderer/parser, gated test layers |
 | [.agents/testing-mcp-apps.md](.agents/testing-mcp-apps.md) | Testing MCP Apps (interactive tool UIs) in the React UI |
 | [.agents/api-endpoints-and-auth.md](.agents/api-endpoints-and-auth.md) | Adding API endpoints, auth middleware, feature permissions, user access control |
 | [.agents/debugging-backends.md](.agents/debugging-backends.md) | Debugging runtime backend failures, dependency conflicts, rebuilding backends |

Dockerfile

@@ -108,6 +108,7 @@ RUN <<EOT bash
         apt-get update && \
         apt-get install -y --no-install-recommends \
             cuda-nvcc-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} \
+            cuda-nvrtc-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} \
             libcufft-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} \
             libcurand-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} \
             libcublas-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} \

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/dllm 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/vibevoice-cpp backends/localvqe backends/tinygrad backends/sherpa-onnx backends/ds4 backends/ds4-darwin backends/liquid-audio
+.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
 
 GOCMD=go
 GOTEST=$(GOCMD) test
@@ -566,6 +566,7 @@ prepare-test-extra: protogen-python
 	$(MAKE) -C backend/python/speaker-recognition
 	$(MAKE) -C backend/rust/kokoros kokoros-grpc
 	$(MAKE) -C backend/go/rfdetr-cpp
+	$(MAKE) -C backend/go/locate-anything-cpp
 
 test-extra: prepare-test-extra
 	$(MAKE) -C backend/python/transformers test
@@ -593,6 +594,9 @@ test-extra: prepare-test-extra
 	$(MAKE) -C backend/python/speaker-recognition test
 	$(MAKE) -C backend/rust/kokoros test
 	$(MAKE) -C backend/go/rfdetr-cpp test
+	$(MAKE) -C backend/go/locate-anything-cpp test
+	$(MAKE) -C backend/go/depth-anything-cpp test
+	$(MAKE) -C backend/go/supertonic test
 
 ##
 ## End-to-end gRPC tests that exercise a built backend container image.
@@ -1160,6 +1164,10 @@ BACKEND_TURBOQUANT = turboquant|turboquant|.|false|false
 # Single-model; hardware-only validation lives at tests/e2e-backends/
 # (BACKEND_BINARY mode); see docs/superpowers/plans/2026-05-11-ds4-backend.md.
 BACKEND_DS4 = ds4|ds4|.|false|false
+# privacy-filter wraps the standalone privacy-filter.cpp GGML engine (the
+# openai-privacy-filter PII/NER token classifier) — the TokenClassify RPC for
+# the PII redactor tier, on stock ggml with no llama.cpp carry-patches.
+BACKEND_PRIVACY_FILTER = privacy-filter|privacy-filter|.|false|false
 
 # Golang backends
 BACKEND_PIPER = piper|golang|.|false|true
@@ -1171,16 +1179,16 @@ BACKEND_STABLEDIFFUSION_GGML = stablediffusion-ggml|golang|.|--progress=plain|tr
 BACKEND_WHISPER = whisper|golang|.|false|true
 BACKEND_CRISPASR = crispasr|golang|.|false|true
 BACKEND_PARAKEET_CPP = parakeet-cpp|golang|.|false|true
-# dllm is mudler/dllm.cpp, the DiffusionGemma block-diffusion engine,
-# wrapped by the purego backend at backend/go/dllm.
-BACKEND_DLLM = dllm|golang|.|false|true
+BACKEND_DEPTH_ANYTHING_CPP = depth-anything-cpp|golang|.|false|true
 BACKEND_VOXTRAL = voxtral|golang|.|false|true
 BACKEND_ACESTEP_CPP = acestep-cpp|golang|.|false|true
 BACKEND_QWEN3_TTS_CPP = qwen3-tts-cpp|golang|.|false|true
+BACKEND_OMNIVOICE_CPP = omnivoice-cpp|golang|.|false|true
 BACKEND_VIBEVOICE_CPP = vibevoice-cpp|golang|.|false|true
 BACKEND_LOCALVQE = localvqe|golang|.|false|true
 BACKEND_OPUS = opus|golang|.|false|true
 BACKEND_SHERPA_ONNX = sherpa-onnx|golang|.|false|true
+BACKEND_SUPERTONIC = supertonic|golang|.|false|true
 
 # Python backends with root context
 BACKEND_RERANKERS = rerankers|python|.|false|true
@@ -1254,6 +1262,7 @@ $(eval $(call generate-docker-build-target,$(BACKEND_LLAMA_CPP)))
 $(eval $(call generate-docker-build-target,$(BACKEND_IK_LLAMA_CPP)))
 $(eval $(call generate-docker-build-target,$(BACKEND_TURBOQUANT)))
 $(eval $(call generate-docker-build-target,$(BACKEND_DS4)))
+$(eval $(call generate-docker-build-target,$(BACKEND_PRIVACY_FILTER)))
 $(eval $(call generate-docker-build-target,$(BACKEND_PIPER)))
 $(eval $(call generate-docker-build-target,$(BACKEND_LOCAL_STORE)))
 $(eval $(call generate-docker-build-target,$(BACKEND_CLOUD_PROXY)))
@@ -1263,7 +1272,7 @@ $(eval $(call generate-docker-build-target,$(BACKEND_STABLEDIFFUSION_GGML)))
 $(eval $(call generate-docker-build-target,$(BACKEND_WHISPER)))
 $(eval $(call generate-docker-build-target,$(BACKEND_CRISPASR)))
 $(eval $(call generate-docker-build-target,$(BACKEND_PARAKEET_CPP)))
-$(eval $(call generate-docker-build-target,$(BACKEND_DLLM)))
+$(eval $(call generate-docker-build-target,$(BACKEND_DEPTH_ANYTHING_CPP)))
 $(eval $(call generate-docker-build-target,$(BACKEND_VOXTRAL)))
 $(eval $(call generate-docker-build-target,$(BACKEND_OPUS)))
 $(eval $(call generate-docker-build-target,$(BACKEND_RERANKERS)))
@@ -1296,6 +1305,7 @@ $(eval $(call generate-docker-build-target,$(BACKEND_WHISPERX)))
 $(eval $(call generate-docker-build-target,$(BACKEND_ACE_STEP)))
 $(eval $(call generate-docker-build-target,$(BACKEND_ACESTEP_CPP)))
 $(eval $(call generate-docker-build-target,$(BACKEND_QWEN3_TTS_CPP)))
+$(eval $(call generate-docker-build-target,$(BACKEND_OMNIVOICE_CPP)))
 $(eval $(call generate-docker-build-target,$(BACKEND_VIBEVOICE_CPP)))
 $(eval $(call generate-docker-build-target,$(BACKEND_LOCALVQE)))
 $(eval $(call generate-docker-build-target,$(BACKEND_MLX)))
@@ -1308,12 +1318,13 @@ $(eval $(call generate-docker-build-target,$(BACKEND_KOKOROS)))
 $(eval $(call generate-docker-build-target,$(BACKEND_SAM3_CPP)))
 $(eval $(call generate-docker-build-target,$(BACKEND_RFDETR_CPP)))
 $(eval $(call generate-docker-build-target,$(BACKEND_SHERPA_ONNX)))
+$(eval $(call generate-docker-build-target,$(BACKEND_SUPERTONIC)))
 
 # Pattern rule for docker-save targets
 docker-save-%: backend-images
 	docker save local-ai-backend:$* -o backend-images/$*.tar
 
-docker-build-backends: docker-build-llama-cpp docker-build-ik-llama-cpp docker-build-turboquant docker-build-ds4 docker-build-rerankers docker-build-vllm docker-build-vllm-omni docker-build-sglang docker-build-transformers docker-build-outetts docker-build-diffusers docker-build-kokoro docker-build-faster-whisper docker-build-crispasr docker-build-coqui docker-build-chatterbox docker-build-vibevoice docker-build-liquid-audio docker-build-moonshine docker-build-pocket-tts docker-build-qwen-tts docker-build-fish-speech docker-build-faster-qwen3-tts docker-build-qwen-asr docker-build-nemo docker-build-voxcpm docker-build-whisperx docker-build-ace-step docker-build-acestep-cpp docker-build-voxtral docker-build-mlx-distributed docker-build-trl docker-build-llama-cpp-quantization docker-build-tinygrad docker-build-kokoros docker-build-sam3-cpp docker-build-rfdetr-cpp docker-build-qwen3-tts-cpp docker-build-vibevoice-cpp docker-build-localvqe docker-build-insightface docker-build-speaker-recognition docker-build-sherpa-onnx docker-build-cloud-proxy
+docker-build-backends: docker-build-llama-cpp docker-build-ik-llama-cpp docker-build-turboquant docker-build-ds4 docker-build-rerankers docker-build-vllm docker-build-vllm-omni docker-build-sglang docker-build-transformers docker-build-outetts docker-build-diffusers docker-build-kokoro docker-build-faster-whisper docker-build-crispasr docker-build-coqui docker-build-chatterbox docker-build-vibevoice docker-build-liquid-audio docker-build-moonshine docker-build-pocket-tts docker-build-qwen-tts docker-build-fish-speech docker-build-faster-qwen3-tts docker-build-qwen-asr docker-build-nemo docker-build-voxcpm docker-build-whisperx docker-build-ace-step docker-build-acestep-cpp docker-build-voxtral docker-build-mlx-distributed docker-build-trl docker-build-llama-cpp-quantization docker-build-tinygrad docker-build-kokoros docker-build-sam3-cpp docker-build-rfdetr-cpp docker-build-qwen3-tts-cpp docker-build-omnivoice-cpp docker-build-vibevoice-cpp docker-build-localvqe docker-build-insightface docker-build-speaker-recognition docker-build-sherpa-onnx docker-build-cloud-proxy docker-build-supertonic docker-build-depth-anything-cpp docker-build-privacy-filter
 
 ########################################################
 ### Mock Backend for E2E Tests

README.md

@@ -29,6 +29,18 @@
 <a href="https://trendshift.io/repositories/5539" target="_blank"><img src="https://trendshift.io/api/badge/repositories/5539" alt="mudler%2FLocalAI | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a>
 </p>
 
+<!-- Keep these links, translations synced daily. -->
+<p align="center">
+<a href="https://zdoc.app/de/mudler/LocalAI">Deutsch</a> |
+<a href="https://zdoc.app/es/mudler/LocalAI">Español</a> |
+<a href="https://zdoc.app/fr/mudler/LocalAI">français</a> |
+<a href="https://zdoc.app/ja/mudler/LocalAI">日本語</a> |
+<a href="https://zdoc.app/ko/mudler/LocalAI">한국어</a> |
+<a href="https://zdoc.app/pt/mudler/LocalAI">Português</a> |
+<a href="https://zdoc.app/ru/mudler/LocalAI">Русский</a> |
+<a href="https://zdoc.app/zh/mudler/LocalAI">中文</a>
+</p>
+
 **LocalAI** is the open-source AI engine. Run any model - LLMs, vision, voice, image, video - on any hardware. No GPU required.
 
 **A small core, not a bundle.** Each backend wraps a best-in-class engine (llama.cpp, vLLM, whisper.cpp, stable-diffusion, MLX...) in its own image, pulled only when a model needs it. You install nothing you don't use.
@@ -165,6 +177,10 @@ For more details, see the [Getting Started guide](https://localai.io/basics/gett
 
 ## Latest News
 
+- **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).
+- **June 2026**: Distributed mode hardening: [prefix-cache-aware routing](https://github.com/mudler/LocalAI/pull/10071), a [production-ready request router with auto-sized embedding/rerank batches](https://github.com/mudler/LocalAI/pull/10104), [ds4 layer-split distributed inference](https://github.com/mudler/LocalAI/pull/10098), [NATS JWT auth + TLS/mTLS](https://github.com/mudler/LocalAI/pull/10159), and [resumable file uploads](https://github.com/mudler/LocalAI/pull/10109).
 - **May 2026**: **LocalAI 4.3.0** - `llama.cpp` [prompt cache on by default](https://github.com/mudler/LocalAI/pull/9925) (repeated system prompts collapse from minutes to seconds), [keyless cosign signing of backend OCI images](https://github.com/mudler/LocalAI/pull/9823), [per-API-key + per-user usage attribution](https://github.com/mudler/LocalAI/pull/9920), Distributed v3 with [per-request replica routing](https://github.com/mudler/LocalAI/pull/9968). [Release notes](https://github.com/mudler/LocalAI/releases/tag/v4.3.0)
 - **May 2026**: **LocalAI 4.2.0** - LocalAI sees and hears: [voice recognition](https://github.com/mudler/LocalAI/pull/9500), [face recognition + antispoofing liveness](https://github.com/mudler/LocalAI/pull/9480), speaker diarization. Plus [drop-in Ollama API](https://github.com/mudler/LocalAI/pull/9284), [video generation](https://github.com/mudler/LocalAI/pull/9420), redesigned UI with i18n + admin-configurable branding, vLLM at feature parity with llama.cpp, and 11 new backends. [Release notes](https://github.com/mudler/LocalAI/releases/tag/v4.2.0)
 - **April 2026**: **LocalAI 4.1.0** - LocalAI becomes a control tower: distributed cluster mode with VRAM-aware smart routing + autoscaling, multi-user platform with OIDC and API keys, per-user quotas with predictive analytics, in-UI fine-tuning with TRL (auto-export to GGUF), on-the-fly quantization backend, visual pipeline editor. [Release notes](https://github.com/mudler/LocalAI/releases/tag/v4.1.0)
@@ -204,10 +220,26 @@ For older news and full release notes, see [GitHub Releases](https://github.com/
 
 ## Supported Backends & Acceleration
 
-LocalAI supports **36+ backends** including llama.cpp, vLLM, transformers, whisper.cpp, diffusers, MLX, MLX-VLM, and many more. Hardware acceleration is available for **NVIDIA** (CUDA 12/13), **AMD** (ROCm), **Intel** (oneAPI/SYCL), **Apple Silicon** (Metal), **Vulkan**, and **NVIDIA Jetson** (L4T). All backends can be installed on-the-fly from the [Backend Gallery](https://localai.io/backends/).
+LocalAI supports **60+ backends** including llama.cpp, vLLM, SGLang, transformers, whisper.cpp, diffusers, MLX, MLX-VLM, and many more. Hardware acceleration is available for **NVIDIA** (CUDA 12/13), **AMD** (ROCm), **Intel** (oneAPI/SYCL), **Apple Silicon** (Metal), **Vulkan**, and **NVIDIA Jetson** (L4T). All backends can be installed on-the-fly from the [Backend Gallery](https://localai.io/backends/).
 
 See the full [Backend & Model Compatibility Table](https://localai.io/model-compatibility/) and [GPU Acceleration guide](https://localai.io/features/gpu-acceleration/).
 
+### Backends built by us
+
+Most backends wrap a best-in-class upstream engine. A handful of them are native C/C++/GGML engines (no Python at inference) developed and maintained by the LocalAI project itself:
+
+| Backend | What it does |
+|---------|-------------|
+| [parakeet.cpp](https://github.com/mudler/parakeet.cpp) | C++/GGML port of NVIDIA NeMo Parakeet ASR (tdt/ctc/rnnt/hybrid), with cache-aware streaming transcription |
+| [voxtral.c](https://github.com/mudler/voxtral.c) | Voxtral Realtime 4B speech-to-text in pure C |
+| [vibevoice.cpp](https://github.com/mudler/vibevoice.cpp) | Native port of Microsoft VibeVoice for TTS (voice cloning) and long-form ASR with speaker diarization |
+| [rf-detr.cpp](https://github.com/mudler/rf-detr.cpp) | Native RF-DETR object detection and instance segmentation |
+| [locate-anything.cpp](https://github.com/mudler/locate-anything.cpp) | Open-vocabulary object detection and visual grounding (LocateAnything-3B) |
+| [depth-anything.cpp](https://github.com/mudler/depth-anything.cpp) | Depth Anything 3 monocular metric depth + camera pose estimation |
+| [privacy-filter.cpp](https://github.com/localai-org/privacy-filter.cpp) | Standalone GGML PII/NER token-classification engine powering LocalAI's PII redaction tier |
+| [LocalVQE](https://github.com/localai-org/LocalVQE) | Joint acoustic echo cancellation, noise suppression, and dereverberation |
+| [local-store](https://github.com/mudler/LocalAI) | Local-first vector database for embeddings (shipped in-tree) |
+
 ## Resources
 
 - [Documentation](https://localai.io/)
@@ -217,7 +249,7 @@ See the full [Backend & Model Compatibility Table](https://localai.io/model-comp
 - [Integrations & community projects](https://localai.io/docs/integrations/)
 - [Installation video walkthrough](https://www.youtube.com/watch?v=cMVNnlqwfw4)
 - [Media & blog posts](https://localai.io/basics/news/#media-blogs-social)
-- [Examples](https://github.com/mudler/LocalAI-examples)
+- [Examples](https://github.com/mudler/LocalAI-examples) — including the [realtime voice assistant demo](https://github.com/localai-org/localai-realtime-demo) (Go client for the Realtime API with tool calling)
 
 ## Team
 

backend/Dockerfile.golang

@@ -206,6 +206,16 @@ RUN if [ "${BACKEND}" = "opus" ]; then \
     apt-get clean && rm -rf /var/lib/apt/lists/*; \
 fi
 
+# CrispASR's piper TTS backend dlopens libespeak-ng at runtime to phonemize
+# non-English text (the MIT-clean path; English uses a built-in G2P). Install
+# the espeak-ng runtime + its libpcaudio/libsonic deps + voice data so
+# package.sh can bundle them into the FROM scratch image.
+RUN if [ "${BACKEND}" = "crispasr" ]; then \
+    apt-get update && apt-get install -y --no-install-recommends \
+        espeak-ng-data libespeak-ng1 libpcaudio0 libsonic0 && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*; \
+fi
+
 COPY . /LocalAI
 
 RUN git config --global --add safe.directory /LocalAI

backend/Dockerfile.privacy-filter

@@ -0,0 +1,109 @@
+ARG BASE_IMAGE=ubuntu:24.04
+# BUILDER_BASE_IMAGE defaults to BASE_IMAGE so the Dockerfile parses when no
+# prebuilt base is supplied; the builder-prebuilt stage is only entered when
+# BUILDER_TARGET=builder-prebuilt, so the fallback content is harmless
+# (BuildKit prunes the unreferenced builder).
+ARG BUILDER_BASE_IMAGE=${BASE_IMAGE}
+# BUILDER_TARGET selects which builder stage the scratch image copies from.
+# Declared before any FROM so it is usable in `FROM ${BUILDER_TARGET}`. The
+# backend_build workflow sets it to builder-prebuilt when the matrix entry
+# provides builder-base-image, else builder-fromsource (the local default).
+ARG BUILDER_TARGET=builder-fromsource
+ARG APT_MIRROR=""
+ARG APT_PORTS_MIRROR=""
+
+# privacy-filter: standalone GGML engine for the openai-privacy-filter PII/NER
+# token classifier, wrapped as a LocalAI gRPC backend.
+#
+# Mirrors backend/Dockerfile.llama-cpp: the build toolchain (gRPC + cmake +
+# protoc + conditional CUDA/Vulkan) comes from the shared
+# .docker/install-base-deps.sh (from-source path) or a prebuilt
+# quay.io/go-skynet/ci-cache:base-grpc-* image (CI path) — nothing GPU-specific
+# is hand-rolled here. BUILD_TYPE selects the engine backend in the Makefile:
+# "" = cpu, "cublas" -> -DPF_CUDA=ON, "vulkan" -> -DPF_VULKAN=ON.
+
+# ============================================================================
+# Stage: builder-fromsource — self-contained build. Runs the same install
+# script backend/Dockerfile.base-grpc-builder runs, so this path is
+# bit-equivalent to the prebuilt base. Used when BUILDER_TARGET=builder-fromsource
+# (the default; local `make backends/privacy-filter`).
+# ============================================================================
+FROM ${BASE_IMAGE} AS builder-fromsource
+ARG BUILD_TYPE
+ARG CUDA_MAJOR_VERSION
+ARG CUDA_MINOR_VERSION
+ARG CMAKE_FROM_SOURCE=false
+# CUDA Toolkit 13.x needs CMake 3.31.9+ for correct toolchain/arch detection.
+ARG CMAKE_VERSION=3.31.10
+ARG GRPC_VERSION=v1.65.0
+ARG GRPC_MAKEFLAGS="-j4 -Otarget"
+ARG SKIP_DRIVERS=false
+ARG TARGETARCH
+ARG UBUNTU_VERSION=2404
+ARG APT_MIRROR
+ARG APT_PORTS_MIRROR
+
+ENV BUILD_TYPE=${BUILD_TYPE} \
+    CUDA_MAJOR_VERSION=${CUDA_MAJOR_VERSION} \
+    CUDA_MINOR_VERSION=${CUDA_MINOR_VERSION} \
+    CMAKE_FROM_SOURCE=${CMAKE_FROM_SOURCE} \
+    CMAKE_VERSION=${CMAKE_VERSION} \
+    GRPC_VERSION=${GRPC_VERSION} \
+    GRPC_MAKEFLAGS=${GRPC_MAKEFLAGS} \
+    SKIP_DRIVERS=${SKIP_DRIVERS} \
+    TARGETARCH=${TARGETARCH} \
+    UBUNTU_VERSION=${UBUNTU_VERSION} \
+    APT_MIRROR=${APT_MIRROR} \
+    APT_PORTS_MIRROR=${APT_PORTS_MIRROR} \
+    DEBIAN_FRONTEND=noninteractive
+# CUDA on PATH (a no-op when CUDA is not installed, e.g. cpu/vulkan builds).
+ENV PATH=/usr/local/cuda/bin:${PATH}
+
+WORKDIR /build
+
+# apt deps + cmake + protoc + gRPC + conditional CUDA/Vulkan, all from the
+# shared script (the source of truth that base-grpc-builder also runs).
+RUN --mount=type=bind,source=.docker/install-base-deps.sh,target=/usr/local/sbin/install-base-deps \
+    --mount=type=bind,source=.docker/apt-mirror.sh,target=/usr/local/sbin/apt-mirror \
+    bash /usr/local/sbin/install-base-deps
+
+# install-base-deps installs gRPC under /opt/grpc; copy it to /usr/local so the
+# backend's find_package(gRPC CONFIG) resolves it at the canonical prefix.
+RUN cp -a /opt/grpc/. /usr/local/
+
+COPY . /LocalAI
+
+RUN --mount=type=cache,target=/root/.ccache,id=privacy-filter-ccache-${TARGETARCH}-${BUILD_TYPE},sharing=locked \
+    make -C /LocalAI/backend/cpp/privacy-filter BUILD_TYPE=${BUILD_TYPE} NATIVE=false grpc-server package
+
+# ============================================================================
+# Stage: builder-prebuilt — FROM a prebuilt
+# quay.io/go-skynet/ci-cache:base-grpc-* image (gRPC at /opt/grpc + apt deps +
+# CUDA/Vulkan already installed). Used in CI when the matrix entry sets
+# builder-base-image.
+# ============================================================================
+FROM ${BUILDER_BASE_IMAGE} AS builder-prebuilt
+ARG BUILD_TYPE
+ARG TARGETARCH
+ENV BUILD_TYPE=${BUILD_TYPE}
+# CUDA on PATH (a no-op for the cpu/vulkan base images).
+ENV PATH=/usr/local/cuda/bin:${PATH}
+
+# Mirror builder-fromsource: the base-grpc image installs gRPC to /opt/grpc but
+# does not copy it to /usr/local.
+RUN cp -a /opt/grpc/. /usr/local/
+
+COPY . /LocalAI
+
+RUN --mount=type=cache,target=/root/.ccache,id=privacy-filter-ccache-${TARGETARCH}-${BUILD_TYPE},sharing=locked \
+    make -C /LocalAI/backend/cpp/privacy-filter BUILD_TYPE=${BUILD_TYPE} NATIVE=false grpc-server package
+
+# ============================================================================
+# Final stage — copy the package output from the selected builder. BuildKit
+# does not expand variables in `COPY --from=`, so alias the chosen builder to a
+# fixed stage name first.
+# ============================================================================
+FROM ${BUILDER_TARGET} AS builder
+
+FROM scratch
+COPY --from=builder /LocalAI/backend/cpp/privacy-filter/package/. ./

backend/Dockerfile.python

@@ -126,6 +126,7 @@ RUN <<EOT bash
         apt-get update && \
         apt-get install -y --no-install-recommends \
             cuda-nvcc-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} \
+            cuda-nvrtc-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} \
             libcufft-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} \
             libcurand-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} \
             libcublas-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} \

backend/backend.proto

@@ -24,6 +24,7 @@ service Backend {
   rpc TokenizeString(PredictOptions) returns (TokenizationResponse) {}
   rpc Status(HealthMessage) returns (StatusResponse) {}
   rpc Detect(DetectOptions) returns (DetectResponse) {}
+  rpc Depth(DepthRequest) returns (DepthResponse) {}
   rpc FaceVerify(FaceVerifyRequest) returns (FaceVerifyResponse) {}
   rpc FaceAnalyze(FaceAnalyzeRequest) returns (FaceAnalyzeResponse) {}
   rpc VoiceVerify(VoiceVerifyRequest) returns (VoiceVerifyResponse) {}
@@ -670,6 +671,35 @@ message DetectResponse {
   repeated Detection Detections = 1;
 }
 
+// --- Depth estimation messages (Depth Anything 3) ---
+
+message DepthRequest {
+  string src = 1;                  // input image (filesystem path or base64-encoded payload)
+  string dst = 2;                  // optional output directory for exports (glb/colmap)
+  bool include_depth = 3;          // return the per-pixel metric depth map
+  bool include_confidence = 4;     // return the per-pixel confidence map (DualDPT)
+  bool include_pose = 5;           // return camera extrinsics/intrinsics (DualDPT)
+  bool include_sky = 6;            // return the per-pixel sky map (mono models)
+  bool include_points = 7;         // back-project to a 3D point cloud (DualDPT)
+  float points_conf_thresh = 8;    // keep points with confidence >= this threshold
+  repeated string exports = 9;     // requested exports: "glb", "colmap"
+}
+
+message DepthResponse {
+  int32 width = 1;                 // processed depth-map width
+  int32 height = 2;                // processed depth-map height
+  repeated float depth = 3;        // width*height row-major metric depth
+  repeated float confidence = 4;   // width*height row-major confidence (DualDPT)
+  repeated float sky = 5;          // width*height row-major sky map (mono)
+  repeated float extrinsics = 6;   // 12 floats, 3x4 row-major (world-to-camera)
+  repeated float intrinsics = 7;   // 9 floats, 3x3 row-major
+  int32 num_points = 8;            // number of 3D points
+  repeated float points = 9;       // num_points*3 xyz, world space
+  bytes point_colors = 10;         // num_points*3 uint8 rgb
+  repeated string export_paths = 11; // paths written for the requested exports
+  bool is_metric = 12;             // depth is in metric units
+}
+
 // --- Face recognition messages ---
 
 message FacialArea {

backend/cpp/ds4/CMakeLists.txt

@@ -9,6 +9,22 @@ option(DS4_NATIVE "Compile with -march=native / -mcpu=native" ON)
 set(DS4_GPU "cpu" CACHE STRING "GPU backend: cpu, cuda, or metal")
 set(DS4_DIR "${CMAKE_CURRENT_SOURCE_DIR}/ds4" CACHE PATH "Path to cloned ds4 source")
 
+if(${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
+    # Homebrew installs protobuf/grpc under a non-default prefix. The generated
+    # backend.pb.cc / backend.grpc.pb.cc pull in google/protobuf and grpcpp
+    # headers, but the hw_grpc_proto library links neither target, so on macOS
+    # the headers (e.g. google/protobuf/runtime_version.h) are never on the
+    # compiler's include path. Add the Homebrew prefix globally, matching the
+    # llama-cpp backend which builds on Darwin CI.
+    if(CMAKE_HOST_SYSTEM_PROCESSOR MATCHES "arm64")
+        set(HOMEBREW_DEFAULT_PREFIX "/opt/homebrew")
+    else()
+        set(HOMEBREW_DEFAULT_PREFIX "/usr/local")
+    endif()
+    link_directories("${HOMEBREW_DEFAULT_PREFIX}/lib")
+    include_directories("${HOMEBREW_DEFAULT_PREFIX}/include")
+endif()
+
 find_package(Threads REQUIRED)
 find_package(Protobuf CONFIG QUIET)
 if(NOT Protobuf_FOUND)

backend/cpp/ds4/Makefile

@@ -1,10 +1,10 @@
 # ds4 backend Makefile.
 #
-# Upstream pin lives below as DS4_VERSION?=8384adf0f9fa0f3bb342dd925372de778b95b263
+# Upstream pin lives below as DS4_VERSION?=80ebbc396aee40eedc1d829222f3362d10fa4c6c
 # (.github/bump_deps.sh) can find and update it - matches the
 # llama-cpp / ik-llama-cpp / turboquant convention.
 
-DS4_VERSION?=8384adf0f9fa0f3bb342dd925372de778b95b263
+DS4_VERSION?=80ebbc396aee40eedc1d829222f3362d10fa4c6c
 DS4_REPO?=https://github.com/antirez/ds4
 
 CURRENT_MAKEFILE_DIR := $(dir $(abspath $(lastword $(MAKEFILE_LIST))))

backend/cpp/ds4/grpc-server.cpp

@@ -25,6 +25,8 @@ extern "C" {
 #include <chrono>
 #include <climits>
 #include <csignal>
+#include <cstddef>
+#include <cstdint>
 #include <cstdlib>
 #include <cstring>
 #include <ctime>
@@ -105,6 +107,130 @@ static bool parse_layers_spec(const std::string &spec, ds4_distributed_layers *o
     return true;
 }
 
+// Parse a boolean LoadModel option. An empty value (a bare flag-style option
+// like "ssd_streaming" with no colon) means true so model YAMLs can write
+// options: ["ssd_streaming"] to enable a switch.
+static bool parse_bool_option(const std::string &s, bool *out) {
+    if (s.empty() || s == "true" || s == "1" || s == "yes" || s == "on") { *out = true; return true; }
+    if (s == "false" || s == "0" || s == "no" || s == "off") { *out = false; return true; }
+    return false;
+}
+
+// Table-driven mapping from LoadModel option keys to ds4_engine_options fields.
+// ds4_engine_options is a fixed C struct with no reflection, so the field set
+// is enumerated once here; adding a future engine knob is a one-line table
+// entry rather than a new branch in LoadModel. Two fields need ds4's own typed
+// parsers (Gib, CacheExperts) so a plain string passthrough can't cover them.
+enum class DsOptType { Bool, Int, Uint, Float, Str, Gib, CacheExperts };
+
+struct DsOptSpec {
+    const char *key;
+    DsOptType   type;
+    size_t      off;      // byte offset into ds4_engine_options
+    size_t      off2;     // second offset (CacheExperts writes experts + bytes)
+    bool        is_path;  // Str values: resolve a relative value against the model dir
+};
+
+static const DsOptSpec kEngineOptSpecs[] = {
+    {"mtp_path",                      DsOptType::Str,          offsetof(ds4_engine_options, mtp_path),                      0, true},
+    {"mtp_draft",                     DsOptType::Int,          offsetof(ds4_engine_options, mtp_draft_tokens),              0},
+    {"mtp_margin",                    DsOptType::Float,        offsetof(ds4_engine_options, mtp_margin),                    0},
+    {"prefill_chunk",                 DsOptType::Uint,         offsetof(ds4_engine_options, prefill_chunk),                 0},
+    {"power_percent",                 DsOptType::Int,          offsetof(ds4_engine_options, power_percent),                 0},
+    {"warm_weights",                  DsOptType::Bool,         offsetof(ds4_engine_options, warm_weights),                  0},
+    {"quality",                       DsOptType::Bool,         offsetof(ds4_engine_options, quality),                       0},
+    {"ssd_streaming",                 DsOptType::Bool,         offsetof(ds4_engine_options, ssd_streaming),                 0},
+    {"ssd_streaming_cold",            DsOptType::Bool,         offsetof(ds4_engine_options, ssd_streaming_cold),            0},
+    {"ssd_streaming_preload_experts", DsOptType::Uint,         offsetof(ds4_engine_options, ssd_streaming_preload_experts), 0},
+    {"ssd_streaming_cache_experts",   DsOptType::CacheExperts, offsetof(ds4_engine_options, ssd_streaming_cache_experts),
+                                                               offsetof(ds4_engine_options, ssd_streaming_cache_bytes)},
+    {"simulate_used_memory",          DsOptType::Gib,          offsetof(ds4_engine_options, simulate_used_memory_bytes),    0},
+    {"expert_profile_path",           DsOptType::Str,          offsetof(ds4_engine_options, expert_profile_path),           0, true},
+    {"directional_steering_file",     DsOptType::Str,          offsetof(ds4_engine_options, directional_steering_file),     0, true},
+    {"directional_steering_attn",     DsOptType::Float,        offsetof(ds4_engine_options, directional_steering_attn),     0},
+    {"directional_steering_ffn",      DsOptType::Float,        offsetof(ds4_engine_options, directional_steering_ffn),      0},
+};
+
+// Apply a single key:value LoadModel option to the engine options struct.
+// Unknown keys are ignored (back-compat: callers pass mixed option sets).
+// String values are copied into `storage`, whose elements the engine reads by
+// pointer during ds4_engine_open; `storage` MUST have reserved capacity so
+// push_back never reallocates and dangles an earlier c_str(). Returns false
+// with `err` set when a recognized key has an invalid value.
+static bool apply_engine_option(ds4_engine_options *opt, const std::string &key,
+                                const std::string &val, const std::string &model_dir,
+                                std::vector<std::string> &storage, std::string &err) {
+    const DsOptSpec *spec = nullptr;
+    for (const auto &s : kEngineOptSpecs) {
+        if (key == s.key) { spec = &s; break; }
+    }
+    if (!spec) return true; // unknown key: ignore
+
+    char *base = reinterpret_cast<char *>(opt);
+    switch (spec->type) {
+    case DsOptType::Bool: {
+        bool b = false;
+        if (!parse_bool_option(val, &b)) { err = key + " must be true/false"; return false; }
+        *reinterpret_cast<bool *>(base + spec->off) = b;
+        return true;
+    }
+    case DsOptType::Int: {
+        char *end = nullptr;
+        long v = std::strtol(val.c_str(), &end, 10);
+        if (val.empty() || !end || *end != '\0') { err = key + " must be an integer"; return false; }
+        *reinterpret_cast<int *>(base + spec->off) = static_cast<int>(v);
+        return true;
+    }
+    case DsOptType::Uint: {
+        char *end = nullptr;
+        long v = std::strtol(val.c_str(), &end, 10);
+        if (val.empty() || !end || *end != '\0' || v < 0 || v > static_cast<long>(UINT32_MAX)) {
+            err = key + " must be a non-negative integer"; return false;
+        }
+        *reinterpret_cast<uint32_t *>(base + spec->off) = static_cast<uint32_t>(v);
+        return true;
+    }
+    case DsOptType::Float: {
+        char *end = nullptr;
+        float f = std::strtof(val.c_str(), &end);
+        if (val.empty() || !end || *end != '\0') { err = key + " must be a number"; return false; }
+        *reinterpret_cast<float *>(base + spec->off) = f;
+        return true;
+    }
+    case DsOptType::Str: {
+        // Resolve a relative path option (e.g. mtp_path: a sibling GGUF the
+        // gallery downloaded next to the model) against the model directory, so
+        // YAMLs reference companion files by name. Absolute values pass through.
+        if (spec->is_path && !model_dir.empty() && !val.empty() && val.front() != '/') {
+            storage.push_back(model_dir + "/" + val);
+        } else {
+            storage.push_back(val);
+        }
+        *reinterpret_cast<const char **>(base + spec->off) = storage.back().c_str();
+        return true;
+    }
+    case DsOptType::Gib: {
+        uint64_t bytes = 0;
+        if (!ds4_parse_gib_arg(val.c_str(), &bytes)) {
+            err = key + " must be a GiB value, e.g. 64GB"; return false;
+        }
+        *reinterpret_cast<uint64_t *>(base + spec->off) = bytes;
+        return true;
+    }
+    case DsOptType::CacheExperts: {
+        uint32_t experts = 0;
+        uint64_t bytes = 0;
+        if (!ds4_parse_streaming_cache_experts_arg(val.c_str(), &experts, &bytes)) {
+            err = key + " must be a positive expert count or a <number>GB budget"; return false;
+        }
+        *reinterpret_cast<uint32_t *>(base + spec->off)  = experts;
+        *reinterpret_cast<uint64_t *>(base + spec->off2) = bytes;
+        return true;
+    }
+    }
+    return true;
+}
+
 // When acting as a distributed coordinator, block until the worker route
 // covers all layers (ds4_session_distributed_route_ready == 1) or the timeout
 // elapses. Returns an empty string on success, or an error message to return
@@ -476,39 +602,10 @@ public:
             return GStatus::OK;
         }
 
-        std::string mtp_path;
-        int mtp_draft = 0;
-        float mtp_margin = 3.0f;
-        std::string ds4_role, ds4_layers, ds4_listen;
-        for (const auto &opt : request->options()) {
-            auto [k, v] = split_option(opt);
-            if (k == "mtp_path") mtp_path = v;
-            else if (k == "mtp_draft") mtp_draft = std::stoi(v);
-            else if (k == "mtp_margin") mtp_margin = std::stof(v);
-            else if (k == "kv_cache_dir") g_kv_cache_dir = v;
-            else if (k == "ds4_role") ds4_role = v;
-            else if (k == "ds4_layers") ds4_layers = v;
-            else if (k == "ds4_listen") ds4_listen = v;
-            else if (k == "ds4_route_timeout") {
-                if (!parse_positive_int(v, &g_route_timeout_sec)) {
-                    result->set_success(false);
-                    result->set_message("ds4: ds4_route_timeout must be a positive integer");
-                    return GStatus::OK;
-                }
-            }
-        }
-
-        g_kv_cache.SetDir(g_kv_cache_dir);
-
         ds4_engine_options opt = {};
         opt.model_path = model_path.c_str();
-        opt.mtp_path = mtp_path.empty() ? nullptr : mtp_path.c_str();
         opt.n_threads = request->threads() > 0 ? request->threads() : 0;
-        opt.mtp_draft_tokens = mtp_draft;
-        opt.mtp_margin = mtp_margin;
-        opt.directional_steering_file = nullptr;
-        opt.warm_weights = false;
-        opt.quality = false;
+        opt.mtp_margin = 3.0f; // ds4 default; overridable via the mtp_margin option
 
 #if defined(DS4_NO_GPU)
         opt.backend = DS4_BACKEND_CPU;
@@ -518,6 +615,46 @@ public:
         opt.backend = DS4_BACKEND_CUDA;
 #endif
 
+        // Stable storage for string-valued engine options. The engine reads
+        // these by pointer during ds4_engine_open, so the std::string backing
+        // store must outlive the call and not reallocate; reserve up front so
+        // push_back keeps every prior c_str() valid. Static + clear() reuses
+        // the buffer across LoadModel calls (the old engine is closed above).
+        static std::vector<std::string> s_opt_strings;
+        s_opt_strings.clear();
+        s_opt_strings.reserve(sizeof(kEngineOptSpecs) / sizeof(kEngineOptSpecs[0]));
+
+        // Directory of the main model, used to resolve relative path options.
+        std::string model_dir;
+        if (auto slash = model_path.find_last_of('/'); slash != std::string::npos) {
+            model_dir = model_path.substr(0, slash);
+        }
+
+        std::string ds4_role, ds4_layers, ds4_listen;
+        for (const auto &o : request->options()) {
+            auto [k, v] = split_option(o);
+            if (k == "kv_cache_dir") { g_kv_cache_dir = v; continue; }
+            else if (k == "ds4_role") { ds4_role = v; continue; }
+            else if (k == "ds4_layers") { ds4_layers = v; continue; }
+            else if (k == "ds4_listen") { ds4_listen = v; continue; }
+            else if (k == "ds4_route_timeout") {
+                if (!parse_positive_int(v, &g_route_timeout_sec)) {
+                    result->set_success(false);
+                    result->set_message("ds4: ds4_route_timeout must be a positive integer");
+                    return GStatus::OK;
+                }
+                continue;
+            }
+            std::string err;
+            if (!apply_engine_option(&opt, k, v, model_dir, s_opt_strings, err)) {
+                result->set_success(false);
+                result->set_message("ds4: " + err);
+                return GStatus::OK;
+            }
+        }
+
+        g_kv_cache.SetDir(g_kv_cache_dir);
+
         // Coordinator wiring. 'ds4_role:coordinator' enables layer-split
         // distributed inference: this process listens on ds4_listen and owns
         // the ds4_layers slice; workers dial in (see `local-ai worker

backend/cpp/ik-llama-cpp/Makefile

@@ -1,5 +1,5 @@
 
-IK_LLAMA_VERSION?=e6f8112f3ba126eed3ff5b30cdd08085414a7516
+IK_LLAMA_VERSION?=71af16a6b7f6fb7315b346b4a51aad530599c3f5
 LLAMA_REPO?=https://github.com/ikawrakow/ik_llama.cpp
 
 CMAKE_ARGS?=

backend/cpp/llama-cpp/Makefile

@@ -1,5 +1,5 @@
 
-LLAMA_VERSION?=039e20a2db9e87b2477c76cc04905f3e1acad77f
+LLAMA_VERSION?=f3e182816421c648188b5eab269853bf1531d950
 LLAMA_REPO?=https://github.com/ggerganov/llama.cpp
 
 CMAKE_ARGS?=

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

@@ -1922,25 +1922,27 @@ public:
                     body_json["min_p"] = data["min_p"];
                 }
 
-                // Pass enable_thinking via chat_template_kwargs (where oaicompat_chat_params_parse reads it)
+                // Forward the chat_template_kwargs the Go layer resolved (model config
+                // chat_template_kwargs + per-request metadata: enable_thinking,
+                // reasoning_effort, preserve_thinking, ...). One generic merge replaces
+                // the previous per-key handling - new template levers need no C++ change.
+                // oaicompat_chat_params_parse reads these from body_json.
                 const auto& metadata = request->metadata();
-                auto et_it = metadata.find("enable_thinking");
-                if (et_it != metadata.end()) {
-                    if (!body_json.contains("chat_template_kwargs")) {
-                        body_json["chat_template_kwargs"] = json::object();
+                auto ctk_it = metadata.find("chat_template_kwargs");
+                if (ctk_it != metadata.end() && !ctk_it->second.empty()) {
+                    try {
+                        json ctk = json::parse(ctk_it->second);
+                        if (ctk.is_object()) {
+                            if (!body_json.contains("chat_template_kwargs")) {
+                                body_json["chat_template_kwargs"] = json::object();
+                            }
+                            for (auto& el : ctk.items()) {
+                                body_json["chat_template_kwargs"][el.key()] = el.value();
+                            }
+                        }
+                    } catch (const std::exception & e) {
+                        SRV_WRN("failed to parse chat_template_kwargs metadata: %s\n", e.what());
                     }
-                    body_json["chat_template_kwargs"]["enable_thinking"] = (et_it->second == "true");
-                }
-
-                // Pass reasoning_effort via chat_template_kwargs too: the lever
-                // jinja templates like gpt-oss (Harmony) / LFM2.5 read, distinct
-                // from enable_thinking which those templates ignore.
-                auto re_it = metadata.find("reasoning_effort");
-                if (re_it != metadata.end() && !re_it->second.empty()) {
-                    if (!body_json.contains("chat_template_kwargs")) {
-                        body_json["chat_template_kwargs"] = json::object();
-                    }
-                    body_json["chat_template_kwargs"]["reasoning_effort"] = re_it->second;
                 }
 
                 // Debug: Print full body_json before template processing (includes messages, tools, tool_choice, etc.)
@@ -2756,25 +2758,26 @@ public:
                     body_json["min_p"] = data["min_p"];
                 }
 
-                // Pass enable_thinking via chat_template_kwargs (where oaicompat_chat_params_parse reads it)
+                // Forward the chat_template_kwargs the Go layer resolved (model config
+                // chat_template_kwargs + per-request metadata: enable_thinking,
+                // reasoning_effort, preserve_thinking, ...). One generic merge replaces
+                // the previous per-key handling - new template levers need no C++ change.
                 const auto& predict_metadata = request->metadata();
-                auto predict_et_it = predict_metadata.find("enable_thinking");
-                if (predict_et_it != predict_metadata.end()) {
-                    if (!body_json.contains("chat_template_kwargs")) {
-                        body_json["chat_template_kwargs"] = json::object();
+                auto predict_ctk_it = predict_metadata.find("chat_template_kwargs");
+                if (predict_ctk_it != predict_metadata.end() && !predict_ctk_it->second.empty()) {
+                    try {
+                        json ctk = json::parse(predict_ctk_it->second);
+                        if (ctk.is_object()) {
+                            if (!body_json.contains("chat_template_kwargs")) {
+                                body_json["chat_template_kwargs"] = json::object();
+                            }
+                            for (auto& el : ctk.items()) {
+                                body_json["chat_template_kwargs"][el.key()] = el.value();
+                            }
+                        }
+                    } catch (const std::exception & e) {
+                        SRV_WRN("failed to parse chat_template_kwargs metadata: %s\n", e.what());
                     }
-                    body_json["chat_template_kwargs"]["enable_thinking"] = (predict_et_it->second == "true");
-                }
-
-                // Pass reasoning_effort via chat_template_kwargs too: the lever
-                // jinja templates like gpt-oss (Harmony) / LFM2.5 read, distinct
-                // from enable_thinking which those templates ignore.
-                auto predict_re_it = predict_metadata.find("reasoning_effort");
-                if (predict_re_it != predict_metadata.end() && !predict_re_it->second.empty()) {
-                    if (!body_json.contains("chat_template_kwargs")) {
-                        body_json["chat_template_kwargs"] = json::object();
-                    }
-                    body_json["chat_template_kwargs"]["reasoning_effort"] = predict_re_it->second;
                 }
 
                 // Debug: Print full body_json before template processing (includes messages, tools, tool_choice, etc.)
@@ -3486,7 +3489,7 @@ public:
         if (body.count("prompt") != 0) {
             const bool add_special = json_value(body, "add_special", false);
 
-            llama_tokens tokens = tokenize_mixed(ctx_server.impl->vocab, body.at("content"), add_special, true);
+            llama_tokens tokens = tokenize_mixed(ctx_server.impl->vocab, body.at("prompt"), add_special, true);
 
 
             for (const auto& token : tokens) {

backend/cpp/privacy-filter/.gitignore

@@ -0,0 +1,9 @@
+/privacy-filter.cpp
+build/
+package/
+grpc-server
+*.o
+backend.pb.cc
+backend.pb.h
+backend.grpc.pb.cc
+backend.grpc.pb.h

backend/cpp/privacy-filter/CMakeLists.txt

@@ -0,0 +1,69 @@
+cmake_minimum_required(VERSION 3.21)
+project(privacy-filter-grpc-server LANGUAGES CXX C)
+
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+set(TARGET grpc-server)
+
+# Path to the privacy-filter.cpp engine sources. The Makefile arranges for this
+# to exist (clone of a pinned commit, or a symlink to PRIVACY_FILTER_SRC).
+set(PRIVACY_FILTER_DIR "${CMAKE_CURRENT_SOURCE_DIR}/privacy-filter.cpp"
+    CACHE PATH "Path to the privacy-filter.cpp engine source tree")
+
+find_package(Threads REQUIRED)
+find_package(Protobuf CONFIG QUIET)
+if(NOT Protobuf_FOUND)
+    find_package(Protobuf REQUIRED)
+endif()
+find_package(gRPC CONFIG QUIET)
+if(NOT gRPC_FOUND)
+    # Ubuntu's apt-installed grpc++ does not ship a CMake config - fall back.
+    find_library(GRPCPP_LIB grpc++ REQUIRED)
+    find_library(GRPCPP_REFLECTION_LIB grpc++_reflection REQUIRED)
+    add_library(gRPC::grpc++ INTERFACE IMPORTED)
+    set_target_properties(gRPC::grpc++ PROPERTIES INTERFACE_LINK_LIBRARIES "${GRPCPP_LIB}")
+    add_library(gRPC::grpc++_reflection INTERFACE IMPORTED)
+    set_target_properties(gRPC::grpc++_reflection PROPERTIES INTERFACE_LINK_LIBRARIES "${GRPCPP_REFLECTION_LIB}")
+endif()
+
+find_program(_PROTOC NAMES protoc REQUIRED)
+find_program(_GRPC_CPP_PLUGIN NAMES grpc_cpp_plugin REQUIRED)
+
+get_filename_component(HW_PROTO "${CMAKE_CURRENT_SOURCE_DIR}/../../backend.proto" ABSOLUTE)
+get_filename_component(HW_PROTO_PATH "${HW_PROTO}" PATH)
+
+set(HW_PROTO_SRCS "${CMAKE_CURRENT_BINARY_DIR}/backend.pb.cc")
+set(HW_PROTO_HDRS "${CMAKE_CURRENT_BINARY_DIR}/backend.pb.h")
+set(HW_GRPC_SRCS  "${CMAKE_CURRENT_BINARY_DIR}/backend.grpc.pb.cc")
+set(HW_GRPC_HDRS  "${CMAKE_CURRENT_BINARY_DIR}/backend.grpc.pb.h")
+
+add_custom_command(
+    OUTPUT "${HW_PROTO_SRCS}" "${HW_PROTO_HDRS}" "${HW_GRPC_SRCS}" "${HW_GRPC_HDRS}"
+    COMMAND ${_PROTOC}
+    ARGS --grpc_out "${CMAKE_CURRENT_BINARY_DIR}"
+         --cpp_out  "${CMAKE_CURRENT_BINARY_DIR}"
+         -I "${HW_PROTO_PATH}"
+         --plugin=protoc-gen-grpc="${_GRPC_CPP_PLUGIN}"
+         "${HW_PROTO}"
+    DEPENDS "${HW_PROTO}")
+
+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 ${CMAKE_CURRENT_BINARY_DIR})
+
+# Build only the pf static lib (+ ggml) from the engine tree — no CLI/bench/tests.
+# PF_VULKAN is honored when passed on the cmake command line (it lands in the
+# shared cache the engine reads).
+set(PF_BUILD_TOOLS OFF CACHE BOOL "" FORCE)
+set(PF_BUILD_TESTS OFF CACHE BOOL "" FORCE)
+add_subdirectory(${PRIVACY_FILTER_DIR} ${CMAKE_CURRENT_BINARY_DIR}/privacy-filter.cpp)
+
+add_executable(${TARGET} grpc-server.cpp)
+target_link_libraries(${TARGET} PRIVATE
+    pf
+    hw_grpc_proto
+    gRPC::grpc++
+    gRPC::grpc++_reflection
+    protobuf::libprotobuf
+    Threads::Threads)

backend/cpp/privacy-filter/Makefile

@@ -0,0 +1,77 @@
+# privacy-filter backend Makefile.
+#
+# Wraps the standalone privacy-filter.cpp GGML engine (the openai-privacy-filter
+# PII/NER token classifier) as a LocalAI gRPC backend. The engine source is
+# fetched at the pin below — .github/workflows/bump_deps.yaml finds and updates
+# PRIVACY_FILTER_VERSION, matching the llama-cpp / ds4 convention.
+#
+# Local development: point at a working checkout instead of cloning, e.g.
+#   make PRIVACY_FILTER_SRC=$HOME/c/privacy-filter.cpp grpc-server
+
+PRIVACY_FILTER_VERSION?=646342f7a59c6b7d195185eac60bad762e572f1d
+PRIVACY_FILTER_REPO?=https://github.com/localai-org/privacy-filter.cpp
+PRIVACY_FILTER_SRC?=
+
+CURRENT_MAKEFILE_DIR := $(dir $(abspath $(lastword $(MAKEFILE_LIST))))
+BUILD_DIR := build
+
+BUILD_TYPE ?=
+NATIVE ?= false
+JOBS ?= $(shell nproc 2>/dev/null || echo 4)
+
+CMAKE_ARGS ?= -DCMAKE_BUILD_TYPE=Release
+
+# GPU backends; the default (cpu) needs no extra flags. 'cublas' is LocalAI's
+# name for the CUDA build (matches llama-cpp / ds4), mapping to the engine's
+# GGML_CUDA path; 'vulkan' selects the ggml Vulkan backend.
+ifeq ($(BUILD_TYPE),cublas)
+    CMAKE_ARGS += -DPF_CUDA=ON
+endif
+ifeq ($(BUILD_TYPE),vulkan)
+    CMAKE_ARGS += -DPF_VULKAN=ON
+endif
+
+# Portable binaries for distribution: disable -march=native unless asked.
+ifneq ($(NATIVE),true)
+    CMAKE_ARGS += -DGGML_NATIVE=OFF
+endif
+
+.PHONY: grpc-server package clean purge test all
+all: grpc-server
+
+# Provide the engine sources at ./privacy-filter.cpp. With PRIVACY_FILTER_SRC
+# set we symlink a local checkout (instant, no network); otherwise we clone the
+# pinned commit and its ggml submodule. The directory/symlink is the target, so
+# make only does this once — run 'make purge && make' to refetch after a bump.
+privacy-filter.cpp:
+ifneq ($(PRIVACY_FILTER_SRC),)
+	ln -sfn $(abspath $(PRIVACY_FILTER_SRC)) privacy-filter.cpp
+else
+	mkdir -p privacy-filter.cpp
+	cd privacy-filter.cpp && \
+	git init -q && \
+	git remote add origin $(PRIVACY_FILTER_REPO) && \
+	git fetch --depth 1 origin $(PRIVACY_FILTER_VERSION) && \
+	git checkout FETCH_HEAD && \
+	git submodule update --init --recursive --depth 1
+endif
+
+grpc-server: privacy-filter.cpp
+	@echo "Building privacy-filter grpc-server ($(BUILD_TYPE)) with $(CMAKE_ARGS)"
+	mkdir -p $(BUILD_DIR)
+	cd $(BUILD_DIR) && cmake $(CMAKE_ARGS) $(CURRENT_MAKEFILE_DIR) && cmake --build . --config Release -j $(JOBS)
+	cp $(BUILD_DIR)/grpc-server grpc-server
+
+package: grpc-server
+	bash package.sh
+
+test:
+	@echo "privacy-filter backend: parity/regression coverage lives in the engine repo"
+
+clean:
+	rm -rf $(BUILD_DIR) grpc-server package
+
+# 'privacy-filter.cpp' may be a symlink (PRIVACY_FILTER_SRC) — rm without a
+# trailing slash removes the link, never the linked-to checkout.
+purge: clean
+	rm -rf privacy-filter.cpp

backend/cpp/privacy-filter/grpc-server.cpp

@@ -0,0 +1,210 @@
+// privacy-filter LocalAI gRPC backend.
+//
+// Thin shim over privacy-filter.cpp's flat C API (include/pf.h): a standalone
+// GGML engine for the openai-privacy-filter token-classification model family
+// (PII NER). It replaces the llama.cpp-patched TokenClassify path for this one
+// model family — same GGUF files, no llama.cpp carry-patches.
+//
+// Only the RPCs the PII tier needs are implemented: LoadModel, TokenClassify,
+// plus Health / Status / Free. Everything else inherits the generated base
+// class default (UNIMPLEMENTED).
+
+#include "backend.pb.h"
+#include "backend.grpc.pb.h"
+
+#include "pf.h"
+
+#include <grpcpp/grpcpp.h>
+#include <grpcpp/server.h>
+#include <grpcpp/server_builder.h>
+#include <grpcpp/ext/proto_server_reflection_plugin.h>
+
+#include <atomic>
+#include <chrono>
+#include <csignal>
+#include <iostream>
+#include <memory>
+#include <mutex>
+#include <string>
+
+using grpc::Server;
+using grpc::ServerBuilder;
+using grpc::ServerContext;
+// NOTE: do NOT alias grpc::Status as Status — the Status RPC method below would
+// shadow the type and break the other method signatures. Use GStatus instead.
+using GStatus = ::grpc::Status;
+using grpc::StatusCode;
+
+namespace {
+
+// The engine is single-model-per-process: LocalAI spawns one backend process
+// per loaded model. g_mu guards (re)load against in-flight classification.
+std::mutex          g_mu;
+pf_ctx *            g_ctx = nullptr;
+std::atomic<Server *> g_server{nullptr};
+
+// Resolve the device string the engine expects ("cpu" / "gpu" / "cuda" /
+// "vulkan", optionally ":N"). Priority: an explicit "device:..." in
+// ModelOptions.Options, then a non-zero NGPULayers as a coarse "use the GPU"
+// signal, else CPU. "gpu" lets the engine pick whichever GPU backend this
+// binary was compiled with (CUDA or Vulkan), so the same config works on
+// either build; pin "device:cuda"/"device:vulkan" to be explicit.
+std::string resolve_device(const backend::ModelOptions * opts) {
+    for (const auto & o : opts->options()) {
+        const std::string prefix = "device:";
+        if (o.rfind(prefix, 0) == 0) {
+            return o.substr(prefix.size());
+        }
+    }
+    if (opts->ngpulayers() > 0) {
+        return "gpu";
+    }
+    return "cpu";
+}
+
+class PrivacyFilterBackend final : public backend::Backend::Service {
+public:
+    GStatus Health(ServerContext *, const backend::HealthMessage *,
+                   backend::Reply * reply) override {
+        reply->set_message("OK");
+        return GStatus::OK;
+    }
+
+    GStatus Status(ServerContext *, const backend::HealthMessage *,
+                   backend::StatusResponse * response) override {
+        std::lock_guard<std::mutex> lock(g_mu);
+        response->set_state(g_ctx ? backend::StatusResponse::READY
+                                  : backend::StatusResponse::UNINITIALIZED);
+        return GStatus::OK;
+    }
+
+    GStatus LoadModel(ServerContext *, const backend::ModelOptions * request,
+                      backend::Result * result) override {
+        std::lock_guard<std::mutex> lock(g_mu);
+
+        // ModelFile is the absolute path LocalAI resolves; Model is the bare
+        // name. Prefer the former, fall back to the latter.
+        const std::string path =
+            !request->modelfile().empty() ? request->modelfile() : request->model();
+        if (path.empty()) {
+            result->set_success(false);
+            result->set_message("no model path supplied");
+            return GStatus::OK;
+        }
+
+        const std::string device = resolve_device(request);
+
+        if (g_ctx) { pf_free(g_ctx); g_ctx = nullptr; }
+
+        pf_ctx * ctx = pf_load(path.c_str(), device.c_str(), request->threads());
+        const char * err = pf_last_error(ctx);
+        if (err) {
+            result->set_success(false);
+            result->set_message(std::string("privacy-filter load failed: ") + err);
+            pf_free(ctx);
+            return GStatus::OK;
+        }
+
+        // ContextSize, when set, becomes the per-forward window. The engine
+        // ignores values that are too small to window (<= 2*halo) and just
+        // runs a single forward, so passing it through is always safe.
+        if (request->contextsize() > 0) {
+            pf_set_window(ctx, request->contextsize());
+        }
+
+        g_ctx = ctx;
+        result->set_success(true);
+        result->set_message("privacy-filter loaded (" + device + ")");
+        return GStatus::OK;
+    }
+
+    GStatus TokenClassify(ServerContext *, const backend::TokenClassifyRequest * request,
+                          backend::TokenClassifyResponse * response) override {
+        std::lock_guard<std::mutex> lock(g_mu);
+        if (!g_ctx) {
+            return GStatus(StatusCode::FAILED_PRECONDITION, "Model not loaded");
+        }
+
+        const std::string & text = request->text();
+        if (text.empty()) {
+            return GStatus::OK;  // no text -> no entities
+        }
+
+        pf_entity * ents = nullptr;
+        size_t      n    = 0;
+        if (pf_classify(g_ctx, text.data(), text.size(), request->threshold(), &ents, &n) != 0) {
+            const char * err = pf_last_error(g_ctx);
+            return GStatus(StatusCode::INTERNAL,
+                           std::string("TokenClassify failed: ") + (err ? err : "unknown"));
+        }
+
+        // Byte offsets are into the original UTF-8 text; the engine already
+        // applied the threshold and whitespace-trimmed span edges.
+        for (size_t i = 0; i < n; i++) {
+            backend::TokenClassifyEntity * ent = response->add_entities();
+            ent->set_entity_group(ents[i].label ? ents[i].label : "");
+            ent->set_start(ents[i].start);
+            ent->set_end(ents[i].end);
+            ent->set_score(ents[i].score);
+            ent->set_text(text.substr((size_t) ents[i].start,
+                                      (size_t) (ents[i].end - ents[i].start)));
+        }
+        pf_entities_free(ents, n);
+        return GStatus::OK;
+    }
+
+    GStatus Free(ServerContext *, const backend::HealthMessage *,
+                 backend::Result * result) override {
+        std::lock_guard<std::mutex> lock(g_mu);
+        if (g_ctx) { pf_free(g_ctx); g_ctx = nullptr; }
+        result->set_success(true);
+        return GStatus::OK;
+    }
+};
+
+void RunServer(const std::string & addr) {
+    PrivacyFilterBackend service;
+    grpc::EnableDefaultHealthCheckService(true);
+    grpc::reflection::InitProtoReflectionServerBuilderPlugin();
+
+    ServerBuilder builder;
+    builder.AddListeningPort(addr, grpc::InsecureServerCredentials());
+    builder.RegisterService(&service);
+    builder.SetMaxReceiveMessageSize(64 * 1024 * 1024);
+    builder.SetMaxSendMessageSize(64 * 1024 * 1024);
+
+    std::unique_ptr<Server> server(builder.BuildAndStart());
+    if (!server) {
+        std::cerr << "privacy-filter grpc-server: failed to bind " << addr << "\n";
+        std::exit(1);
+    }
+    g_server = server.get();
+    std::cerr << "privacy-filter grpc-server listening on " << addr << "\n";
+    server->Wait();
+}
+
+void signal_handler(int) {
+    if (auto * srv = g_server.load()) {
+        srv->Shutdown(std::chrono::system_clock::now() + std::chrono::seconds(3));
+    }
+}
+
+} // namespace
+
+int main(int argc, char * argv[]) {
+    std::string addr = "127.0.0.1:50051";
+    for (int i = 1; i < argc; ++i) {
+        std::string a = argv[i];
+        const std::string addr_flag = "--addr=";
+        if (a.rfind(addr_flag, 0) == 0)      addr = a.substr(addr_flag.size());
+        else if (a == "--addr" && i + 1 < argc) addr = argv[++i];
+        else if (a == "--help" || a == "-h") {
+            std::cout << "Usage: grpc-server --addr=HOST:PORT\n";
+            return 0;
+        }
+    }
+    std::signal(SIGINT,  signal_handler);
+    std::signal(SIGTERM, signal_handler);
+    RunServer(addr);
+    return 0;
+}

backend/cpp/privacy-filter/package.sh

@@ -0,0 +1,39 @@
+#!/bin/bash
+# Assemble package/ for the from-scratch backend image: the grpc-server binary,
+# run.sh, the dynamic loader, and every shared library the binary needs.
+set -e
+CURDIR=$(dirname "$(realpath "$0")")
+REPO_ROOT="${CURDIR}/../../.."
+
+mkdir -p "$CURDIR/package/lib"
+cp -avf "$CURDIR/grpc-server" "$CURDIR/package/"
+cp -rfv "$CURDIR/run.sh"      "$CURDIR/package/"
+
+# The dynamic loader, renamed to lib/ld.so so run.sh can invoke it explicitly
+# (makes the image independent of the host's glibc layout).
+if [ -f "/lib64/ld-linux-x86-64.so.2" ]; then
+    cp -arfLv /lib64/ld-linux-x86-64.so.2 "$CURDIR/package/lib/ld.so"
+elif [ -f "/lib/ld-linux-aarch64.so.1" ]; then
+    cp -arfLv /lib/ld-linux-aarch64.so.1 "$CURDIR/package/lib/ld.so"
+else
+    echo "package.sh: unknown architecture" >&2; exit 1
+fi
+
+# Bundle the binary's transitive shared deps (libstdc++, libgomp, and the apt
+# grpc++/protobuf/absl stack) by walking ldd — robust to whichever of those are
+# linked shared vs static. The loader line (no "=>") is skipped; ld.so above
+# already covers it.
+ldd "$CURDIR/grpc-server" | awk '$2 == "=>" && $3 ~ /^\// { print $3 }' | sort -u | \
+while read -r so; do
+    [ -f "$so" ] && cp -arfLv "$so" "$CURDIR/package/lib/"
+done
+
+# Vulkan loader / GPU libs when building the GPU variant.
+GPU_LIB_SCRIPT="${REPO_ROOT}/scripts/build/package-gpu-libs.sh"
+if [ -f "$GPU_LIB_SCRIPT" ]; then
+    source "$GPU_LIB_SCRIPT" "$CURDIR/package/lib"
+    package_gpu_libs
+fi
+
+echo "privacy-filter package contents:"
+ls -lah "$CURDIR/package/" "$CURDIR/package/lib/"

backend/cpp/privacy-filter/run.sh

@@ -0,0 +1,9 @@
+#!/bin/bash
+# Entry point for the privacy-filter backend image / BACKEND_BINARY mode.
+set -e
+CURDIR=$(dirname "$(realpath "$0")")
+export LD_LIBRARY_PATH="$CURDIR/lib:$LD_LIBRARY_PATH"
+if [ -f "$CURDIR/lib/ld.so" ]; then
+    exec "$CURDIR/lib/ld.so" "$CURDIR/grpc-server" "$@"
+fi
+exec "$CURDIR/grpc-server" "$@"

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?=c29f6653a516a3001d923944dad8892072cc7334
+CRISPASR_VERSION?=d745bda4386ae0f9d1d2f23fff8ec95d76428221
 SO_TARGET?=libgocrispasr.so
 
 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF

backend/go/crispasr/gocrispasr.go

@@ -11,6 +11,7 @@ import (
 
 	"github.com/go-audio/audio"
 	"github.com/go-audio/wav"
+	gguf "github.com/gpustack/gguf-parser-go"
 	"github.com/mudler/LocalAI/pkg/grpc/base"
 	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
 	"github.com/mudler/LocalAI/pkg/utils"
@@ -37,6 +38,39 @@ var (
 
 type CrispASR struct {
 	base.SingleThread
+	// sampleRate is the output rate (Hz) of the loaded TTS engine's PCM, used to
+	// write a correct WAV header. Most CrispASR TTS backends emit 24 kHz, but
+	// piper returns its model's native rate (16 kHz for x_low/low voices,
+	// 22.05 kHz for medium/high), so it is read from the GGUF metadata at Load.
+	sampleRate int
+}
+
+// defaultTTSSampleRate is the output rate assumed for CrispASR TTS engines that
+// don't advertise one in GGUF metadata (vibevoice/orpheus/chatterbox/qwen3-tts
+// all emit 24 kHz). piper is the exception and carries piper.sample_rate.
+const defaultTTSSampleRate = 24000
+
+// piperSampleRate reads the piper.sample_rate metadata key from a GGUF model.
+// CrispASR's piper backend returns PCM at the model's native rate without
+// resampling, so the WAV header must match it. Returns ok=false for non-piper
+// models (key absent) or an unreadable file, letting the caller fall back to
+// defaultTTSSampleRate.
+func piperSampleRate(modelPath string) (int, bool) {
+	// Only scalar architecture keys are read, so skip the large array metadata
+	// (phoneme map) and mmap the header - same rationale as pkg/vram's reader.
+	f, err := gguf.ParseGGUFFile(modelPath, gguf.UseMMap(), gguf.SkipLargeMetadata())
+	if err != nil {
+		return 0, false
+	}
+	kv, ok := f.Header.MetadataKV.Get("piper.sample_rate")
+	if !ok || kv.ValueType != gguf.GGUFMetadataValueTypeUint32 {
+		return 0, false
+	}
+	rate := int(kv.ValueUint32())
+	if rate <= 0 {
+		return 0, false
+	}
+	return rate, true
 }
 
 // splitOption splits a "prefix:value" model option into its key and value,
@@ -103,6 +137,14 @@ func (w *CrispASR) Load(opts *pb.ModelOptions) error {
 		return fmt.Errorf("Failed to load CrispASR transcription model")
 	}
 
+	// Determine the TTS output sample rate for the WAV header. piper voices
+	// carry their native rate in GGUF metadata and CrispASR does not resample;
+	// every other engine emits the 24 kHz default.
+	w.sampleRate = defaultTTSSampleRate
+	if rate, ok := piperSampleRate(opts.ModelFile); ok {
+		w.sampleRate = rate
+	}
+
 	// Load the companion file (codec/tokenizer/s3gen) after the session is open.
 	// rc==0 means success or "not applicable" for the active backend; only a
 	// negative code is fatal.
@@ -390,7 +432,7 @@ func (w *CrispASR) synthesize(text string) ([]float32, error) {
 	}
 	defer CppTTSFree(ptr)
 	src := unsafe.Slice((*float32)(unsafe.Pointer(ptr)), int(n)) //nolint:govet // ptr addresses C-allocated PCM returned across the purego boundary; copied out immediately below, before tts_free.
-	out := make([]float32, int(n)) // copy out of C memory before free
+	out := make([]float32, int(n))                               // copy out of C memory before free
 	copy(out, src)
 	return out, nil
 }
@@ -417,7 +459,7 @@ func (w *CrispASR) TTS(req *pb.TTSRequest) error {
 	if err != nil {
 		return err
 	}
-	return writeWAV24k(req.Dst, pcm)
+	return writeWAV(req.Dst, pcm, w.sampleRate)
 }
 
 // TTSStream is the streaming counterpart to TTS. CrispASR has no progressive
@@ -447,7 +489,7 @@ func (w *CrispASR) TTSStream(req *pb.TTSRequest, results chan []byte) error {
 	}
 	defer func() { _ = os.Remove(dst) }()
 
-	if err := writeWAV24k(dst, pcm); err != nil {
+	if err := writeWAV(dst, pcm, w.sampleRate); err != nil {
 		return err
 	}
 
@@ -459,14 +501,14 @@ func (w *CrispASR) TTSStream(req *pb.TTSRequest, results chan []byte) error {
 	return nil
 }
 
-// writeWAV24k writes pcm as a 24000 Hz, mono, 16-bit PCM WAV at dst.
-func writeWAV24k(dst string, pcm []float32) error {
+// writeWAV writes pcm as a sampleRate Hz, mono, 16-bit PCM WAV at dst.
+func writeWAV(dst string, pcm []float32, sampleRate int) error {
 	f, err := os.Create(dst)
 	if err != nil {
 		return fmt.Errorf("crispasr: create %q: %w", dst, err)
 	}
 
-	enc := wav.NewEncoder(f, 24000, 16, 1, 1)
+	enc := wav.NewEncoder(f, sampleRate, 16, 1, 1)
 	ints := make([]int, len(pcm))
 	for i, s := range pcm {
 		if s > 1 {
@@ -477,7 +519,7 @@ func writeWAV24k(dst string, pcm []float32) error {
 		ints[i] = int(s * 32767)
 	}
 	buf := &audio.IntBuffer{
-		Format:         &audio.Format{NumChannels: 1, SampleRate: 24000},
+		Format:         &audio.Format{NumChannels: 1, SampleRate: sampleRate},
 		Data:           ints,
 		SourceBitDepth: 16,
 	}

backend/go/crispasr/gocrispasr_samplerate_test.go

@@ -0,0 +1,164 @@
+package main
+
+import (
+	"bytes"
+	"encoding/binary"
+	"os"
+	"path/filepath"
+
+	"github.com/go-audio/wav"
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+// GGUF metadata value type tags (subset) from the GGUF spec.
+const (
+	ggufTypeUint32 uint32 = 4
+	ggufTypeString uint32 = 8
+)
+
+type ggufKV struct {
+	key   string
+	vtype uint32
+	val   any
+}
+
+// writeMinimalGGUF emits a valid, tensor-less GGUF file carrying only the given
+// metadata key-values. Enough for the header-only parse path piperSampleRate
+// uses; avoids pulling a real multi-MB voice into the test.
+func writeMinimalGGUF(path string, kvs []ggufKV) error {
+	var b bytes.Buffer
+	b.WriteString("GGUF")                                // magic
+	_ = binary.Write(&b, binary.LittleEndian, uint32(3)) // version
+	_ = binary.Write(&b, binary.LittleEndian, uint64(0)) // tensor count
+	_ = binary.Write(&b, binary.LittleEndian, uint64(len(kvs)))
+	for _, kv := range kvs {
+		_ = binary.Write(&b, binary.LittleEndian, uint64(len(kv.key)))
+		b.WriteString(kv.key)
+		_ = binary.Write(&b, binary.LittleEndian, kv.vtype)
+		switch v := kv.val.(type) {
+		case uint32:
+			_ = binary.Write(&b, binary.LittleEndian, v)
+		case string:
+			_ = binary.Write(&b, binary.LittleEndian, uint64(len(v)))
+			b.WriteString(v)
+		}
+	}
+	return os.WriteFile(path, b.Bytes(), 0o644)
+}
+
+// wavSampleRate decodes the WAV header at path and returns its sample rate.
+func wavSampleRate(path string) (int, error) {
+	f, err := os.Open(path)
+	if err != nil {
+		return 0, err
+	}
+	defer func() { _ = f.Close() }()
+	dec := wav.NewDecoder(f)
+	dec.ReadInfo()
+	return int(dec.SampleRate), nil
+}
+
+var _ = Describe("piper sample rate", func() {
+	Context("piperSampleRate", func() {
+		It("reads piper.sample_rate from a piper GGUF (medium = 22050)", func() {
+			p := filepath.Join(GinkgoT().TempDir(), "voice.gguf")
+			Expect(writeMinimalGGUF(p, []ggufKV{
+				{key: "general.architecture", vtype: ggufTypeString, val: "piper"},
+				{key: "piper.sample_rate", vtype: ggufTypeUint32, val: uint32(22050)},
+			})).To(Succeed())
+
+			rate, ok := piperSampleRate(p)
+			Expect(ok).To(BeTrue(), "piper.sample_rate should be found")
+			Expect(rate).To(Equal(22050))
+		})
+
+		It("reads the low-quality rate (16000)", func() {
+			p := filepath.Join(GinkgoT().TempDir(), "voice.gguf")
+			Expect(writeMinimalGGUF(p, []ggufKV{
+				{key: "piper.sample_rate", vtype: ggufTypeUint32, val: uint32(16000)},
+			})).To(Succeed())
+
+			rate, ok := piperSampleRate(p)
+			Expect(ok).To(BeTrue())
+			Expect(rate).To(Equal(16000))
+		})
+
+		It("returns ok=false for a non-piper GGUF (no piper.sample_rate key)", func() {
+			p := filepath.Join(GinkgoT().TempDir(), "other.gguf")
+			Expect(writeMinimalGGUF(p, []ggufKV{
+				{key: "general.architecture", vtype: ggufTypeString, val: "vibevoice"},
+			})).To(Succeed())
+
+			_, ok := piperSampleRate(p)
+			Expect(ok).To(BeFalse())
+		})
+
+		It("returns ok=false for an unreadable/non-GGUF file", func() {
+			p := filepath.Join(GinkgoT().TempDir(), "garbage.gguf")
+			Expect(os.WriteFile(p, []byte("not a gguf"), 0o644)).To(Succeed())
+
+			_, ok := piperSampleRate(p)
+			Expect(ok).To(BeFalse())
+		})
+	})
+
+	// End-to-end through the built .so. Gated on CRISPASR_PIPER_MODEL_PATH (a
+	// real piper voice GGUF) like the other model-backed specs; never runs in
+	// default CI. Proves CrispASR's piper backend output rate flows into the
+	// WAV header instead of the hardcoded 24 kHz default.
+	Context("piper TTS end-to-end", func() {
+		It("writes the WAV at the model's native piper.sample_rate", func() {
+			model := os.Getenv("CRISPASR_PIPER_MODEL_PATH")
+			if model == "" {
+				Skip("set CRISPASR_PIPER_MODEL_PATH to run the piper e2e spec")
+			}
+			ensureLibLoaded()
+
+			expected, ok := piperSampleRate(model)
+			Expect(ok).To(BeTrue(), "model should carry piper.sample_rate metadata")
+
+			w := &CrispASR{}
+			Expect(w.Load(&pb.ModelOptions{
+				ModelFile: model,
+				Options:   []string{"backend:piper"},
+				Threads:   4,
+			})).To(Succeed())
+
+			dst := filepath.Join(GinkgoT().TempDir(), "piper.wav")
+			Expect(w.TTS(&pb.TTSRequest{Text: "Hello from CrispASR piper.", Dst: dst})).To(Succeed())
+
+			info, err := os.Stat(dst)
+			Expect(err).ToNot(HaveOccurred())
+			Expect(info.Size()).To(BeNumerically(">", 1024), "expected a non-trivial WAV")
+
+			rate, err := wavSampleRate(dst)
+			Expect(err).ToNot(HaveOccurred())
+			Expect(rate).To(Equal(expected),
+				"WAV header rate must equal the model's native piper.sample_rate, not the 24k default")
+		})
+	})
+
+	Context("writeWAV", func() {
+		It("writes the WAV header at the given sample rate (22050 for piper, not the 24k default)", func() {
+			dst := filepath.Join(GinkgoT().TempDir(), "out.wav")
+			pcm := make([]float32, 220) // 10 ms of silence is enough for a header
+			Expect(writeWAV(dst, pcm, 22050)).To(Succeed())
+
+			rate, err := wavSampleRate(dst)
+			Expect(err).ToNot(HaveOccurred())
+			Expect(rate).To(Equal(22050))
+		})
+
+		It("writes a 16000 Hz header for low-quality piper voices", func() {
+			dst := filepath.Join(GinkgoT().TempDir(), "out.wav")
+			pcm := make([]float32, 160)
+			Expect(writeWAV(dst, pcm, 16000)).To(Succeed())
+
+			rate, err := wavSampleRate(dst)
+			Expect(err).ToNot(HaveOccurred())
+			Expect(rate).To(Equal(16000))
+		})
+	})
+})

backend/go/crispasr/package.sh

@@ -51,6 +51,32 @@ else
     exit 1
 fi
 
+# Bundle espeak-ng (+ its libpcaudio/libsonic runtime deps) and its voice data so
+# the piper TTS backend can phonemize non-English text. CrispASR dlopens
+# libespeak-ng.so.1 at runtime (the MIT-clean path); the dlopen succeeds loading
+# libespeak-ng but FAILS if libpcaudio/libsonic are absent, so all three .so are
+# required. run.sh points CRISPASR_ESPEAK_DATA_PATH at the bundled data dir.
+# Best-effort: only copied when present, so a local dev build without espeak-ng
+# installed still packages the rest (English voices keep working).
+ESPEAK_LIBDIR=""
+for d in /usr/lib/x86_64-linux-gnu /usr/lib/aarch64-linux-gnu; do
+    if [ -f "$d/libespeak-ng.so.1" ]; then
+        ESPEAK_LIBDIR="$d"
+        break
+    fi
+done
+if [ -n "$ESPEAK_LIBDIR" ]; then
+    echo "Bundling espeak-ng from $ESPEAK_LIBDIR ..."
+    cp -arfLv "$ESPEAK_LIBDIR/libespeak-ng.so.1" $CURDIR/package/lib/
+    cp -arfLv "$ESPEAK_LIBDIR/libpcaudio.so.0" $CURDIR/package/lib/
+    cp -arfLv "$ESPEAK_LIBDIR/libsonic.so.0" $CURDIR/package/lib/
+    if [ -d "$ESPEAK_LIBDIR/espeak-ng-data" ]; then
+        cp -arfLv "$ESPEAK_LIBDIR/espeak-ng-data" $CURDIR/package/
+    fi
+else
+    echo "espeak-ng not found; non-English piper voices will not phonemize"
+fi
+
 # Package GPU libraries based on BUILD_TYPE
 # The GPU library packaging script will detect BUILD_TYPE and copy appropriate GPU libraries
 GPU_LIB_SCRIPT="${REPO_ROOT}/scripts/build/package-gpu-libs.sh"

backend/go/crispasr/run.sh

@@ -41,6 +41,11 @@ 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
+
 # If there is a lib/ld.so, use it
 if [ -f $CURDIR/lib/ld.so ]; then
 	echo "Using lib/ld.so"

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

@@ -0,0 +1,7 @@
+sources/
+build*/
+package/
+libdepthanythingcpp*.so
+depth-anything-cpp
+test-models/
+test-data/

backend/go/depth-anything-cpp/CMakeLists.txt

@@ -0,0 +1,28 @@
+cmake_minimum_required(VERSION 3.18)
+project(libdepthanythingcpp LANGUAGES C CXX)
+
+set(CMAKE_POSITION_INDEPENDENT_CODE ON)
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+
+# Static-link ggml into the depth-anything shared library so the resulting .so
+# has no runtime dependency on an external libggml — only on
+# libc/libstdc++/libgomp, which the LocalAI package step bundles into the
+# docker image.
+set(BUILD_SHARED_LIBS OFF CACHE BOOL "Build static libraries" FORCE)
+
+# depth-anything.cpp build switches: skip CLI/tests, but build libdepthanything
+# itself as a SHARED library (DA_SHARED) while ggml stays static
+# (BUILD_SHARED_LIBS OFF above). The da_capi_* C ABI is compiled into
+# src/da_capi.cpp and re-exported by that shared library, so no extra MODULE
+# wrapper is needed (unlike locate-anything.cpp).
+set(DA_BUILD_CLI OFF CACHE BOOL "Disable depth-anything CLI" FORCE)
+set(DA_BUILD_TESTS OFF CACHE BOOL "Disable depth-anything tests" FORCE)
+set(DA_SHARED ON CACHE BOOL "Build libdepthanything as a shared lib" FORCE)
+
+add_subdirectory(./sources/depth-anything.cpp)
+
+# Emit libdepthanything.so into the top-level build dir so the Makefile can
+# rename it to the per-variant libdepthanythingcpp-<variant>.so.
+set_target_properties(depthanything PROPERTIES
+    LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR})

backend/go/depth-anything-cpp/Makefile

@@ -0,0 +1,139 @@
+CMAKE_ARGS?=
+BUILD_TYPE?=
+NATIVE?=false
+
+GOCMD?=go
+GO_TAGS?=
+JOBS?=$(shell nproc --ignore=1)
+
+# depth-anything.cpp. Pin to a specific commit for a stable build; a squash
+# merge upstream can orphan a branch, so the native version is pinned by SHA.
+# This SHA adds the nested two-file metric C-API (abi_version 4,
+# da_capi_load_nested) required by the depth-anything-3-nested gallery model;
+# tag it (e.g. v0.1.3) upstream to keep the SHA alive.
+DEPTHANYTHING_REPO?=https://github.com/mudler/depth-anything.cpp.git
+DEPTHANYTHING_VERSION?=cce5edc395fd1843806093d7ccc0c8b0d0b97b72
+
+ifeq ($(NATIVE),false)
+	CMAKE_ARGS+=-DGGML_NATIVE=OFF
+endif
+
+# Forward LocalAI's BUILD_TYPE to the matching ggml backend switch. depth-anything.cpp
+# force-sets GGML_CUDA/GGML_VULKAN/GGML_METAL from its own DA_GGML_* options, so
+# those must be toggled via the DA_GGML_* names (a bare -DGGML_CUDA=ON would be
+# overridden); the remaining ggml switches pass straight through.
+ifeq ($(BUILD_TYPE),cublas)
+	CMAKE_ARGS+=-DGGML_CUDA=ON -DDA_GGML_CUDA=ON
+else ifeq ($(BUILD_TYPE),openblas)
+	CMAKE_ARGS+=-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
+else ifeq ($(BUILD_TYPE),clblas)
+	CMAKE_ARGS+=-DGGML_CLBLAST=ON
+else ifeq ($(BUILD_TYPE),hipblas)
+	ROCM_HOME ?= /opt/rocm
+	ROCM_PATH ?= /opt/rocm
+	export CXX=$(ROCM_HOME)/llvm/bin/clang++
+	export CC=$(ROCM_HOME)/llvm/bin/clang
+	AMDGPU_TARGETS?=gfx908,gfx90a,gfx942,gfx950,gfx1030,gfx1100,gfx1101,gfx1102,gfx1200,gfx1201
+	CMAKE_ARGS+=-DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=$(AMDGPU_TARGETS)
+else ifeq ($(BUILD_TYPE),vulkan)
+	CMAKE_ARGS+=-DGGML_VULKAN=ON -DDA_GGML_VULKAN=ON
+else ifeq ($(OS),Darwin)
+	ifneq ($(BUILD_TYPE),metal)
+		CMAKE_ARGS+=-DGGML_METAL=OFF
+	else
+		CMAKE_ARGS+=-DGGML_METAL=ON
+		CMAKE_ARGS+=-DGGML_METAL_EMBED_LIBRARY=ON
+		CMAKE_ARGS+=-DDA_GGML_METAL=ON
+	endif
+endif
+
+ifeq ($(BUILD_TYPE),sycl_f16)
+	CMAKE_ARGS+=-DGGML_SYCL=ON \
+		-DCMAKE_C_COMPILER=icx \
+		-DCMAKE_CXX_COMPILER=icpx \
+		-DGGML_SYCL_F16=ON
+endif
+
+ifeq ($(BUILD_TYPE),sycl_f32)
+	CMAKE_ARGS+=-DGGML_SYCL=ON \
+		-DCMAKE_C_COMPILER=icx \
+		-DCMAKE_CXX_COMPILER=icpx
+endif
+
+sources/depth-anything.cpp:
+	mkdir -p sources && \
+	git clone --recursive $(DEPTHANYTHING_REPO) sources/depth-anything.cpp && \
+	cd sources/depth-anything.cpp && \
+	git checkout $(DEPTHANYTHING_VERSION) && \
+	git submodule update --init --recursive --depth 1 --single-branch
+
+# Detect OS
+UNAME_S := $(shell uname -s)
+
+# Only build CPU variants on Linux
+ifeq ($(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
+endif
+
+depth-anything-cpp: main.go godepthanythingcpp.go $(VARIANT_TARGETS)
+	CGO_ENABLED=0 $(GOCMD) build -tags "$(GO_TAGS)" -o depth-anything-cpp ./
+
+package: depth-anything-cpp
+	bash package.sh
+
+build: package
+
+clean: purge
+	rm -rf libdepthanythingcpp*.so depth-anything-cpp package sources
+
+purge:
+	rm -rf build*
+
+# Build all variants (Linux only)
+ifeq ($(UNAME_S),Linux)
+libdepthanythingcpp-avx.so: sources/depth-anything.cpp
+	rm -rfv build-$@
+	$(info ${GREEN}I depth-anything-cpp build info:avx${RESET})
+	SO_TARGET=$@ CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off" $(MAKE) libdepthanythingcpp-custom
+	rm -rfv build-$@
+
+libdepthanythingcpp-avx2.so: sources/depth-anything.cpp
+	rm -rfv build-$@
+	$(info ${GREEN}I depth-anything-cpp build info:avx2${RESET})
+	SO_TARGET=$@ CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on -DGGML_BMI2=on" $(MAKE) libdepthanythingcpp-custom
+	rm -rfv build-$@
+
+libdepthanythingcpp-avx512.so: sources/depth-anything.cpp
+	rm -rfv build-$@
+	$(info ${GREEN}I depth-anything-cpp build info:avx512${RESET})
+	SO_TARGET=$@ CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=on -DGGML_FMA=on -DGGML_F16C=on -DGGML_BMI2=on" $(MAKE) libdepthanythingcpp-custom
+	rm -rfv build-$@
+endif
+
+# Build fallback variant (all platforms)
+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" $(MAKE) libdepthanythingcpp-custom
+	rm -rfv build-$@
+
+libdepthanythingcpp-custom: CMakeLists.txt
+	mkdir -p build-$(SO_TARGET) && \
+	cd build-$(SO_TARGET) && \
+	cmake .. $(CMAKE_ARGS) && \
+	cmake --build . --config Release -j$(JOBS) && \
+	cd .. && \
+	mv build-$(SO_TARGET)/libdepthanything.so ./$(SO_TARGET)
+
+all: depth-anything-cpp package
+
+# `test` is invoked by the top-level Makefile's `test-extra` target. It builds
+# the backend binary + the fallback shared library (needed for dlopen at
+# runtime), then runs test.sh which downloads a small GGUF + a test image and
+# exercises the gRPC Load/Predict wire path via the Go smoke test in
+# main_test.go.
+test: depth-anything-cpp libdepthanythingcpp-fallback.so
+	bash test.sh

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

@@ -0,0 +1,556 @@
+package main
+
+// godepthanythingcpp.go - gRPC handlers (Load, Predict, GenerateImage) for the
+// depth-anything-cpp backend, wrapping the Depth Anything 3 ggml C-API
+// (libdepthanythingcpp-<variant>.so) via purego.
+//
+// Embeds base.SingleThread to default the unimplemented RPCs to "not supported"
+// and to serialize calls — the C side shares a ggml graph allocator and is NOT
+// reentrant, so all inference must run one-at-a-time.
+//
+// Depth has no native OpenAI endpoint, so the model is exposed two ways:
+//
+//   - GenerateImage(src, dst): run depth on the src image and write a
+//     min-max-normalised grayscale depth PNG to dst.
+//   - Predict(images[0]): run depth+pose and return a JSON blob with the depth
+//     dimensions, depth stats and the camera extrinsics (3x4) / intrinsics (3x3).
+
+import (
+	"encoding/base64"
+	"encoding/json"
+	"fmt"
+	"image"
+	"image/png"
+	"math"
+	"os"
+	"path/filepath"
+	"strings"
+	"unsafe"
+
+	"github.com/mudler/LocalAI/pkg/grpc/base"
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+)
+
+// C-API function pointers, registered in main.go via purego. The da_capi_*
+// symbols live inside libdepthanything (src/da_capi.cpp) and are re-exported by
+// the DA_SHARED build.
+var (
+	// da_capi_load(const char* gguf_path, int n_threads) -> da_ctx* (0 = fail)
+	CapiLoad func(gguf string, nThreads int32) uintptr
+	// da_capi_load_nested(const char* anyview_gguf, const char* metric_gguf,
+	//   int n_threads) -> da_ctx* (0 = fail). The returned ctx serves the nested
+	//   metric model: depth/pose calls produce final metric-scale depth + scaled pose.
+	CapiLoadNested func(anyview string, metric string, nThreads int32) uintptr
+	// da_capi_free(da_ctx* ctx) — safe on a 0 handle.
+	CapiFree func(handle uintptr)
+	// da_capi_last_error(da_ctx* ctx) -> const char* (owned by ctx, "" if none).
+	// purego marshals the returned C string into a Go string (a copy), so we
+	// never free it.
+	CapiLastError func(handle uintptr) string
+	// da_capi_depth_path(ctx, image_path, out_h*, out_w*) -> float* depth map
+	// (row-major H*W); nil on error. Caller frees via da_capi_free_floats.
+	CapiDepthPath func(handle uintptr, imagePath string, outH *int32, outW *int32) *float32
+	// da_capi_free_floats(float* p)
+	CapiFreeFloats func(p *float32)
+	// da_capi_pose_path(ctx, image_path, out_ext[12], out_intr[9]) -> 0 ok, -1 err
+	CapiPosePath func(handle uintptr, imagePath string, outExt *float32, outIntr *float32) int32
+	// da_capi_depth_dense(ctx, image_path, out_h*, out_w*, out_depth**, out_conf**,
+	//   out_sky**, out_ext[12], out_intr[9], out_is_metric*) -> 0 ok, -1 err.
+	// Each non-NULL out_depth/out_conf/out_sky receives a malloc'd float[H*W] (free
+	// via da_capi_free_floats); buffers the model doesn't produce are set NULL.
+	CapiDepthDense func(handle uintptr, imagePath string,
+		outH, outW *int32,
+		outDepth, outConf, outSky **float32,
+		outExt, outIntr *float32,
+		outIsMetric *int32) int32
+	// da_capi_points(ctx, image_path, conf_thresh, out_n*, out_xyz**, out_rgb**) ->
+	//   0 ok, -1 err. *out_xyz = malloc'd float[3*N] (free via da_capi_free_floats),
+	//   *out_rgb = malloc'd uint8[3*N] (free via da_capi_free_bytes).
+	CapiPoints func(handle uintptr, imagePath string, confThresh float32,
+		outN *int32, outXyz **float32, outRgb **byte) int32
+	// da_capi_free_bytes(unsigned char* p)
+	CapiFreeBytes func(p *byte)
+	// da_capi_export_glb(ctx, image_path, out_glb) -> 0 ok, -1 err
+	CapiExportGlb func(handle uintptr, imagePath string, outGlb string) int32
+	// da_capi_export_colmap(ctx, image_path, out_dir, binary) -> 0 ok, -1 err
+	CapiExportColmap func(handle uintptr, imagePath string, outDir string, binary int32) int32
+)
+
+type DepthAnythingCpp struct {
+	base.SingleThread
+	handle uintptr
+}
+
+// Load loads the GGUF model at opts.ModelFile (joined with opts.ModelPath if
+// relative) and stores the da_ctx handle for later inference calls.
+func (r *DepthAnythingCpp) Load(opts *pb.ModelOptions) error {
+	modelFile := opts.ModelFile
+	if modelFile == "" {
+		modelFile = opts.Model
+	}
+	if modelFile == "" {
+		return fmt.Errorf("depth-anything-cpp: ModelFile is empty")
+	}
+
+	resolve := func(name string) string {
+		if filepath.IsAbs(name) {
+			return name
+		}
+		return filepath.Join(opts.ModelPath, name)
+	}
+	modelPath := resolve(modelFile)
+
+	if _, err := os.Stat(modelPath); err != nil {
+		return fmt.Errorf("depth-anything-cpp: model file not found: %s: %w", modelPath, err)
+	}
+
+	// Nested metric models are a two-file pair: the main model is the anyview
+	// (GIANT) branch and the metric (ViT-L + DPT/sky) branch is named via a
+	// "metric_model:<filename>" entry in opts.Options. When present we load both
+	// branches so the engine runs the nested metric alignment.
+	metricFile := optionValue(opts.Options, "metric_model")
+
+	threads := opts.Threads
+	if threads <= 0 {
+		threads = 4
+	}
+
+	// Release previous model if any (re-Load).
+	if r.handle != 0 {
+		CapiFree(r.handle)
+		r.handle = 0
+	}
+
+	var h uintptr
+	if metricFile != "" {
+		metricPath := resolve(metricFile)
+		if _, err := os.Stat(metricPath); err != nil {
+			return fmt.Errorf("depth-anything-cpp: metric_model file not found: %s: %w", metricPath, err)
+		}
+		h = CapiLoadNested(modelPath, metricPath, threads)
+		if h == 0 {
+			if msg := CapiLastError(0); msg != "" {
+				return fmt.Errorf("depth-anything-cpp: da_capi_load_nested failed for %s + %s: %s", modelPath, metricPath, msg)
+			}
+			return fmt.Errorf("depth-anything-cpp: da_capi_load_nested failed for %s + %s", modelPath, metricPath)
+		}
+	} else {
+		h = CapiLoad(modelPath, threads)
+		if h == 0 {
+			// da_capi_last_error needs a ctx; on a failed load we have none (it
+			// returns "" for a null ctx), so the text is best-effort.
+			if msg := CapiLastError(0); msg != "" {
+				return fmt.Errorf("depth-anything-cpp: da_capi_load failed for %s: %s", modelPath, msg)
+			}
+			return fmt.Errorf("depth-anything-cpp: da_capi_load failed for %s", modelPath)
+		}
+	}
+	r.handle = h
+	return nil
+}
+
+// optionValue returns the value of the first "key:value" entry in opts whose key
+// matches (case-sensitive), or "" if absent. Mirrors how other LocalAI backends
+// read ModelOptions.Options.
+func optionValue(opts []string, key string) string {
+	prefix := key + ":"
+	for _, o := range opts {
+		if strings.HasPrefix(o, prefix) {
+			return strings.TrimSpace(o[len(prefix):])
+		}
+	}
+	return ""
+}
+
+// depthResult is the JSON payload returned by Predict.
+type depthResult struct {
+	DepthW     int         `json:"depth_w"`
+	DepthH     int         `json:"depth_h"`
+	DepthMin   float32     `json:"depth_min"`
+	DepthMax   float32     `json:"depth_max"`
+	Extrinsics [12]float32 `json:"extrinsics"` // 3x4 row-major
+	Intrinsics [9]float32  `json:"intrinsics"` // 3x3 row-major
+}
+
+// Predict runs depth+pose on the first supplied image and returns depth
+// statistics + camera pose as a JSON string. LocalAI wraps the string into the
+// Reply.Message of the gRPC response. The image in Images[0] may be a
+// filesystem path or a base64-encoded payload.
+func (r *DepthAnythingCpp) Predict(opts *pb.PredictOptions) (string, error) {
+	imgs := opts.GetImages()
+	if len(imgs) == 0 {
+		return "", fmt.Errorf("depth-anything-cpp: Predict requires an image in Images[]")
+	}
+
+	imgPath, cleanup, err := materializeImage(imgs[0])
+	if err != nil {
+		return "", fmt.Errorf("depth-anything-cpp: %w", err)
+	}
+	defer cleanup()
+
+	depth, h, w, ext, intr, err := r.runDepthPose(imgPath)
+	if err != nil {
+		return "", err
+	}
+
+	dmin, dmax := minMax(depth)
+	payload, err := json.Marshal(depthResult{
+		DepthW: w, DepthH: h,
+		DepthMin: dmin, DepthMax: dmax,
+		Extrinsics: ext, Intrinsics: intr,
+	})
+	if err != nil {
+		return "", fmt.Errorf("depth-anything-cpp: marshal: %w", err)
+	}
+	return string(payload), nil
+}
+
+// GenerateImage runs depth on req.Src and writes a normalised grayscale depth
+// PNG to req.Dst.
+func (r *DepthAnythingCpp) GenerateImage(req *pb.GenerateImageRequest) error {
+	if req.GetSrc() == "" {
+		return fmt.Errorf("depth-anything-cpp: GenerateImage requires src")
+	}
+	if req.GetDst() == "" {
+		return fmt.Errorf("depth-anything-cpp: GenerateImage requires dst")
+	}
+
+	imgPath, cleanup, err := materializeImage(req.GetSrc())
+	if err != nil {
+		return fmt.Errorf("depth-anything-cpp: %w", err)
+	}
+	defer cleanup()
+
+	depth, h, w, _, _, err := r.runDepthPose(imgPath)
+	if err != nil {
+		return err
+	}
+	return writeDepthPNG(req.GetDst(), depth, h, w)
+}
+
+// Depth is the typed Depth RPC. It runs the Depth Anything 3 pipeline on the
+// request's src image and fills a DepthResponse honoring the include_* flags and
+// exports: per-pixel metric depth + confidence (DualDPT) or depth + sky (mono),
+// camera extrinsics/intrinsics, an optional back-projected 3D point cloud and
+// glb/COLMAP exports. The src may be a filesystem path or a base64 payload.
+func (r *DepthAnythingCpp) Depth(in *pb.DepthRequest) (pb.DepthResponse, error) {
+	// Accumulate into locals and return a single composite literal at the end:
+	// returning a named pb.DepthResponse value would copy its embedded mutex
+	// (go vet copylocks).
+	if r.handle == 0 {
+		return pb.DepthResponse{}, fmt.Errorf("depth-anything-cpp: model not loaded")
+	}
+	if in.GetSrc() == "" {
+		return pb.DepthResponse{}, fmt.Errorf("depth-anything-cpp: Depth requires src")
+	}
+
+	imgPath, cleanup, err := materializeImage(in.GetSrc())
+	if err != nil {
+		return pb.DepthResponse{}, fmt.Errorf("depth-anything-cpp: %w", err)
+	}
+	defer cleanup()
+
+	// Dense per-pixel output + pose. Pass buffer pointers only for the
+	// requested maps so the native side can skip unrequested work; ext/intr
+	// must always point at 12/9 floats per the C ABI.
+	var (
+		h, w, isMetric      int32
+		depthPtr, confPtr   *float32
+		skyPtr              *float32
+		ext                 [12]float32
+		intr                [9]float32
+		pDepth, pConf, pSky **float32
+	)
+	if in.GetIncludeDepth() {
+		pDepth = &depthPtr
+	}
+	if in.GetIncludeConfidence() {
+		pConf = &confPtr
+	}
+	if in.GetIncludeSky() {
+		pSky = &skyPtr
+	}
+
+	rc := CapiDepthDense(r.handle, imgPath, &h, &w, pDepth, pConf, pSky, &ext[0], &intr[0], &isMetric)
+	if rc != 0 {
+		return pb.DepthResponse{}, fmt.Errorf("depth-anything-cpp: da_capi_depth_dense failed (rc=%d): %s", rc, r.lastError())
+	}
+
+	n := int(h) * int(w)
+	var (
+		depth, conf, sky      []float32
+		extrinsics, intrinsic []float32
+		numPoints             int32
+		points                []float32
+		pointColors           []byte
+		exportPaths           []string
+	)
+
+	if depthPtr != nil {
+		depth = copyFloats(depthPtr, n)
+		CapiFreeFloats(depthPtr)
+	}
+	if confPtr != nil {
+		conf = copyFloats(confPtr, n)
+		CapiFreeFloats(confPtr)
+	}
+	if skyPtr != nil {
+		sky = copyFloats(skyPtr, n)
+		CapiFreeFloats(skyPtr)
+	}
+	if in.GetIncludePose() {
+		extrinsics = append([]float32(nil), ext[:]...)
+		intrinsic = append([]float32(nil), intr[:]...)
+	}
+
+	// 3D point cloud (DualDPT / pose-capable models only).
+	if in.GetIncludePoints() {
+		var (
+			np     int32
+			xyzPtr *float32
+			rgbPtr *byte
+		)
+		if rc := CapiPoints(r.handle, imgPath, in.GetPointsConfThresh(), &np, &xyzPtr, &rgbPtr); rc != 0 {
+			return pb.DepthResponse{}, fmt.Errorf("depth-anything-cpp: da_capi_points failed (rc=%d): %s", rc, r.lastError())
+		}
+		numPoints = np
+		if xyzPtr != nil {
+			points = copyFloats(xyzPtr, int(np)*3)
+			CapiFreeFloats(xyzPtr)
+		}
+		if rgbPtr != nil {
+			pointColors = copyBytes(rgbPtr, int(np)*3)
+			CapiFreeBytes(rgbPtr)
+		}
+	}
+
+	// Exports (glb / colmap). They are written under in.Dst (a directory); a
+	// temp dir is used when Dst is empty.
+	if len(in.GetExports()) > 0 {
+		exportPaths, err = r.runExports(imgPath, in.GetDst(), in.GetExports())
+		if err != nil {
+			return pb.DepthResponse{}, err
+		}
+	}
+
+	return pb.DepthResponse{
+		Width:       w,
+		Height:      h,
+		Depth:       depth,
+		Confidence:  conf,
+		Sky:         sky,
+		Extrinsics:  extrinsics,
+		Intrinsics:  intrinsic,
+		NumPoints:   numPoints,
+		Points:      points,
+		PointColors: pointColors,
+		ExportPaths: exportPaths,
+		IsMetric:    isMetric != 0,
+	}, nil
+}
+
+// runExports writes the requested exports for imgPath into dstDir and returns
+// the written paths. Supported exports: "glb", "colmap".
+func (r *DepthAnythingCpp) runExports(imgPath, dstDir string, exports []string) ([]string, error) {
+	if dstDir == "" {
+		tmp, err := os.MkdirTemp("", "depth-anything-export-*")
+		if err != nil {
+			return nil, fmt.Errorf("depth-anything-cpp: mkdir export dir: %w", err)
+		}
+		dstDir = tmp
+	} else if err := os.MkdirAll(dstDir, 0o750); err != nil {
+		return nil, fmt.Errorf("depth-anything-cpp: mkdir %s: %w", dstDir, err)
+	}
+
+	var paths []string
+	for _, exp := range exports {
+		switch exp {
+		case "glb":
+			out := filepath.Join(dstDir, "pointcloud.glb")
+			if rc := CapiExportGlb(r.handle, imgPath, out); rc != 0 {
+				return nil, fmt.Errorf("depth-anything-cpp: da_capi_export_glb failed (rc=%d): %s", rc, r.lastError())
+			}
+			paths = append(paths, out)
+		case "colmap":
+			out := filepath.Join(dstDir, "colmap")
+			if err := os.MkdirAll(out, 0o750); err != nil {
+				return nil, fmt.Errorf("depth-anything-cpp: mkdir %s: %w", out, err)
+			}
+			if rc := CapiExportColmap(r.handle, imgPath, out, 1); rc != 0 {
+				return nil, fmt.Errorf("depth-anything-cpp: da_capi_export_colmap failed (rc=%d): %s", rc, r.lastError())
+			}
+			paths = append(paths, out)
+		default:
+			return nil, fmt.Errorf("depth-anything-cpp: unknown export %q (want glb|colmap)", exp)
+		}
+	}
+	return paths, nil
+}
+
+// copyFloats copies n float32 values from a C heap pointer into a fresh Go
+// slice so the C buffer can be freed afterwards.
+func copyFloats(p *float32, n int) []float32 {
+	if p == nil || n <= 0 {
+		return nil
+	}
+	src := unsafe.Slice(p, n)
+	out := make([]float32, n)
+	copy(out, src)
+	return out
+}
+
+// copyBytes copies n bytes from a C heap pointer into a fresh Go slice.
+func copyBytes(p *byte, n int) []byte {
+	if p == nil || n <= 0 {
+		return nil
+	}
+	src := unsafe.Slice(p, n)
+	out := make([]byte, n)
+	copy(out, src)
+	return out
+}
+
+// runDepthPose runs depth estimation then pose recovery on an image file. It
+// returns the row-major depth map (length h*w), its dimensions, the 3x4
+// extrinsics (12 floats) and 3x3 intrinsics (9 floats).
+// runDepthPose returns depth + camera pose via two C-API calls (depth then pose).
+// For a nested metric model both calls run the full two-branch pipeline, so this
+// path infers twice; the typed Depth RPC (single da_capi_depth_dense call) is the
+// efficient path for nested models.
+func (r *DepthAnythingCpp) runDepthPose(imagePath string) (depth []float32, h, w int, ext [12]float32, intr [9]float32, err error) {
+	if r.handle == 0 {
+		err = fmt.Errorf("depth-anything-cpp: model not loaded")
+		return
+	}
+
+	var ch, cw int32
+	ptr := CapiDepthPath(r.handle, imagePath, &ch, &cw)
+	if ptr == nil {
+		err = fmt.Errorf("depth-anything-cpp: da_capi_depth_path failed: %s", r.lastError())
+		return
+	}
+	h, w = int(ch), int(cw)
+	n := h * w
+	if n > 0 {
+		src := unsafe.Slice(ptr, n)
+		depth = make([]float32, n)
+		copy(depth, src)
+	}
+	CapiFreeFloats(ptr)
+
+	if rc := CapiPosePath(r.handle, imagePath, &ext[0], &intr[0]); rc != 0 {
+		err = fmt.Errorf("depth-anything-cpp: da_capi_pose_path failed (rc=%d): %s", rc, r.lastError())
+		return
+	}
+	return
+}
+
+// lastError returns the context's last error string, or "" if none.
+func (r *DepthAnythingCpp) lastError() string {
+	if CapiLastError == nil || r.handle == 0 {
+		return ""
+	}
+	return CapiLastError(r.handle)
+}
+
+// materializeImage returns a filesystem path for an image argument that may be
+// either an existing path or a base64-encoded payload. When the input is
+// base64 it is decoded into a temp file; cleanup removes it (no-op for a path).
+func materializeImage(arg string) (path string, cleanup func(), err error) {
+	cleanup = func() {}
+	if _, statErr := os.Stat(arg); statErr == nil {
+		return arg, cleanup, nil
+	}
+	// Strip an optional data URL prefix (data:image/...;base64,<payload>).
+	b64 := arg
+	if i := indexComma(b64); i >= 0 && hasDataPrefix(b64) {
+		b64 = b64[i+1:]
+	}
+	data, decErr := base64.StdEncoding.DecodeString(b64)
+	if decErr != nil {
+		return "", cleanup, fmt.Errorf("image is neither an existing path nor valid base64: %v", decErr)
+	}
+	f, tErr := os.CreateTemp("", "depth-anything-*.img")
+	if tErr != nil {
+		return "", cleanup, tErr
+	}
+	if _, wErr := f.Write(data); wErr != nil {
+		_ = f.Close()
+		_ = os.Remove(f.Name())
+		return "", cleanup, wErr
+	}
+	_ = f.Close()
+	name := f.Name()
+	return name, func() { _ = os.Remove(name) }, nil
+}
+
+func hasDataPrefix(s string) bool {
+	return len(s) >= 5 && s[:5] == "data:"
+}
+
+func indexComma(s string) int {
+	for i := 0; i < len(s); i++ {
+		if s[i] == ',' {
+			return i
+		}
+	}
+	return -1
+}
+
+// writeDepthPNG min-max normalises a depth map and writes it as an 8-bit
+// grayscale PNG. Near = bright (255), far = dark (0), matching the usual
+// depth-map convention for inverse-depth-like outputs.
+func writeDepthPNG(dst string, depth []float32, h, w int) error {
+	if h <= 0 || w <= 0 || len(depth) < h*w {
+		return fmt.Errorf("depth-anything-cpp: writeDepthPNG: bad dims h=%d w=%d len=%d", h, w, len(depth))
+	}
+	dmin, dmax := minMax(depth)
+	span := dmax - dmin
+	if span <= 0 || math.IsNaN(float64(span)) {
+		span = 1
+	}
+	img := image.NewGray(image.Rect(0, 0, w, h))
+	for y := 0; y < h; y++ {
+		for x := 0; x < w; x++ {
+			v := depth[y*w+x]
+			n := (v - dmin) / span // 0..1
+			if math.IsNaN(float64(n)) {
+				n = 0
+			}
+			if n < 0 {
+				n = 0
+			} else if n > 1 {
+				n = 1
+			}
+			img.Pix[y*img.Stride+x] = uint8(n * 255)
+		}
+	}
+	// dst is the gRPC-provided output path chosen by the LocalAI core (the
+	// intended write destination for the rendered depth map), not
+	// attacker-controlled input, so the variable path is expected here.
+	f, err := os.Create(dst) // #nosec G304
+	if err != nil {
+		return err
+	}
+	defer func() { _ = f.Close() }()
+	return png.Encode(f, img)
+}
+
+func minMax(v []float32) (mn, mx float32) {
+	if len(v) == 0 {
+		return 0, 0
+	}
+	mn, mx = v[0], v[0]
+	for _, x := range v {
+		if math.IsNaN(float64(x)) || math.IsInf(float64(x), 0) {
+			continue
+		}
+		if x < mn {
+			mn = x
+		}
+		if x > mx {
+			mx = x
+		}
+	}
+	return mn, mx
+}

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

@@ -0,0 +1,62 @@
+package main
+
+// main.go - entry point for the depth-anything-cpp gRPC backend.
+//
+// Dlopens libdepthanythingcpp-<variant>.so via purego at the path in
+// DEPTHANYTHING_LIBRARY (set by run.sh based on /proc/cpuinfo), registers the
+// da_capi_* C ABI symbols, then starts the gRPC server.
+
+import (
+	"flag"
+	"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() {
+	// Get library name from environment variable, default to fallback
+	libName := os.Getenv("DEPTHANYTHING_LIBRARY")
+	if libName == "" {
+		libName = "./libdepthanythingcpp-fallback.so"
+	}
+
+	lib, err := purego.Dlopen(libName, purego.RTLD_NOW|purego.RTLD_GLOBAL)
+	if err != nil {
+		panic(err)
+	}
+
+	libFuncs := []LibFuncs{
+		{&CapiLoad, "da_capi_load"},
+		{&CapiLoadNested, "da_capi_load_nested"},
+		{&CapiFree, "da_capi_free"},
+		{&CapiLastError, "da_capi_last_error"},
+		{&CapiDepthPath, "da_capi_depth_path"},
+		{&CapiFreeFloats, "da_capi_free_floats"},
+		{&CapiPosePath, "da_capi_pose_path"},
+		{&CapiDepthDense, "da_capi_depth_dense"},
+		{&CapiPoints, "da_capi_points"},
+		{&CapiFreeBytes, "da_capi_free_bytes"},
+		{&CapiExportGlb, "da_capi_export_glb"},
+		{&CapiExportColmap, "da_capi_export_colmap"},
+	}
+
+	for _, lf := range libFuncs {
+		purego.RegisterLibFunc(lf.FuncPtr, lib, lf.Name)
+	}
+
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &DepthAnythingCpp{}); err != nil {
+		panic(err)
+	}
+}

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

@@ -0,0 +1,167 @@
+package main
+
+// main_test.go - end-to-end smoke test for the depth-anything-cpp gRPC backend.
+//
+// Spawns the compiled depth-anything-cpp binary on a free local port, dials it
+// via gRPC, and exercises LoadModel + Predict against the test fixtures
+// downloaded by test.sh: the small (vits) f32 GGUF of Depth Anything 3 and a
+// real photo. Asserts that Predict returns a JSON payload with a positive
+// depth-map width/height.
+//
+// The spec Skip()s cleanly if its fixtures (the model, the test image, the
+// built binary, or the fallback .so) are missing, so the test target stays
+// usable on a fresh checkout / on CI runners where the model hasn't been
+// downloaded.
+
+import (
+	"context"
+	"encoding/base64"
+	"encoding/json"
+	"fmt"
+	"net"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"testing"
+	"time"
+
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+	"google.golang.org/grpc"
+	"google.golang.org/grpc/credentials/insecure"
+)
+
+func TestDepth(t *testing.T) {
+	RegisterFailHandler(Fail)
+	RunSpecs(t, "depth-anything-cpp backend smoke suite")
+}
+
+// freePort grabs an ephemeral TCP port and immediately releases it so the
+// spawned backend can bind to it. There is a tiny TOCTOU window here but in
+// practice it's adequate for a smoke test on a quiet runner.
+func freePort() int {
+	l, err := net.Listen("tcp", "127.0.0.1:0")
+	Expect(err).ToNot(HaveOccurred(), "freePort listen")
+	port := l.Addr().(*net.TCPAddr).Port
+	Expect(l.Close()).To(Succeed())
+	return port
+}
+
+// startBackend spawns the depth-anything-cpp binary on the given port and waits
+// until it accepts TCP connections (up to 10s). It mirrors how main.go resolves
+// the purego library: the DEPTHANYTHING_LIBRARY env var points the dlopen at the
+// freshly built fallback .so. The returned cleanup func kills the process.
+func startBackend(port int) func() {
+	binary, err := filepath.Abs("./depth-anything-cpp")
+	Expect(err).ToNot(HaveOccurred())
+	if _, err := os.Stat(binary); err != nil {
+		Skip(fmt.Sprintf("backend binary not built: %s (run `make depth-anything-cpp` first)", binary))
+	}
+
+	libPath, err := filepath.Abs("./libdepthanythingcpp-fallback.so")
+	Expect(err).ToNot(HaveOccurred())
+	if _, err := os.Stat(libPath); err != nil {
+		Skip(fmt.Sprintf("fallback library not built: %s (run `make libdepthanythingcpp-fallback.so` first)", libPath))
+	}
+
+	addr := fmt.Sprintf("127.0.0.1:%d", port)
+	cmd := exec.Command(binary, "--addr", addr)
+	cmd.Env = append(os.Environ(), "DEPTHANYTHING_LIBRARY="+libPath)
+	cmd.Stdout = os.Stderr
+	cmd.Stderr = os.Stderr
+	Expect(cmd.Start()).To(Succeed())
+
+	cleanup := func() {
+		if cmd.Process != nil {
+			_ = cmd.Process.Kill()
+			_, _ = cmd.Process.Wait()
+		}
+	}
+
+	deadline := time.Now().Add(10 * time.Second)
+	for time.Now().Before(deadline) {
+		c, err := net.DialTimeout("tcp", addr, 200*time.Millisecond)
+		if err == nil {
+			_ = c.Close()
+			return cleanup
+		}
+		time.Sleep(200 * time.Millisecond)
+	}
+
+	cleanup()
+	Fail(fmt.Sprintf("backend did not become ready on %s within 10s", addr))
+	return func() {}
+}
+
+// loadTestImage reads the test image downloaded by test.sh and returns its
+// base64-encoded content (one of the wire formats accepted by Predict).
+func loadTestImage() string {
+	imgPath, err := filepath.Abs("test-data/test.jpg")
+	Expect(err).ToNot(HaveOccurred())
+	imgBytes, err := os.ReadFile(imgPath)
+	if err != nil {
+		Skip(fmt.Sprintf("test image not present: %s (run test.sh first)", imgPath))
+	}
+	return base64.StdEncoding.EncodeToString(imgBytes)
+}
+
+// dialBackend opens a gRPC client connection to the spawned backend.
+func dialBackend(port int) (pb.BackendClient, func()) {
+	addr := fmt.Sprintf("127.0.0.1:%d", port)
+	conn, err := grpc.NewClient(addr, grpc.WithTransportCredentials(insecure.NewCredentials()))
+	Expect(err).ToNot(HaveOccurred())
+	return pb.NewBackendClient(conn), func() { _ = conn.Close() }
+}
+
+// modelPathOrSkip resolves the model file under ./test-models/ and Skip()s the
+// current spec if it's missing (not present on a fresh checkout / on CI runners
+// without the download).
+func modelPathOrSkip(name string) string {
+	modelDir, err := filepath.Abs("test-models")
+	Expect(err).ToNot(HaveOccurred())
+	modelPath := filepath.Join(modelDir, name)
+	if _, err := os.Stat(modelPath); err != nil {
+		Skip(fmt.Sprintf("model not present: %s (run test.sh first)", modelPath))
+	}
+	return modelPath
+}
+
+var _ = Describe("depth-anything-cpp backend", func() {
+	It("runs depth+pose against a known-good image", func() {
+		modelPath := modelPathOrSkip("depth-anything-small-f32.gguf")
+		imgB64 := loadTestImage()
+
+		port := freePort()
+		cleanup := startBackend(port)
+		defer cleanup()
+
+		client, closeConn := dialBackend(port)
+		defer closeConn()
+
+		ctx, cancel := context.WithTimeout(context.Background(), 20*time.Minute)
+		defer cancel()
+
+		loadResp, err := client.LoadModel(ctx, &pb.ModelOptions{
+			Model:     "depth-anything-small-f32.gguf",
+			ModelFile: modelPath,
+			Threads:   4,
+		})
+		Expect(err).ToNot(HaveOccurred(), "LoadModel")
+		Expect(loadResp.GetSuccess()).To(BeTrue(), "LoadModel reported failure: %s", loadResp.GetMessage())
+
+		// Predict runs depth+pose and returns the JSON depthResult in Reply.Message.
+		reply, err := client.Predict(ctx, &pb.PredictOptions{
+			Images: []string{imgB64},
+		})
+		Expect(err).ToNot(HaveOccurred(), "Predict")
+
+		var res depthResult
+		Expect(json.Unmarshal(reply.GetMessage(), &res)).To(Succeed(), "Predict returned non-JSON: %q", string(reply.GetMessage()))
+		Expect(res.DepthW).To(BeNumerically(">", 0), "depth width should be positive")
+		Expect(res.DepthH).To(BeNumerically(">", 0), "depth height should be positive")
+
+		_, _ = fmt.Fprintf(GinkgoWriter, "depth OK: %dx%d min=%.3f max=%.3f\n",
+			res.DepthW, res.DepthH, res.DepthMin, res.DepthMax)
+	})
+})

backend/go/depth-anything-cpp/nested_e2e_test.go

@@ -0,0 +1,64 @@
+package main
+
+// nested_e2e_test.go - e2e smoke for the nested two-file metric model. Loads the
+// anyview branch as the main model and points the metric branch via the
+// "metric_model:<file>" option (exactly as the depth-anything-3-nested gallery
+// entry does), then exercises the typed Depth RPC and asserts a metric depth map.
+//
+// Skips cleanly unless both nested GGUFs are present under ./test-models/ and the
+// backend binary + fallback .so are built.
+
+import (
+	"context"
+	"fmt"
+	"path/filepath"
+	"time"
+
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+var _ = Describe("depth-anything-cpp nested metric model", func() {
+	It("loads the two-file pair via the metric_model option and returns metric depth", func() {
+		anyviewPath := modelPathOrSkip("depth-anything-nested-anyview.gguf")
+		_ = modelPathOrSkip("depth-anything-nested-metric.gguf")
+		imgB64 := loadTestImage()
+
+		port := freePort()
+		cleanup := startBackend(port)
+		defer cleanup()
+
+		client, closeConn := dialBackend(port)
+		defer closeConn()
+
+		ctx, cancel := context.WithTimeout(context.Background(), 25*time.Minute)
+		defer cancel()
+
+		loadResp, err := client.LoadModel(ctx, &pb.ModelOptions{
+			Model:     "depth-anything-nested-anyview.gguf",
+			ModelFile: anyviewPath,
+			ModelPath: filepath.Dir(anyviewPath),
+			Options:   []string{"metric_model:depth-anything-nested-metric.gguf"},
+			Threads:   8,
+		})
+		Expect(err).ToNot(HaveOccurred(), "LoadModel(nested)")
+		Expect(loadResp.GetSuccess()).To(BeTrue(), "LoadModel reported failure: %s", loadResp.GetMessage())
+
+		resp, err := client.Depth(ctx, &pb.DepthRequest{
+			Src:          imgB64,
+			IncludeDepth: true,
+			IncludePose:  true,
+		})
+		Expect(err).ToNot(HaveOccurred(), "Depth(nested)")
+		Expect(resp.GetWidth()).To(BeNumerically(">", 0), "depth width")
+		Expect(resp.GetHeight()).To(BeNumerically(">", 0), "depth height")
+		Expect(resp.GetIsMetric()).To(BeTrue(), "nested output must be metric")
+		Expect(len(resp.GetDepth())).To(Equal(int(resp.GetWidth())*int(resp.GetHeight())), "dense depth length")
+		Expect(len(resp.GetExtrinsics())).To(Equal(12), "extrinsics 3x4")
+		Expect(resp.GetIntrinsics()[0]).To(BeNumerically(">", 0), "fx > 0")
+
+		_, _ = fmt.Fprintf(GinkgoWriter, "nested depth OK: %dx%d is_metric=%v fx=%.2f\n",
+			resp.GetWidth(), resp.GetHeight(), resp.GetIsMetric(), resp.GetIntrinsics()[0])
+	})
+})

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

@@ -0,0 +1,20 @@
+package main
+
+import (
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+var _ = DescribeTable("optionValue",
+	func(opts []string, key, want string) {
+		Expect(optionValue(opts, key)).To(Equal(want))
+	},
+	Entry("present", []string{"foo:bar", "metric_model:m.gguf"}, "metric_model", "m.gguf"),
+	Entry("absent", []string{"foo:bar"}, "metric_model", ""),
+	Entry("nil", []string(nil), "metric_model", ""),
+	Entry("trims space", []string{"metric_model:  m.gguf  "}, "metric_model", "m.gguf"),
+	Entry("value with colon", []string{"metric_model:a:b.gguf"}, "metric_model", "a:b.gguf"),
+	Entry("first wins", []string{"metric_model:first.gguf", "metric_model:second.gguf"}, "metric_model", "first.gguf"),
+	Entry("empty value", []string{"metric_model:"}, "metric_model", ""),
+	Entry("prefix not key", []string{"metric_model_extra:x"}, "metric_model", ""),
+)

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

@@ -0,0 +1,59 @@
+#!/bin/bash
+
+# Script to copy the appropriate libraries based on architecture
+
+set -e
+
+CURDIR=$(dirname "$(realpath $0)")
+REPO_ROOT="${CURDIR}/../../.."
+
+# Create lib directory
+mkdir -p $CURDIR/package/lib
+
+cp -avf $CURDIR/libdepthanythingcpp-*.so $CURDIR/package/
+cp -avf $CURDIR/depth-anything-cpp $CURDIR/package/
+cp -fv $CURDIR/run.sh $CURDIR/package/
+
+# Detect architecture and copy appropriate libraries
+if [ -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
+    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
+    # ARM64 architecture
+    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 [ $(uname -s) = "Darwin" ]; then
+    echo "Detected Darwin"
+else
+    echo "Error: Could not detect architecture"
+    exit 1
+fi
+
+# Package GPU libraries based on BUILD_TYPE
+GPU_LIB_SCRIPT="${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/
+ls -liah $CURDIR/package/lib/

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

@@ -0,0 +1,52 @@
+#!/bin/bash
+set -ex
+
+# Get the absolute current dir where the script is located
+CURDIR=$(dirname "$(realpath $0)")
+
+cd /
+
+echo "CPU info:"
+if [ "$(uname)" != "Darwin" ]; then
+	grep -e "model\sname" /proc/cpuinfo | head -1
+	grep -e "flags" /proc/cpuinfo | head -1
+fi
+
+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 $CURDIR/libdepthanythingcpp-avx.so ]; then
+			LIBRARY="$CURDIR/libdepthanythingcpp-avx.so"
+		fi
+	fi
+
+	if grep -q -e "\savx2\s" /proc/cpuinfo ; then
+		echo "CPU:    AVX2   found OK"
+		if [ -e $CURDIR/libdepthanythingcpp-avx2.so ]; then
+			LIBRARY="$CURDIR/libdepthanythingcpp-avx2.so"
+		fi
+	fi
+
+	# Check avx 512
+	if grep -q -e "\savx512f\s" /proc/cpuinfo ; then
+		echo "CPU:    AVX512F found OK"
+		if [ -e $CURDIR/libdepthanythingcpp-avx512.so ]; then
+			LIBRARY="$CURDIR/libdepthanythingcpp-avx512.so"
+		fi
+	fi
+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
+	echo "Using lib/ld.so"
+	echo "Using library: $LIBRARY"
+	exec $CURDIR/lib/ld.so $CURDIR/depth-anything-cpp "$@"
+fi
+
+echo "Using library: $LIBRARY"
+exec $CURDIR/depth-anything-cpp "$@"

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

@@ -0,0 +1,45 @@
+#!/bin/bash
+set -e
+
+CURDIR=$(dirname "$(realpath $0)")
+
+echo "Running depth-anything-cpp backend tests..."
+
+# Test model from the mudler/depth-anything.cpp-gguf HuggingFace repo. The small
+# (vits) f32 GGUF is the lightest backbone (~131 MB), so it keeps the download
+# cheap. It is resumed with `curl -C -` and skipped entirely if already present.
+DEPTHANYTHING_MODEL_DIR="${DEPTHANYTHING_MODEL_DIR:-$CURDIR/test-models}"
+
+DEPTHANYTHING_MODEL_FILE="${DEPTHANYTHING_MODEL_FILE:-depth-anything-small-f32.gguf}"
+DEPTHANYTHING_MODEL_URL="${DEPTHANYTHING_MODEL_URL:-https://huggingface.co/mudler/depth-anything.cpp-gguf/resolve/main/depth-anything-small-f32.gguf}"
+
+mkdir -p "$DEPTHANYTHING_MODEL_DIR"
+
+if [ ! -f "$DEPTHANYTHING_MODEL_DIR/$DEPTHANYTHING_MODEL_FILE" ]; then
+    echo "Downloading depth-anything small f32 model (~131 MB)..."
+    # -C - resumes a partial download so an interrupted run doesn't restart from 0.
+    curl -L -C - -o "$DEPTHANYTHING_MODEL_DIR/$DEPTHANYTHING_MODEL_FILE" "$DEPTHANYTHING_MODEL_URL" --progress-bar
+fi
+
+# Use a real photo (people + cars) from the upstream rf-detr.cpp repo (~46 KB).
+# Depth estimation needs real content; a synthetic image would be degenerate.
+TEST_IMAGE_DIR="$CURDIR/test-data"
+TEST_IMAGE_FILE="$TEST_IMAGE_DIR/test.jpg"
+TEST_IMAGE_URL="${TEST_IMAGE_URL:-https://raw.githubusercontent.com/mudler/rf-detr.cpp/main/tests/fixtures/ci/test_image.jpg}"
+
+mkdir -p "$TEST_IMAGE_DIR"
+if [ ! -f "$TEST_IMAGE_FILE" ]; then
+    echo "Downloading test image..."
+    curl -L -o "$TEST_IMAGE_FILE" "$TEST_IMAGE_URL" --progress-bar
+fi
+
+echo "depth-anything-cpp test setup complete."
+echo "  model:      $DEPTHANYTHING_MODEL_DIR/$DEPTHANYTHING_MODEL_FILE"
+echo "  test image: $TEST_IMAGE_FILE"
+
+# Run the Go smoke test: spawns the backend binary on a free port, calls
+# LoadModel + Predict via gRPC against the downloaded GGUF + image.
+echo ""
+echo "Running Go smoke test..."
+cd "$CURDIR"
+go test -v -timeout 30m ./...

backend/go/dllm/.gitignore

@@ -1,10 +0,0 @@
-.cache/
-sources/
-build/
-package/
-dllm-grpc
-# build artifacts staged in-tree by the Makefile (cp from sources/) or
-# symlinked for local dev; the real sources live in dllm.cpp upstream.
-*.so
-*.so.*
-compile_commands.json

backend/go/dllm/Makefile

@@ -1,101 +0,0 @@
-# dllm backend Makefile.
-#
-# Upstream pin lives below as DLLM_VERSION?=<sha> so .github/bump_deps.sh
-# can find and update it - matches the whisper.cpp / parakeet-cpp / ds4
-# convention.
-#
-# Local dev shortcut: if you already have an out-of-tree dllm.cpp build,
-# you can symlink the .so into this directory and skip the clone/cmake
-# steps entirely, e.g.:
-#
-#   ln -sf /path/to/dllm.cpp/build/libdllm.so .
-#   go build -o dllm-grpc .
-#
-# That's what the gated C-ABI binding smoke uses (DLLM_TEST_LIBRARY). The
-# default target below does the proper clone-at-pin + cmake build so CI
-# doesn't need a side-checkout.
-#
-# NOTE: github.com/mudler/dllm.cpp is still private (publishing is planned);
-# until then the anonymous clone below fails. Use the symlink shortcut above
-# with a local checkout, or a git credential helper with access to the repo.
-
-# The pin below is the P5 performance-parity head (device-resident
-# self-conditioning, full-GPU placement at ngl >= n_layer, graph reuse,
-# device-side EB reductions: ~8x per-step on GB10, see dllm.cpp
-# docs/validation.md section 10). C-ABI unchanged (still version 1). It
-# also carries the multimodal entry points (dllm_capi_generate_mm /
-# dllm_capi_generate_stream_mm) the image-input path probes for; older
-# libs still load, but image requests then fail with "library predates
-# the multimodal entry points".
-DLLM_VERSION?=320b57756efc3460169b8ea9e8c782867198f2a5
-DLLM_REPO?=https://github.com/mudler/dllm.cpp
-
-GOCMD?=go
-GO_TAGS?=
-JOBS?=$(shell nproc 2>/dev/null || sysctl -n hw.ncpu 2>/dev/null || echo 4)
-
-BUILD_TYPE?=
-NATIVE?=false
-
-# libdllm.so is self-contained: dllm.cpp's CMakeLists statically absorbs ggml
-# (BUILD_SHARED_LIBS=OFF + PIC) into the shared lib, so dlopen needs no
-# libggml*.so alongside it, only system libs (libstdc++/libgomp/libc) the
-# runtime image already provides. Tests/CLI are upstream-only concerns.
-CMAKE_ARGS?=-DCMAKE_BUILD_TYPE=Release -DDLLM_BUILD_TESTS=OFF
-
-ifeq ($(NATIVE),false)
-	CMAKE_ARGS+=-DGGML_NATIVE=OFF
-endif
-
-# Same arch set the sibling ggml backends (acestep/vibevoice/qwen3-tts) bake
-# for their cublas images; override for a native build.
-CUDA_ARCHITECTURES?=75-virtual;80-virtual;86-real;89-real
-
-# dllm.cpp gates CUDA behind DLLM_CUDA (set(GGML_CUDA ... CACHE FORCE)), so
-# forward that instead of a bare -DGGML_CUDA=ON.
-ifeq ($(BUILD_TYPE),cublas)
-	CMAKE_ARGS+=-DDLLM_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES="$(CUDA_ARCHITECTURES)"
-endif
-
-.PHONY: dllm-grpc package build clean purge test all
-
-all: dllm-grpc
-
-# Clone the upstream dllm.cpp source at the pinned commit (ggml comes in as
-# a submodule). Directory acts as the target so make only re-clones when
-# missing. After a DLLM_VERSION bump, run 'make purge && make' to refetch.
-sources/dllm.cpp:
-	mkdir -p sources/dllm.cpp
-	cd sources/dllm.cpp && \
-	git init -q && \
-	git remote add origin $(DLLM_REPO) && \
-	git fetch --depth 1 origin $(DLLM_VERSION) && \
-	git checkout FETCH_HEAD && \
-	git submodule update --init --recursive --depth 1 --single-branch
-
-# Build the shared lib out-of-tree, then stage it next to the Go sources so
-# purego.Dlopen("libdllm.so") and the packaging step both pick it up.
-libdllm.so: sources/dllm.cpp
-	cmake -B sources/dllm.cpp/build -S sources/dllm.cpp $(CMAKE_ARGS)
-	cmake --build sources/dllm.cpp/build --config Release -j$(JOBS)
-	cp -fv sources/dllm.cpp/build/libdllm.so ./
-
-dllm-grpc: libdllm.so main.go capi.go
-	CGO_ENABLED=0 $(GOCMD) build -tags "$(GO_TAGS)" -o dllm-grpc .
-
-package: dllm-grpc
-	bash package.sh
-
-build: package
-
-# Test target. The C-ABI binding smoke is gated on DLLM_TEST_LIBRARY +
-# DLLM_TEST_TINY_MODEL; without them the gated specs auto-skip and only the
-# pure-Go helper specs run.
-test:
-	LD_LIBRARY_PATH=$(CURDIR):$$LD_LIBRARY_PATH $(GOCMD) test ./... -count=1
-
-clean: purge
-	rm -rf libdllm.so* package dllm-grpc
-
-purge:
-	rm -rf sources/dllm.cpp

backend/go/dllm/capi.go

@@ -1,326 +0,0 @@
-package main
-
-// Typed Go wrappers over dllm.cpp's flat C-ABI (include/dllm_capi.h, ABI v1).
-//
-// Contract highlights the wrappers encode (see the header + src/capi.cpp):
-//   - tokenize_json/generate return malloc'd char* the CALLER owns: bound as
-//     uintptr, copied with goStringFromCPtr, released via dllm_capi_free_string.
-//   - last_error returns a BORROWED pointer (valid until the next call on the
-//     same ctx): bound as a plain string (purego copies), never freed, and only
-//     read AFTER the failing call has returned - reading it while a generate is
-//     in flight on the same ctx violates the per-ctx serialization contract.
-//   - All entry points except dllm_capi_cancel must be externally serialized
-//     per ctx (one ctx = one concurrent generate/tokenize). Cancel only flips
-//     an atomic and may be called from any goroutine mid-generate.
-//   - No C++ exception crosses the boundary; failures land in last_error.
-
-import (
-	"encoding/json"
-	"errors"
-	"fmt"
-	"sync"
-	"sync/atomic"
-	"unsafe"
-
-	"github.com/ebitengine/purego"
-)
-
-// dllmABIVersion is the DLLM_CAPI_ABI_VERSION this binding was written
-// against; main.go refuses to start against a libdllm.so reporting another.
-const dllmABIVersion = 1
-
-// purego-bound entry points from libdllm.so. Names match dllm_capi.h
-// exactly; loadCAPI (main.go) fills these in at boot.
-var (
-	cppAbiVersion func() int32
-	cppLoad       func(ggufPath, paramsJSON string) uintptr
-	cppFree       func(ctx uintptr)
-	cppLastError  func(ctx uintptr) string // borrowed pointer: purego copies, do NOT free
-	cppFreeString func(s uintptr)
-	// malloc'd char* returns, hence uintptr (see loadCAPI's doc comment).
-	cppTokenizeJSON func(ctx uintptr, text string) uintptr
-	cppGenerate     func(ctx uintptr, prompt, optsJSON string) uintptr
-	// on_block/on_step are C function pointers produced by purego.NewCallback;
-	// userData carries the streamCallStates registry key.
-	cppGenerateStream func(ctx uintptr, prompt, optsJSON string, onBlock, onStep, userData uintptr) int32
-	cppCancel         func(ctx uintptr)
-)
-
-// Optional multimodal entry points (dllm_capi.h's P4 surface). The ABI
-// version stays 1: presence is detected by PROBING the symbols with Dlsym at
-// boot (loadCAPI, mirroring the parakeet-cpp optional-symbol pattern). nil
-// means the loaded libdllm.so predates the mm surface; the wrappers below
-// then fail with errMMUnsupported instead of crashing on a nil call.
-var (
-	cppGenerateMM       func(ctx uintptr, prompt, imagesJSON, optsJSON string) uintptr
-	cppGenerateStreamMM func(ctx uintptr, prompt, imagesJSON, optsJSON string, onBlock, onStep, userData uintptr) int32
-)
-
-// mmImageMarker is the literal placeholder dllm_capi_generate_mm expands to
-// <boi> + soft-token placeholders + <eoi> (dllm_capi.h placeholder contract;
-// capi.cpp MM_MARKER). The prompt must carry exactly one marker per
-// images_json entry, in image order.
-const mmImageMarker = "<image>"
-
-// errMMUnsupported is returned for image-bearing requests against an old
-// text-only libdllm.so (the Dlsym probe found no mm symbols).
-var errMMUnsupported = errors.New(
-	"dllm: image input requires libdllm.so with the multimodal entry points (dllm_capi_generate_mm), but the loaded library predates them - rebuild/upgrade the dllm backend to use images")
-
-// cMMSupported reports whether the loaded libdllm.so carries the multimodal
-// generate pair. Both symbols ship together (same dllm.cpp commit), but the
-// guard requires both anyway so a half-present surface can never dispatch.
-func cMMSupported() bool {
-	return cppGenerateMM != nil && cppGenerateStreamMM != nil
-}
-
-// cAbiVersion returns the library's DLLM_CAPI_ABI_VERSION.
-func cAbiVersion() int32 {
-	return cppAbiVersion()
-}
-
-// cLoad opens the GGUF at path with the flat params JSON (e.g.
-// {"n_gpu_layers":99}). Returns 0 on failure; per the header contract there
-// is no ctx to carry the reason, the C side logs it to stderr (and
-// cLastError(0) only yields the static NULL-ctx message).
-func cLoad(path, paramsJSON string) uintptr {
-	return cppLoad(path, paramsJSON)
-}
-
-// cFree releases a ctx; safe on 0 (delete nullptr).
-func cFree(h uintptr) {
-	cppFree(h)
-}
-
-// cLastError returns the ctx's last error message (or the static NULL-ctx
-// message for h==0). The C pointer is borrowed and only valid until the next
-// call on the same ctx; purego's string return copies it immediately, so the
-// returned Go string is safe to keep. Must not be called while another call
-// on the same ctx is in flight.
-func cLastError(h uintptr) string {
-	return cppLastError(h)
-}
-
-// lastErrorOr is cLastError with a fallback for the empty-message case, so
-// wrapped errors never end in ": ".
-func lastErrorOr(h uintptr, fallback string) string {
-	if msg := cLastError(h); msg != "" {
-		return msg
-	}
-	return fallback
-}
-
-// cTokenizeJSON tokenizes text (the C side prepends bos per vocab.add_bos)
-// and returns the token ids as a JSON array string, e.g. "[2,18]".
-func cTokenizeJSON(h uintptr, text string) (string, error) {
-	ret := cppTokenizeJSON(h, text)
-	if ret == 0 {
-		return "", fmt.Errorf("dllm: tokenize failed: %s", lastErrorOr(h, "unknown error"))
-	}
-	out := goStringFromCPtr(ret)
-	cppFreeString(ret)
-	return out, nil
-}
-
-// cGenerate runs a blocking generation and returns the detokenized text.
-// optsJSON must be a FLAT JSON object of scalars (use buildOptsJSON); the C
-// parser rejects nested objects/arrays. NULL return -> last_error (read only
-// after the call returned, per the serialization contract); a cancelled call
-// surfaces as the "cancelled" message.
-func cGenerate(h uintptr, prompt, optsJSON string) (string, error) {
-	ret := cppGenerate(h, prompt, optsJSON)
-	if ret == 0 {
-		return "", fmt.Errorf("dllm: generate failed: %s", lastErrorOr(h, "unknown error"))
-	}
-	out := goStringFromCPtr(ret)
-	cppFreeString(ret)
-	return out, nil
-}
-
-// cGenerateMM is cGenerate's multimodal counterpart. imagesJSON is the flat
-// JSON array of image entries (data: base64 URIs here; the C side also takes
-// file paths) and the prompt must carry one mmImageMarker per entry - the
-// engine enforces the 1:1 match and reports mismatches through last_error.
-func cGenerateMM(h uintptr, prompt, imagesJSON, optsJSON string) (string, error) {
-	if !cMMSupported() {
-		return "", errMMUnsupported
-	}
-	ret := cppGenerateMM(h, prompt, imagesJSON, optsJSON)
-	if ret == 0 {
-		return "", fmt.Errorf("dllm: generate_mm failed: %s", lastErrorOr(h, "unknown error"))
-	}
-	out := goStringFromCPtr(ret)
-	cppFreeString(ret)
-	return out, nil
-}
-
-// streamCallState carries the Go callbacks for one in-flight
-// cGenerateStream call; the registry key travels through C as user_data.
-// The map shape mirrors the whisper backend's streamCallStates: only one
-// entry per ctx is ever live (the C-ABI is serialized per ctx), but keying
-// by call survives multiple models/processes sharing the package.
-type streamCallState struct {
-	onBlock func(text string)
-	onStep  func(step, total int, preview string)
-}
-
-var (
-	streamCallStates sync.Map // uint64 -> *streamCallState
-	streamCallSeq    atomic.Uint64
-
-	// purego.NewCallback allocates a finite, never-released callback slot, so
-	// the two trampolines are created exactly once and reused across calls.
-	streamCbOnce sync.Once
-	blockCbPtr   uintptr
-	stepCbPtr    uintptr
-)
-
-// onBlockTrampoline is the Go side of dllm_block_cb. It runs on the C
-// calling thread, mid-generate: keep it tiny and non-blocking (callers that
-// bridge to goroutines must hand off via buffered channels). The text
-// pointer is only valid for the duration of the invocation, so it is copied
-// to a Go string immediately.
-func onBlockTrampoline(text uintptr, userData uintptr) {
-	v, ok := streamCallStates.Load(uint64(userData))
-	if !ok {
-		return // call already torn down
-	}
-	state := v.(*streamCallState)
-	if state.onBlock != nil {
-		state.onBlock(goStringFromCPtr(text))
-	}
-}
-
-// onStepTrampoline is the Go side of dllm_step_cb; same threading and
-// lifetime caveats as onBlockTrampoline.
-func onStepTrampoline(step int32, totalSteps int32, canvasPreview uintptr, userData uintptr) {
-	v, ok := streamCallStates.Load(uint64(userData))
-	if !ok {
-		return
-	}
-	state := v.(*streamCallState)
-	if state.onStep != nil {
-		state.onStep(int(step), int(totalSteps), goStringFromCPtr(canvasPreview))
-	}
-}
-
-// withStreamCallbacks registers onBlock/onStep in the trampoline registry
-// for the duration of one streaming C call and invokes call with the C
-// function pointers (NULL for absent callbacks, so the C side skips the
-// per-block / per-step detokenize work entirely) plus the registry key to
-// pass as user_data. Shared by the text and multimodal stream wrappers.
-func withStreamCallbacks(onBlock func(text string), onStep func(step, total int, preview string), call func(blockPtr, stepPtr, userData uintptr) int32) int32 {
-	streamCbOnce.Do(func() {
-		blockCbPtr = purego.NewCallback(onBlockTrampoline)
-		stepCbPtr = purego.NewCallback(onStepTrampoline)
-	})
-
-	id := streamCallSeq.Add(1)
-	streamCallStates.Store(id, &streamCallState{onBlock: onBlock, onStep: onStep})
-	defer streamCallStates.Delete(id)
-
-	var blockPtr, stepPtr uintptr
-	if onBlock != nil {
-		blockPtr = blockCbPtr
-	}
-	if onStep != nil {
-		stepPtr = stepCbPtr
-	}
-	return call(blockPtr, stepPtr, uintptr(id))
-}
-
-// cGenerateStream runs a generation with per-committed-block (onBlock) and
-// per-denoising-step (onStep) callbacks; either may be nil. The callbacks
-// run on the C thread (see the trampoline docs). Returns an error carrying
-// last_error on failure; cancellation surfaces as the "cancelled" message.
-func cGenerateStream(h uintptr, prompt, optsJSON string, onBlock func(text string), onStep func(step, total int, preview string)) error {
-	rc := withStreamCallbacks(onBlock, onStep, func(blockPtr, stepPtr, userData uintptr) int32 {
-		return cppGenerateStream(h, prompt, optsJSON, blockPtr, stepPtr, userData)
-	})
-	if rc != 0 {
-		return fmt.Errorf("dllm: generate_stream failed: %s", lastErrorOr(h, "unknown error"))
-	}
-	return nil
-}
-
-// cGenerateStreamMM is cGenerateStream's multimodal counterpart; see
-// cGenerateMM for the imagesJSON/marker contract.
-func cGenerateStreamMM(h uintptr, prompt, imagesJSON, optsJSON string, onBlock func(text string), onStep func(step, total int, preview string)) error {
-	if !cMMSupported() {
-		return errMMUnsupported
-	}
-	rc := withStreamCallbacks(onBlock, onStep, func(blockPtr, stepPtr, userData uintptr) int32 {
-		return cppGenerateStreamMM(h, prompt, imagesJSON, optsJSON, blockPtr, stepPtr, userData)
-	})
-	if rc != 0 {
-		return fmt.Errorf("dllm: generate_stream_mm failed: %s", lastErrorOr(h, "unknown error"))
-	}
-	return nil
-}
-
-// cCancel requests cancellation of the in-flight generate on h. This is the
-// ONE entry point safe to call from any goroutine while a generate runs (it
-// only flips an atomic). Note the cancel-reset race from the header: each
-// generate resets the flag on entry, so a watchdog should re-issue cancel if
-// the call has not returned.
-func cCancel(h uintptr) {
-	cppCancel(h)
-}
-
-// buildOptsJSON renders generation options as the flat JSON object the
-// C-ABI expects (known keys: n_predict, blocks, seed, eb_*, kv_cache). The
-// C-side scanner only understands scalar number/string values and rejects
-// nested objects/arrays loudly; bools are rejected here too because the
-// scanner has no concept of them. Fail loud rather than let an option be
-// silently misread.
-//
-// CAVEAT: json.Marshal HTML-escapes <, > and & inside string values (e.g.
-// "<" becomes the six-byte \u003c sequence). None of the known string-valued keys
-// (kv_cache: auto|on|off) can contain those bytes today; if one ever does,
-// switch to an Encoder with SetEscapeHTML(false) like gemma4JSONString.
-func buildOptsJSON(opts map[string]any) (string, error) {
-	if len(opts) == 0 {
-		return "{}", nil
-	}
-	for k, v := range opts {
-		switch v.(type) {
-		case string,
-			int, int8, int16, int32, int64,
-			uint, uint8, uint16, uint32, uint64,
-			float32, float64,
-			json.Number:
-			// scalar: fine
-		default:
-			return "", fmt.Errorf("dllm: opts key %q has non-scalar value %T (the C-ABI only accepts flat number/string scalars)", k, v)
-		}
-	}
-	b, err := json.Marshal(opts)
-	if err != nil {
-		return "", fmt.Errorf("dllm: marshal opts: %w", err)
-	}
-	return string(b), nil
-}
-
-// goStringFromCPtr copies a NUL-terminated C string into Go memory. cptr is
-// the raw pointer returned by purego from the C-ABI (a malloc'd buffer the
-// caller owns, or a callback argument only valid during the invocation);
-// owning callers must free it via cppFreeString after the copy lands.
-//
-// A direct unsafe.Pointer(cptr) conversion trips go vet's unsafeptr check,
-// which can't distinguish a C-owned heap pointer from Go-managed memory (the
-// parakeet-cpp and whisper backends tolerate that warning). Reinterpreting
-// through &cptr below is equivalent at runtime and keeps plain `go vet`
-// clean. It is safe either way: the pointer addresses C memory the Go GC
-// neither tracks nor moves, and we dereference it immediately to copy the
-// bytes out.
-func goStringFromCPtr(cptr uintptr) string {
-	if cptr == 0 {
-		return ""
-	}
-	p := *(*unsafe.Pointer)(unsafe.Pointer(&cptr)) // C-owned 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/dllm/dllm.go

@@ -1,622 +0,0 @@
-package main
-
-// LocalAI gRPC backend for dllm.cpp (DiffusionGemma block-diffusion models).
-//
-// Wiring overview:
-//   - Load opens the GGUF via dllm_capi_load and starts the per-model worker
-//     goroutine that serializes every C call (see submit).
-//   - PredictRich / PredictStreamRich implement grpc.AIModelRich: when the
-//     request carries raw messages (use_tokenizer_template), the backend owns
-//     templating (RenderGemma4) and output parsing (Gemma4Parser) and replies
-//     with ChatDeltas, like the llama.cpp autoparser and the ds4 backend.
-//   - The legacy Predict / PredictStream methods delegate to the rich pair
-//     (cloud-proxy precedent); the gRPC server prefers the rich path anyway.
-
-import (
-	"encoding/json"
-	"errors"
-	"fmt"
-	"strconv"
-	"strings"
-	"sync"
-	"unicode/utf8"
-
-	grpc "github.com/mudler/LocalAI/pkg/grpc"
-	"github.com/mudler/LocalAI/pkg/grpc/base"
-	"github.com/mudler/LocalAI/pkg/grpc/grpcerrors"
-	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
-	"github.com/mudler/xlog"
-)
-
-// The gRPC server cancels in-flight generations on client disconnect only
-// for backends advertising the Cancellable capability; keep Dllm pinned to
-// it so a signature drift fails the build, not the disconnect path.
-var _ grpc.Cancellable = (*Dllm)(nil)
-
-// generator is the seam between the backend wiring and the dllm.cpp C-ABI:
-// the real implementation (capiGenerator) wraps the cGenerate/cTokenizeJSON
-// family, while tests substitute a fake to exercise prompt construction,
-// parsing and serialization without libdllm.so.
-type generator interface {
-	generate(prompt, optsJSON string) (string, error)
-	// generateStream invokes onBlock once per committed diffusion block, on
-	// the thread running the C call, before returning.
-	generateStream(prompt, optsJSON string, onBlock func(text string)) error
-	// generateMM / generateStreamMM are the multimodal counterparts:
-	// imagesJSON is a flat JSON array of data: base64 URIs and the prompt
-	// carries one mmImageMarker per entry (dllm_capi.h placeholder
-	// contract). Against an old text-only libdllm.so they fail with
-	// errMMUnsupported.
-	generateMM(prompt, imagesJSON, optsJSON string) (string, error)
-	generateStreamMM(prompt, imagesJSON, optsJSON string, onBlock func(text string)) error
-	tokenizeJSON(text string) (string, error)
-	// cancel is the ONE entry point safe to call concurrently with an
-	// in-flight generate on the same ctx (dllm_capi.h: it only flips an
-	// atomic; everything else must be externally serialized per ctx).
-	cancel()
-	free()
-}
-
-// capiGenerator is the production generator over one dllm_ctx handle.
-type capiGenerator struct {
-	h uintptr
-}
-
-func (g *capiGenerator) generate(prompt, optsJSON string) (string, error) {
-	return cGenerate(g.h, prompt, optsJSON)
-}
-
-func (g *capiGenerator) generateStream(prompt, optsJSON string, onBlock func(text string)) error {
-	// on_step (per-denoise-step canvas preview, dllm.cpp's --visual) is
-	// passed as nil for now: a future progress hook for the React UI can
-	// plumb it through without touching the C binding.
-	return cGenerateStream(g.h, prompt, optsJSON, onBlock, nil)
-}
-
-func (g *capiGenerator) generateMM(prompt, imagesJSON, optsJSON string) (string, error) {
-	return cGenerateMM(g.h, prompt, imagesJSON, optsJSON)
-}
-
-func (g *capiGenerator) generateStreamMM(prompt, imagesJSON, optsJSON string, onBlock func(text string)) error {
-	// on_step is nil for the same reason as generateStream.
-	return cGenerateStreamMM(g.h, prompt, imagesJSON, optsJSON, onBlock, nil)
-}
-
-func (g *capiGenerator) tokenizeJSON(text string) (string, error) {
-	return cTokenizeJSON(g.h, text)
-}
-
-func (g *capiGenerator) cancel() {
-	cCancel(g.h)
-}
-
-func (g *capiGenerator) free() {
-	cFree(g.h)
-}
-
-// Dllm is the gRPC backend instance: one per loaded model (LocalAI starts
-// one backend process per model).
-type Dllm struct {
-	base.Base
-
-	gen generator
-	// genOpts holds the model-level generation overrides parsed from
-	// ModelOptions.Options at Load (eb_*, blocks, kv_cache). The C-ABI takes
-	// them per-generate, not per-load, so they are merged into every
-	// request's opts JSON (requestOptsJSON).
-	genOpts map[string]any
-
-	// jobs is the per-model worker queue. dllm_capi.h requires every entry
-	// point EXCEPT dllm_capi_cancel to be externally serialized per ctx (one
-	// ctx = one concurrent generate/tokenize; last_error is unsafe to read
-	// while a call is in flight). A single goroutine owning all C calls makes
-	// that contract structural instead of relying on lock discipline.
-	jobs     chan func()
-	workerWG sync.WaitGroup
-
-	// genMu guards gen against Free racing in-flight requests: requests hold
-	// the read lock for their full duration (they stay concurrent with each
-	// other - the worker still serializes the C calls), Free takes the write
-	// lock so it can only run when no request is in flight.
-	genMu sync.RWMutex
-}
-
-func (d *Dllm) startWorker() {
-	d.jobs = make(chan func())
-	d.workerWG.Add(1)
-	go func() {
-		defer d.workerWG.Done()
-		for job := range d.jobs {
-			job()
-		}
-	}()
-}
-
-// submit runs job on the worker goroutine and waits for it to finish.
-// Concurrent gRPC requests therefore queue up and execute one at a time
-// against the single dllm_ctx.
-func (d *Dllm) submit(job func()) {
-	done := make(chan struct{})
-	d.jobs <- func() {
-		defer close(done)
-		job()
-	}
-	<-done
-}
-
-// Load opens the GGUF and prepares the worker. Load-time engine parameters
-// travel as the flat params JSON of dllm_capi_load; generation overrides
-// from Options are stored for per-request opts JSON instead (the C-ABI has
-// no per-load sampler state).
-func (d *Dllm) Load(opts *pb.ModelOptions) error {
-	if d.gen != nil {
-		return errors.New("dllm: model already loaded")
-	}
-
-	params := map[string]any{
-		"n_gpu_layers": opts.GetNGPULayers(),
-	}
-	if opts.GetThreads() > 0 {
-		params["n_threads"] = opts.GetThreads()
-	}
-	if opts.GetContextSize() > 0 {
-		params["ctx_len"] = opts.GetContextSize()
-	}
-	paramsJSON, err := buildOptsJSON(params)
-	if err != nil {
-		return err
-	}
-
-	d.genOpts = parseModelGenOpts(opts.GetOptions())
-
-	h := cLoad(opts.GetModelFile(), paramsJSON)
-	if h == 0 {
-		// No ctx exists on load failure, so last_error(NULL) only carries the
-		// static NULL-ctx message; the real reason is on the backend's stderr.
-		return fmt.Errorf("dllm: load %q failed: %s (see backend log for details)",
-			opts.GetModelFile(), lastErrorOr(0, "unknown error"))
-	}
-	d.gen = &capiGenerator{h: h}
-	d.startWorker()
-	xlog.Info("dllm: model loaded", "model", opts.GetModelFile(), "params", paramsJSON, "gen_opts", d.genOpts)
-	return nil
-}
-
-// Free releases the dllm ctx and stops the worker. Safe when never loaded.
-//
-// The write lock is essential: the gRPC server (pkg/grpc/server.go, see the
-// model-unload path around line 764) calls Free with no locking of its own,
-// and base.Base provides none either. Without it a request racing Free would
-// panic sending on the closed jobs channel - or worse, generate on a freed C
-// ctx. Holding genMu until gen is nil also turns post-Free requests into a
-// clean "model not loaded" error instead of a crash.
-func (d *Dllm) Free() error {
-	d.genMu.Lock()
-	defer d.genMu.Unlock()
-	if d.gen == nil {
-		return nil
-	}
-	d.submit(d.gen.free)
-	close(d.jobs)
-	d.workerWG.Wait()
-	d.gen = nil
-	return nil
-}
-
-// Cancel requests cancellation of the in-flight generate (the
-// grpc.Cancellable capability). The gRPC server arms it via
-// context.AfterFunc on the request/stream context, so a client
-// disconnect or timeout aborts the generation server-side - the same
-// semantics the llama.cpp C++ backend gets from polling IsCancelled().
-// It deliberately bypasses the worker queue: dllm_capi_cancel is the one
-// call the C-ABI allows from any goroutine mid-generate (it only flips
-// an atomic).
-//
-// Note dllm_capi.h's cancel-reset race: each generate resets the flag on
-// entry, so a Cancel racing a NEW generate on the same ctx can be lost
-// (and, with requests queued on the worker, it aborts whichever generate
-// is currently running). The single-flag granularity is acceptable here
-// because the server de-registers the hook on normal completion and one
-// backend process serves one model.
-func (d *Dllm) Cancel() {
-	// RLock so a server-side AfterFunc firing in the window between a
-	// request finishing and a model unload cannot touch a freed C ctx
-	// (Free holds the write lock while tearing gen down). cancel() is the
-	// one C call that is safe concurrently with an in-flight generate, so
-	// taking a read lock here cannot deadlock against request holders.
-	d.genMu.RLock()
-	defer d.genMu.RUnlock()
-	if d.gen != nil {
-		d.gen.cancel()
-	}
-}
-
-// dllmGenOptKeys are the ModelOptions.Options keys this backend forwards to
-// the engine. Options is a shared free-form bag (other layers put their own
-// entries there), so unknown keys are skipped with a warning, not an error.
-var dllmGenOptKeys = map[string]bool{
-	"blocks":   true,
-	"kv_cache": true, // "auto"|"on"|"off"; honored by the engine from P3
-}
-
-// parseModelGenOpts parses "key:value" Options entries into the flat scalar
-// map merged into every generate's opts JSON. eb_* (Entropy-Bound sampler
-// knobs) and the keys in dllmGenOptKeys are recognized; values are typed by
-// first successful parse (int, then float, else string) to match the C
-// scanner's number/string scalars.
-func parseModelGenOpts(options []string) map[string]any {
-	out := map[string]any{}
-	for _, o := range options {
-		key, val, found := strings.Cut(o, ":")
-		if !found {
-			xlog.Warn("dllm: ignoring malformed option (want key:value)", "option", o)
-			continue
-		}
-		if !strings.HasPrefix(key, "eb_") && !dllmGenOptKeys[key] {
-			xlog.Debug("dllm: ignoring unrecognized option", "key", key)
-			continue
-		}
-		out[key] = parseScalarOpt(val)
-	}
-	return out
-}
-
-func parseScalarOpt(v string) any {
-	if iv, err := strconv.ParseInt(v, 10, 64); err == nil {
-		return iv
-	}
-	if fv, err := strconv.ParseFloat(v, 64); err == nil {
-		return fv
-	}
-	return v
-}
-
-// metadataEnableThinking reads the enable_thinking gate. Unlike ds4 (default
-// ON, matching ds4-server), dllm defaults OFF: DiffusionGemma's chat
-// template guards every thinking branch with `enable_thinking is defined and
-// enable_thinking`, i.e. thinking is opt-in for this model family, and the
-// no-thinking render pre-closes an empty thought channel that the OFF
-// default must produce.
-func metadataEnableThinking(opts *pb.PredictOptions) bool {
-	v := opts.GetMetadata()["enable_thinking"]
-	return v == "true" || v == "1"
-}
-
-// buildPrompt resolves the prompt for a request. With use_tokenizer_template
-// and raw messages the backend owns templating (RenderGemma4, including the
-// mmImageMarker injection for opts.Images) and the output is in the known
-// gemma4 format, so parse=true. Without it the caller templated the prompt
-// themselves (LocalAI's Go templates + PEG fallback, or a bare completion):
-// the prompt passes through verbatim - for image requests it must already
-// carry one literal mmImageMarker per image (the engine enforces the 1:1
-// match) - and the output is NOT gemma4-parsed - it is emitted as plain
-// content and the Go side's extraction applies, as for any non-autoparsing
-// backend.
-func buildPrompt(opts *pb.PredictOptions) (prompt string, parse bool, err error) {
-	if opts.GetUseTokenizerTemplate() && len(opts.GetMessages()) > 0 {
-		prompt, err = RenderGemma4(opts.GetMessages(), opts.GetTools(), len(opts.GetImages()), metadataEnableThinking(opts), true)
-		return prompt, true, err
-	}
-	return opts.GetPrompt(), false, nil
-}
-
-// imagesJSON renders opts.Images as the flat JSON array of data: URIs the mm
-// C-ABI expects, or "" when the request carries no images. The entries arrive
-// as RAW base64 payloads: LocalAI's OpenAI layer decodes every image_url /
-// image content part (URL download or data: URI) to plain base64 via
-// utils.GetContentURIAsBase64 (core/http/middleware/request.go) and core
-// flattens them into PredictOptions.Images (core/backend/llm.go). The
-// hardcoded image/jpeg mime mirrors the llama.cpp backend's re-wrapping
-// convention (grpc-server.cpp, "data:image/jpeg;base64," + images(i)); the
-// engine ignores the declared mime and sniffs the real format from the
-// decoded bytes (stb_image), so PNG/BMP payloads work through it too.
-func imagesJSON(images []string) (string, error) {
-	if len(images) == 0 {
-		return "", nil
-	}
-	uris := make([]string, len(images))
-	for i, img := range images {
-		// dllm_capi.h: array entries are read VERBATIM up to the closing
-		// quote, with NO escape handling. json.Marshal would escape these
-		// bytes and the C side would misparse the entry, so fail loud (they
-		// can never appear in genuine base64 anyway).
-		if strings.ContainsAny(img, "\"\\") {
-			return "", fmt.Errorf("dllm: image %d is not base64 (contains a quote or backslash; PredictOptions.Images entries must be raw base64 payloads)", i)
-		}
-		uris[i] = "data:image/jpeg;base64," + img
-	}
-	b, err := json.Marshal(uris)
-	if err != nil {
-		return "", fmt.Errorf("dllm: marshal images: %w", err)
-	}
-	return string(b), nil
-}
-
-// requestOptsJSON merges the model-level overrides with the request's
-// sampling fields into the flat opts JSON for one generate call.
-func (d *Dllm) requestOptsJSON(opts *pb.PredictOptions) (string, error) {
-	m := make(map[string]any, len(d.genOpts)+2)
-	for k, v := range d.genOpts {
-		m[k] = v
-	}
-	if n := opts.GetTokens(); n > 0 {
-		// The engine rounds n_predict UP to a whole number of diffusion
-		// blocks (the canvas is denoised block-wise), so the completion may
-		// run slightly past the requested budget. Tokens==0 omits the key so
-		// the C-ABI default of 256 applies (hardcoded in capi.cpp's
-		// parse_gen_opts, independent of canvas_length).
-		m["n_predict"] = n
-	}
-	if s := opts.GetSeed(); s > 0 {
-		// The engine seeds mt19937 with explicit non-negative seeds. Seed<=0
-		// is omitted: proto3 cannot distinguish 0 from unset, and negative
-		// values conventionally mean "random" across LocalAI backends.
-		m["seed"] = s
-	}
-	return buildOptsJSON(m)
-}
-
-// prepareRequest is the shared prologue of the rich methods: resolve the
-// prompt (and whether the output gets gemma4-parsed) and build the per-call
-// opts JSON plus the images JSON ("" for text-only requests, which routes
-// the call through the text generate entry points).
-func (d *Dllm) prepareRequest(opts *pb.PredictOptions) (prompt string, parse bool, optsJSON, imgJSON string, err error) {
-	// Fail loud on media the engine has no path for, instead of silently
-	// generating from a prompt that ignores them.
-	if len(opts.GetVideos()) > 0 || len(opts.GetAudios()) > 0 {
-		return "", false, "", "", errors.New("dllm: video/audio input is not supported (images only)")
-	}
-	prompt, parse, err = buildPrompt(opts)
-	if err != nil {
-		return "", false, "", "", err
-	}
-	optsJSON, err = d.requestOptsJSON(opts)
-	if err != nil {
-		return "", false, "", "", err
-	}
-	imgJSON, err = imagesJSON(opts.GetImages())
-	if err != nil {
-		return "", false, "", "", err
-	}
-	return prompt, parse, optsJSON, imgJSON, nil
-}
-
-// sanitizeUTF8 makes s safe for a proto3 string field. Block-boundary
-// detokenization and byte-fallback tokens can produce invalid UTF-8, and
-// grpc-go refuses to marshal it ("string field contains invalid UTF-8"), so
-// every string destined for a Reply/ChatDelta must pass through here (or
-// through splitValidUTF8, which calls it). Lone malformed bytes are genuinely
-// undecodable: replace with U+FFFD rather than crash the stream.
-func sanitizeUTF8(s string) string {
-	if utf8.ValidString(s) {
-		return s
-	}
-	return strings.ToValidUTF8(s, "<22>")
-}
-
-// utf8SeqLen returns the declared sequence length of a UTF-8 leading byte
-// (1 for bytes that can never lead a multi-byte sequence, so they are never
-// held back and fall through to sanitizeUTF8's replacement).
-func utf8SeqLen(b byte) int {
-	switch {
-	case b&0xE0 == 0xC0:
-		return 2
-	case b&0xF0 == 0xE0:
-		return 3
-	case b&0xF8 == 0xF0:
-		return 4
-	default:
-		return 1
-	}
-}
-
-// splitValidUTF8 prepends the previous block's carry to the new block and
-// splits the result into text safe to emit now and a trailing INCOMPLETE
-// UTF-8 sequence (at most utf8.UTFMax-1 bytes) to carry into the next block:
-// the per-block detokenize can split a multi-byte character across block
-// boundaries (llama.cpp's grpc-server holds back the same way). Only a
-// suffix that can still become a valid rune is withheld; bytes that are
-// already undecodable are replaced immediately so the carry stays bounded.
-func splitValidUTF8(carry, block string) (emit, newCarry string) {
-	s := carry + block
-	cut := len(s)
-	for i := len(s) - 1; i >= 0 && len(s)-i < utf8.UTFMax; i-- {
-		b := s[i]
-		if b < utf8.RuneSelf {
-			break // ASCII: everything before the tail scan is complete
-		}
-		if !utf8.RuneStart(b) {
-			continue // continuation byte: keep looking for its leading byte
-		}
-		// Leading byte: hold the sequence back iff it declares more bytes
-		// than the stream has produced so far (it may complete next block).
-		if utf8SeqLen(b) > len(s)-i {
-			cut = i
-		}
-		break
-	}
-	return sanitizeUTF8(s[:cut]), s[cut:]
-}
-
-// PredictRich is the non-streaming inference path (grpc.AIModelRich).
-// Returns one Reply whose Message is the aggregated assistant content and
-// whose ChatDeltas carry the parsed content/reasoning/tool-call events.
-func (d *Dllm) PredictRich(opts *pb.PredictOptions) (*pb.Reply, error) {
-	d.genMu.RLock()
-	defer d.genMu.RUnlock()
-	if d.gen == nil {
-		return nil, grpcerrors.ModelNotLoaded("dllm")
-	}
-	prompt, parse, optsJSON, imgJSON, err := d.prepareRequest(opts)
-	if err != nil {
-		return nil, err
-	}
-
-	var out string
-	var genErr error
-	d.submit(func() {
-		if imgJSON != "" {
-			out, genErr = d.gen.generateMM(prompt, imgJSON, optsJSON)
-		} else {
-			out, genErr = d.gen.generate(prompt, optsJSON)
-		}
-	})
-	if genErr != nil {
-		return nil, genErr
-	}
-	// Byte-fallback tokens can detokenize to invalid UTF-8; proto3 strings
-	// must be valid or grpc-go fails the whole reply at marshal time.
-	out = sanitizeUTF8(out)
-
-	if !parse {
-		// Raw-prompt mode: plain content, no gemma4 parsing (see buildPrompt).
-		return &pb.Reply{Message: []byte(out), ChatDeltas: []*pb.ChatDelta{{Content: out}}}, nil
-	}
-
-	// The prompt renders with add_generation_prompt; both thinking modes
-	// leave the model starting in content state (see the Gemma4Parser header
-	// comment), hence NewGemma4Parser(false).
-	parser := NewGemma4Parser(false)
-	if reply := replyFromDeltas(append(parser.Feed(out), parser.Close()...)); reply != nil {
-		return reply, nil
-	}
-	// Everything was markers (or out was empty): an empty but non-nil Reply.
-	return &pb.Reply{}, nil
-}
-
-// PredictStreamRich is the streaming counterpart (grpc.AIModelRich): one
-// Reply per committed diffusion block that produced deltas. Per the
-// interface contract the channel is only sent into here - the gRPC server
-// closes it after this returns (opposite to legacy PredictStream).
-func (d *Dllm) PredictStreamRich(opts *pb.PredictOptions, results chan<- *pb.Reply) error {
-	d.genMu.RLock()
-	defer d.genMu.RUnlock()
-	if d.gen == nil {
-		return grpcerrors.ModelNotLoaded("dllm")
-	}
-	prompt, parse, optsJSON, imgJSON, err := d.prepareRequest(opts)
-	if err != nil {
-		return err
-	}
-
-	var parser *Gemma4Parser
-	if parse {
-		parser = NewGemma4Parser(false)
-	}
-	// emit runs inside onBlock, i.e. on the thread driving the C generate.
-	// Sending on results can block on a slow consumer, but the server-side
-	// pump (pkg/grpc/server.go PredictStream) drains continuously and drops
-	// undeliverable sends, so this backpressure is brief and bounded - and
-	// pausing the diffusion loop under it is the desired behavior anyway.
-	emit := func(text string) {
-		if !parse {
-			if text != "" {
-				results <- &pb.Reply{Message: []byte(text), ChatDeltas: []*pb.ChatDelta{{Content: text}}}
-			}
-			return
-		}
-		deltas := parser.Feed(text)
-		if reply := replyFromDeltas(deltas); reply != nil {
-			results <- reply
-		}
-	}
-	// onBlock guards emit (and through it the parser) against invalid UTF-8:
-	// a multi-byte character split across block boundaries is held back until
-	// it completes (see splitValidUTF8), so proto3 marshaling never fails.
-	var carry string
-	onBlock := func(block string) {
-		var text string
-		text, carry = splitValidUTF8(carry, block)
-		emit(text)
-	}
-
-	var genErr error
-	d.submit(func() {
-		if imgJSON != "" {
-			genErr = d.gen.generateStreamMM(prompt, imgJSON, optsJSON, onBlock)
-		} else {
-			genErr = d.gen.generateStream(prompt, optsJSON, onBlock)
-		}
-	})
-	if genErr != nil {
-		return genErr
-	}
-	if carry != "" {
-		// The stream ended mid-sequence: the held-back bytes can no longer
-		// complete, so flush them through the U+FFFD last resort.
-		emit(sanitizeUTF8(carry))
-	}
-	if parse {
-		if reply := replyFromDeltas(parser.Close()); reply != nil {
-			results <- reply
-		}
-	}
-	return nil
-}
-
-// replyFromDeltas wraps one batch of parsed deltas into a streaming Reply,
-// or nil when the batch is empty (markers consumed, nothing emitted yet).
-// Message mirrors the batch's content text so legacy chan-string consumers
-// see exactly the displayed tokens.
-func replyFromDeltas(deltas []*pb.ChatDelta) *pb.Reply {
-	if len(deltas) == 0 {
-		return nil
-	}
-	var content strings.Builder
-	for _, delta := range deltas {
-		content.WriteString(delta.GetContent())
-	}
-	return &pb.Reply{Message: []byte(content.String()), ChatDeltas: deltas}
-}
-
-// Predict is the legacy (string, error) signature; the gRPC server prefers
-// PredictRich, this exists for non-rich callers (cloud-proxy precedent).
-func (d *Dllm) Predict(opts *pb.PredictOptions) (string, error) {
-	reply, err := d.PredictRich(opts)
-	if err != nil {
-		return "", err
-	}
-	return string(reply.GetMessage()), nil
-}
-
-// PredictStream is the legacy chan-string path: rich replies reduced to
-// their content text. Note the inverted channel ownership - the LEGACY
-// contract requires the impl to close the channel.
-func (d *Dllm) PredictStream(opts *pb.PredictOptions, results chan string) error {
-	defer close(results)
-	richCh := make(chan *pb.Reply)
-	errCh := make(chan error, 1)
-	go func() {
-		errCh <- d.PredictStreamRich(opts, richCh)
-		close(richCh)
-	}()
-	for reply := range richCh {
-		if msg := reply.GetMessage(); len(msg) > 0 {
-			results <- string(msg)
-		}
-	}
-	return <-errCh
-}
-
-// TokenizeString tokenizes opts.Prompt via dllm_capi_tokenize_json (the C
-// side prepends bos per the vocab) and decodes the returned id array.
-func (d *Dllm) TokenizeString(opts *pb.PredictOptions) (pb.TokenizationResponse, error) {
-	d.genMu.RLock()
-	defer d.genMu.RUnlock()
-	if d.gen == nil {
-		return pb.TokenizationResponse{}, grpcerrors.ModelNotLoaded("dllm")
-	}
-	var out string
-	var tokErr error
-	d.submit(func() {
-		out, tokErr = d.gen.tokenizeJSON(opts.GetPrompt())
-	})
-	if tokErr != nil {
-		return pb.TokenizationResponse{}, tokErr
-	}
-	var tokens []int32
-	if err := json.Unmarshal([]byte(out), &tokens); err != nil {
-		return pb.TokenizationResponse{}, fmt.Errorf("dllm: decode tokenize result %q: %w", out, err)
-	}
-	return pb.TokenizationResponse{Length: int32(len(tokens)), Tokens: tokens}, nil
-}

backend/go/dllm/gemma4_parser.go

@@ -1,562 +0,0 @@
-// Gemma4 (DiffusionGemma) streaming output parser: raw model text, fed in
-// arbitrary fragments (per committed diffusion block; a fragment can split
-// anywhere, including mid-marker and mid-payload), is turned into
-// pb.ChatDelta events (content / reasoning_content / tool_calls).
-//
-// Normative sources:
-//   - The chat template embedded at the top of gemma4_renderer.go ("tpl L<n>"
-//     citations below refer to its numbered lines). The OUTPUT format mirrors
-//     what the template renders for assistant history: thought channels
-//     (<|channel>thought\n ... <channel|>, tpl L240), tool calls
-//     (<|tool_call>call:name{...}<tool_call|>, tpl L246-L257) and turn ends
-//     (<turn|>, tpl L351).
-//   - vLLM PR #45163: vllm/tool_parsers/gemma4_tool_parser.py (marker
-//     handling, the call:name{...} argument grammar and its decoder, ported
-//     below) and vllm/reasoning/gemma4_reasoning_parser.py (channel markers,
-//     the "thought\n" role label, is_reasoning_end semantics).
-//
-// Initial state (derived from the generation prompt, tpl L356-L362, see
-// RenderGemma4):
-//   - enable_thinking=false: the prompt ends with "<|turn>model\n" +
-//     "<|channel>thought\n<channel|>" - an EMPTY thought channel, pre-opened
-//     AND pre-closed by the template. The model's output therefore starts in
-//     plain content. Use NewGemma4Parser(false).
-//   - enable_thinking=true: the prompt ends at "<|turn>model\n" and the model
-//     opens and closes its own thought channel in the OUTPUT
-//     ("<|channel>thought\n...reasoning...<channel|>final answer", per the
-//     vLLM Gemma4ReasoningParser docstring). The parser still starts in
-//     content state - the channel markers in the output drive the switch.
-//     Use NewGemma4Parser(false) here too.
-//   - NewGemma4Parser(true) is for callers that pre-open the thought channel
-//     in the prompt themselves (appending "<|channel>thought\n" after the
-//     generation prompt to force thinking): the output then begins mid-thought
-//     and everything is reasoning until the first <channel|>.
-//
-// State diagram (markers are consumed, never emitted):
-//
-//	             <|channel>                  \n (channel name dropped: the
-//	[content] --------------> [chan-header] ----> [thought]   "thought\n" role
-//	   ^ |  <channel|> (stray close: swallowed,                label, stripped
-//	   +-+  strip_thinking semantics, tpl L148-L158)           like vLLM does)
-//	   ^                  <channel|>
-//	   +----------------------------------------- [thought]
-//	   ^                  <tool_call|>                 | <|tool_call> (implicit
-//	   +-------------- [tool-call] <-------------------+  reasoning end, vLLM
-//	   |  <|tool_call>     ^                               is_reasoning_end)
-//	   +-------------------+
-//	[content]/[thought] --- <turn|> ---> [done]  (everything after is dropped)
-//
-// Buffering rules:
-//   - content/thought states hold back at most len(longest marker)-1 bytes:
-//     the longest tail that is still a proper prefix of a watched marker.
-//     Content is otherwise emitted immediately (no unbounded buffering).
-//   - the tool-call state buffers the whole payload until <tool_call|>. This
-//     is unbounded in principle but bounded in practice by the model's
-//     diffusion canvas, and is required because the call:name{...} payload
-//     only becomes decodable (and trustworthy) once complete - the same
-//     reason vLLM's parser accumulates before parsing.
-//   - Close() flushes whatever is still held: partial markers come out as
-//     content/reasoning (per the state that held them); an unterminated
-//     channel header or tool-call payload is re-emitted RAW (including its
-//     opening marker) as content - malformed output is never silently
-//     dropped (mirrors vLLM extract_tool_calls returning the raw text as
-//     content when its regex does not match).
-//
-// Streaming granularity DIVERGENCE from vLLM: vLLM re-parses the partial
-// payload on every token and streams argument-JSON diffs (its `partial=True`
-// decoder mode plus withholding logic exist only for that). Our fragments are
-// whole committed diffusion blocks, so each completed tool call is emitted
-// once, as a single ToolCallDelta carrying index + id + name + the full
-// arguments JSON - exactly the shape backend/python/vllm/backend.py emits
-// per call and pkg/functions.ToolCallsFromChatDeltas re-accumulates.
-package main
-
-import (
-	"encoding/json"
-	"regexp"
-	"strconv"
-	"strings"
-
-	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
-)
-
-// gemma4CallRE is vLLM's tool_call_regex
-// (`<\|tool_call>call:([\w\-\.]+)\{(.*?)\}<tool_call\|>`, DOTALL) anchored to
-// a single already-extracted payload: name charset [\w\-.], braces mandatory.
-var gemma4CallRE = regexp.MustCompile(`(?s)^call:([\w\-.]+)\{(.*)\}$`)
-
-type g4State int
-
-const (
-	g4Content g4State = iota
-	g4ChanHeader
-	g4Thought
-	g4ToolCall
-	g4Done
-)
-
-// Markers watched per emitting state. A stray <tool_call|> outside a tool
-// call is deliberately NOT watched: it passes through verbatim, consistent
-// with the malformed-payload fallback re-emitting it as content.
-var (
-	gemma4ContentMarkers = []string{gemma4ChannelOpen, gemma4ChannelClose, gemma4ToolCallOpen, gemma4TurnEnd}
-	gemma4ThoughtMarkers = []string{gemma4ChannelClose, gemma4ToolCallOpen, gemma4TurnEnd}
-)
-
-type Gemma4Parser struct {
-	state g4State
-	// held is the per-state carry-over between Feed calls: a partial marker
-	// (content/thought), a partial channel header (chan-header) or the
-	// payload accumulated so far (tool-call).
-	held    string
-	toolIdx int
-}
-
-// NewGemma4Parser returns a parser positioned per the initial-state rules in
-// the header comment: startInThought=true only when the caller pre-opened a
-// thought channel in the prompt.
-func NewGemma4Parser(startInThought bool) *Gemma4Parser {
-	state := g4Content
-	if startInThought {
-		state = g4Thought
-	}
-	return &Gemma4Parser{state: state}
-}
-
-// Feed consumes the next output fragment and returns the deltas it completes.
-func (p *Gemma4Parser) Feed(text string) []*pb.ChatDelta {
-	if text == "" || p.state == g4Done {
-		return nil
-	}
-	pending := p.held + text
-	p.held = ""
-	var em g4Emitter
-	for pending != "" {
-		switch p.state {
-		case g4Content, g4Thought:
-			markers := gemma4ContentMarkers
-			if p.state == g4Thought {
-				markers = gemma4ThoughtMarkers
-			}
-			idx, marker := findEarliestGemma4Marker(pending, markers)
-			if idx == -1 {
-				hold := gemma4MarkerHoldback(pending, markers)
-				p.emitText(&em, pending[:len(pending)-hold])
-				p.held = pending[len(pending)-hold:]
-				pending = ""
-				continue
-			}
-			p.emitText(&em, pending[:idx])
-			pending = pending[idx+len(marker):]
-			switch marker {
-			case gemma4ChannelOpen:
-				p.state = g4ChanHeader
-			case gemma4ChannelClose:
-				// In thought: channel ends. In content: stray close,
-				// swallowed (strip_thinking keeps both sides, tpl L148-L158).
-				p.state = g4Content
-			case gemma4ToolCallOpen:
-				p.state = g4ToolCall
-			case gemma4TurnEnd:
-				p.state = g4Done
-			}
-		case g4ChanHeader:
-			// The channel header is "<name>\n"; the template only ever writes
-			// "thought" (tpl L240/L360) and the label is structural, so it is
-			// dropped, not emitted (vLLM strips the same "thought\n" prefix).
-			nl := strings.IndexByte(pending, '\n')
-			if nl == -1 {
-				p.held = pending
-				pending = ""
-				continue
-			}
-			pending = pending[nl+1:]
-			p.state = g4Thought
-		case g4ToolCall:
-			end := strings.Index(pending, gemma4ToolCallClose)
-			if end == -1 {
-				p.held = pending
-				pending = ""
-				continue
-			}
-			p.emitToolCall(&em, pending[:end])
-			pending = pending[end+len(gemma4ToolCallClose):]
-			p.state = g4Content
-		case g4Done:
-			pending = ""
-		}
-	}
-	return em.deltas
-}
-
-// Close flushes held-back partials. Incomplete structures (open channel
-// header, unterminated tool payload) are re-emitted raw as content rather
-// than dropped. The parser is finished afterwards.
-func (p *Gemma4Parser) Close() []*pb.ChatDelta {
-	var em g4Emitter
-	switch p.state {
-	case g4Content:
-		em.content(p.held)
-	case g4Thought:
-		em.reasoning(p.held)
-	case g4ChanHeader:
-		em.content(gemma4ChannelOpen + p.held)
-	case g4ToolCall:
-		em.content(gemma4ToolCallOpen + p.held)
-	case g4Done:
-	}
-	p.held = ""
-	p.state = g4Done
-	return em.deltas
-}
-
-func (p *Gemma4Parser) emitText(em *g4Emitter, s string) {
-	if p.state == g4Thought {
-		em.reasoning(s)
-		return
-	}
-	em.content(s)
-}
-
-// emitToolCall decodes one complete <|tool_call>...<tool_call|> payload. On a
-// payload that does not match call:name{...} the raw text (markers included)
-// is emitted as content, mirroring vLLM's extract_tool_calls fallback.
-func (p *Gemma4Parser) emitToolCall(em *g4Emitter, payload string) {
-	m := gemma4CallRE.FindStringSubmatch(payload)
-	if m == nil {
-		em.content(gemma4ToolCallOpen + payload + gemma4ToolCallClose)
-		return
-	}
-	// Index-based ids: deterministic (the split-invariance property relies
-	// on it) and matching the call_<n> convention of pkg/grpc/rich_test.go;
-	// core only needs ids to be non-empty and unique within the response.
-	em.tool(p.toolIdx, "call_"+strconv.Itoa(p.toolIdx), m[1], decodeGemma4Args(m[2], 0))
-	p.toolIdx++
-}
-
-// g4Emitter collects ChatDeltas; empty text events are dropped.
-type g4Emitter struct {
-	deltas []*pb.ChatDelta
-}
-
-func (e *g4Emitter) content(s string) {
-	if s != "" {
-		e.deltas = append(e.deltas, &pb.ChatDelta{Content: s})
-	}
-}
-
-func (e *g4Emitter) reasoning(s string) {
-	if s != "" {
-		e.deltas = append(e.deltas, &pb.ChatDelta{ReasoningContent: s})
-	}
-}
-
-func (e *g4Emitter) tool(index int, id, name, argsJSON string) {
-	e.deltas = append(e.deltas, &pb.ChatDelta{ToolCalls: []*pb.ToolCallDelta{{
-		Index:     int32(index),
-		Id:        id,
-		Name:      name,
-		Arguments: argsJSON,
-	}}})
-}
-
-// findEarliestGemma4Marker returns the position and value of the first
-// complete marker occurrence, or (-1, "").
-func findEarliestGemma4Marker(s string, markers []string) (int, string) {
-	best, bestMarker := -1, ""
-	for _, m := range markers {
-		if idx := strings.Index(s, m); idx >= 0 && (best == -1 || idx < best) {
-			best, bestMarker = idx, m
-		}
-	}
-	return best, bestMarker
-}
-
-// gemma4MarkerHoldback returns the length of the longest suffix of s that is
-// a proper prefix of a watched marker - the only bytes that may still grow
-// into a marker and therefore must not be emitted yet (bounded by the
-// longest marker, so content is never buffered unboundedly).
-func gemma4MarkerHoldback(s string, markers []string) int {
-	maxHold := 0
-	for _, m := range markers {
-		if len(m)-1 > maxHold {
-			maxHold = len(m) - 1
-		}
-	}
-	if len(s) < maxHold {
-		maxHold = len(s)
-	}
-	for k := maxHold; k >= 1; k-- {
-		tail := s[len(s)-k:]
-		for _, m := range markers {
-			if strings.HasPrefix(m, tail) {
-				return k
-			}
-		}
-	}
-	return 0
-}
-
-// ---------------------------------------------------------------------------
-// call:name{...} argument decoder
-//
-// Port of vLLM's _parse_gemma4_args / _parse_gemma4_array /
-// _parse_gemma4_value (gemma4_tool_parser.py) in non-partial mode only: this
-// parser decodes exclusively COMPLETE payloads (incomplete ones fall back to
-// raw content at Close), so vLLM's partial-withholding machinery
-// (trailing-dot floats, withheld bare tails) is intentionally not ported.
-//
-// Grammar (inverse of the renderer's formatGemma4Argument, tpl L118-L147):
-//
-//	args    := pair (',' pair)*
-//	pair    := key ':' value          (keys unquoted, up to the first ':')
-//	value   := string | object | array | bare
-//	string  := '<|"|>' ... '<|"|>'    (no escapes; unterminated -> rest)
-//	object  := '{' args '}'           (delimited strings skipped when
-//	array   := '[' value,* ']'         counting braces/brackets)
-//	bare    := true | false | null/none/nil | number | bare-string
-//
-// Output is a JSON object/array string with keys in payload order (Python
-// dict insertion order), built with HTML escaping off so payload text
-// survives byte-for-byte.
-// ---------------------------------------------------------------------------
-
-func isGemma4Space(c byte) bool { return c == ' ' || c == '\n' || c == '\t' }
-
-// gemma4MaxArgsDepth caps the mutual recursion between decodeGemma4Args and
-// decodeGemma4Array. Defense against model-generated deep nesting: a Go stack
-// overflow is a fatal process kill, not a recoverable error, so past the cap
-// a nested body gracefully degrades to a JSON string of its raw text.
-const gemma4MaxArgsDepth = 100
-
-// decodeGemma4Args decodes one args body (the text between the outer braces
-// of call:name{...}) into a JSON object string. depth is the current nesting
-// level (0 at the payload root); see gemma4MaxArgsDepth.
-func decodeGemma4Args(s string, depth int) string {
-	if depth > gemma4MaxArgsDepth {
-		return gemma4JSONString(s)
-	}
-	var b strings.Builder
-	b.WriteString("{")
-	first := true
-	pair := func(key, val string) {
-		if !first {
-			b.WriteString(",")
-		}
-		first = false
-		b.WriteString(gemma4JSONString(key))
-		b.WriteString(":")
-		b.WriteString(val)
-	}
-	i, n := 0, len(s)
-	for i < n {
-		for i < n && (isGemma4Space(s[i]) || s[i] == ',') {
-			i++
-		}
-		if i >= n {
-			break
-		}
-		keyStart := i
-		for i < n && s[i] != ':' {
-			i++
-		}
-		if i >= n {
-			break // no ':' -> trailing junk, dropped (vLLM does the same)
-		}
-		key := strings.TrimSpace(s[keyStart:i])
-		i++ // skip ':'
-		for i < n && isGemma4Space(s[i]) {
-			i++
-		}
-		if i >= n {
-			pair(key, `""`) // "key:" with nothing after -> empty string
-			break
-		}
-		switch {
-		case strings.HasPrefix(s[i:], gemma4StringDelim):
-			i += len(gemma4StringDelim)
-			if end := strings.Index(s[i:], gemma4StringDelim); end == -1 {
-				pair(key, gemma4JSONString(s[i:])) // unterminated -> take rest
-				i = n
-			} else {
-				pair(key, gemma4JSONString(s[i:i+end]))
-				i += end + len(gemma4StringDelim)
-			}
-		case s[i] == '{':
-			inner, next := scanGemma4Balanced(s, i, '{', '}')
-			pair(key, decodeGemma4Args(inner, depth+1))
-			i = next
-		case s[i] == '[':
-			inner, next := scanGemma4Balanced(s, i, '[', ']')
-			pair(key, decodeGemma4Array(inner, depth+1))
-			i = next
-		default:
-			valStart := i
-			for i < n && s[i] != ',' && s[i] != '}' && s[i] != ']' {
-				i++
-			}
-			if i == valStart {
-				// No progress (value starts on a stray '}'/']'): abort on
-				// malformed input rather than loop, like vLLM.
-				i = n
-				continue
-			}
-			pair(key, decodeGemma4Bare(s[valStart:i]))
-		}
-	}
-	b.WriteString("}")
-	return b.String()
-}
-
-// decodeGemma4Array decodes one array body (the text between '[' and ']')
-// into a JSON array string. depth is the current nesting level; see
-// gemma4MaxArgsDepth.
-func decodeGemma4Array(s string, depth int) string {
-	if depth > gemma4MaxArgsDepth {
-		return gemma4JSONString(s)
-	}
-	var b strings.Builder
-	b.WriteString("[")
-	first := true
-	item := func(val string) {
-		if !first {
-			b.WriteString(",")
-		}
-		first = false
-		b.WriteString(val)
-	}
-	i, n := 0, len(s)
-	for i < n {
-		for i < n && (isGemma4Space(s[i]) || s[i] == ',') {
-			i++
-		}
-		if i >= n {
-			break
-		}
-		switch {
-		case strings.HasPrefix(s[i:], gemma4StringDelim):
-			i += len(gemma4StringDelim)
-			if end := strings.Index(s[i:], gemma4StringDelim); end == -1 {
-				item(gemma4JSONString(s[i:]))
-				i = n
-			} else {
-				item(gemma4JSONString(s[i : i+end]))
-				i += end + len(gemma4StringDelim)
-			}
-		case s[i] == '{':
-			inner, next := scanGemma4Balanced(s, i, '{', '}')
-			item(decodeGemma4Args(inner, depth+1))
-			i = next
-		case s[i] == '[':
-			inner, next := scanGemma4Balanced(s, i, '[', ']')
-			item(decodeGemma4Array(inner, depth+1))
-			i = next
-		default:
-			valStart := i
-			for i < n && s[i] != ',' && s[i] != ']' {
-				i++
-			}
-			if i == valStart {
-				i = n // no progress: abort on malformed input, like vLLM
-				continue
-			}
-			item(decodeGemma4Bare(s[valStart:i]))
-		}
-	}
-	b.WriteString("]")
-	return b.String()
-}
-
-// scanGemma4Balanced scans a brace/bracket-balanced span starting at the
-// opener s[start], skipping over <|"|>-delimited strings so structural
-// characters inside them do not count (vLLM's depth scan). Returns the inner
-// text and the index just past the closer; an unterminated span yields the
-// rest of the string (the inner decoder still extracts what is there - this
-// path is only reachable from genuinely malformed complete payloads).
-func scanGemma4Balanced(s string, start int, open, close byte) (string, int) {
-	depth := 1
-	i := start + 1
-	innerStart := i
-	n := len(s)
-	for i < n && depth > 0 {
-		if strings.HasPrefix(s[i:], gemma4StringDelim) {
-			i += len(gemma4StringDelim)
-			if nd := strings.Index(s[i:], gemma4StringDelim); nd == -1 {
-				i = n
-			} else {
-				i += nd + len(gemma4StringDelim)
-			}
-			continue
-		}
-		switch s[i] {
-		case open:
-			depth++
-		case close:
-			depth--
-		}
-		i++
-	}
-	if depth > 0 {
-		return s[innerStart:], n
-	}
-	return s[innerStart : i-1], i
-}
-
-// decodeGemma4Bare maps an undelimited value to its JSON form: booleans,
-// null aliases (null/none/nil, case-insensitive - the renderer writes
-// Python None as "None", tpl L144-L145 via format_argument's else branch),
-// numbers (vLLM's rule: a '.' tries float, otherwise int; anything that
-// fails parses as a bare string).
-func decodeGemma4Bare(raw string) string {
-	v := strings.TrimSpace(raw)
-	if v == "" {
-		return `""`
-	}
-	if v == "true" || v == "false" {
-		return v
-	}
-	switch strings.ToLower(v) {
-	case "null", "none", "nil":
-		return "null"
-	}
-	if strings.Contains(v, ".") {
-		if f, err := strconv.ParseFloat(v, 64); err == nil {
-			return formatGemma4Float(f)
-		}
-	} else if iv, err := strconv.ParseInt(v, 10, 64); err == nil {
-		return strconv.FormatInt(iv, 10)
-	}
-	return gemma4JSONString(v)
-}
-
-// formatGemma4Float renders like Python's json.dumps(float): integral floats
-// keep a ".0" suffix ("108." decodes to 108.0, not 108), so the arguments
-// JSON matches what vLLM would have produced for the same payload.
-func formatGemma4Float(f float64) string {
-	s := strconv.FormatFloat(f, 'g', -1, 64)
-	if !strings.ContainsAny(s, ".eE") {
-		s += ".0"
-	}
-	return s
-}
-
-// gemma4JSONString encodes a JSON string WITHOUT HTML escaping (json.Marshal
-// would escape the angle brackets in "<div>" to \u003c / \u003e sequences;
-// payload text should survive
-// byte-for-byte, like Python's json.dumps(ensure_ascii=False)).
-func gemma4JSONString(s string) string {
-	var sb strings.Builder
-	enc := json.NewEncoder(&sb)
-	enc.SetEscapeHTML(false)
-	if err := enc.Encode(s); err != nil {
-		// Unreachable for plain strings; fall back to default escaping
-		// rather than emitting invalid JSON.
-		b, mErr := json.Marshal(s)
-		if mErr != nil {
-			return `""`
-		}
-		return string(b)
-	}
-	// Encode appends a trailing newline.
-	return strings.TrimSuffix(sb.String(), "\n")
-}

backend/go/dllm/gemma4_parser_test.go

@@ -1,592 +0,0 @@
-package main
-
-// Parser specs for Gemma4Parser (model output text -> pb.ChatDelta events).
-//
-// Fixture provenance:
-//   - Entries marked "vLLM: <name>" are direct ports of the named test from
-//     vLLM PR #45163, tests/tool_parsers/test_gemma4_tool_parser.py (the
-//     authoritative test-suite for the gemma4 tool-call wire format). The
-//     streaming tests' chunk lists are reused verbatim as Feed fragments.
-//   - Decoder entries port the TestParseGemma4Args / TestParseGemma4Array
-//     classes from the same file (non-partial mode only; this parser never
-//     decodes partial payloads, see the divergence note in gemma4_parser.go).
-//   - Channel/turn-marker expectations come from the chat template embedded
-//     in gemma4_renderer.go (tpl L356-L362 generation prompt, L148-L158
-//     strip_thinking) and vLLM's Gemma4ReasoningParser
-//     (vllm/reasoning/gemma4_reasoning_parser.py).
-
-import (
-	"encoding/json"
-	"fmt"
-	"strings"
-
-	. "github.com/onsi/ginkgo/v2"
-	. "github.com/onsi/gomega"
-
-	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
-)
-
-// flatGemma4Tool is one accumulated tool call, mirroring how LocalAI core
-// folds ToolCallDelta streams (pkg/functions/chat_deltas.go
-// ToolCallsFromChatDeltas: name/id latch on first non-empty, arguments
-// concatenate per index). Tests flatten through the same rules so they
-// assert exactly what core will reconstruct.
-type flatGemma4Tool struct {
-	id   string
-	name string
-	args string
-}
-
-func flattenGemma4Deltas(deltas []*pb.ChatDelta) (string, string, []flatGemma4Tool) {
-	var content, reasoning strings.Builder
-	byIndex := map[int32]*flatGemma4Tool{}
-	maxIdx := int32(-1)
-	for _, d := range deltas {
-		content.WriteString(d.GetContent())
-		reasoning.WriteString(d.GetReasoningContent())
-		for _, tc := range d.GetToolCalls() {
-			acc, ok := byIndex[tc.GetIndex()]
-			if !ok {
-				acc = &flatGemma4Tool{}
-				byIndex[tc.GetIndex()] = acc
-			}
-			if tc.GetName() != "" {
-				acc.name = tc.GetName()
-			}
-			if tc.GetId() != "" {
-				acc.id = tc.GetId()
-			}
-			acc.args += tc.GetArguments()
-			if tc.GetIndex() > maxIdx {
-				maxIdx = tc.GetIndex()
-			}
-		}
-	}
-	var tools []flatGemma4Tool
-	for i := int32(0); i <= maxIdx; i++ {
-		if acc, ok := byIndex[i]; ok {
-			tools = append(tools, *acc)
-		}
-	}
-	return content.String(), reasoning.String(), tools
-}
-
-type wantGemma4Tool struct {
-	name     string
-	argsJSON string // compared with MatchJSON (key order irrelevant)
-}
-
-type parseGemma4Case struct {
-	startInThought bool
-	fragments      []string
-	wantContent    string
-	wantReasoning  string
-	wantTools      []wantGemma4Tool
-}
-
-func parseGemma4Fragments(startInThought bool, fragments []string) []*pb.ChatDelta {
-	p := NewGemma4Parser(startInThought)
-	var all []*pb.ChatDelta
-	for _, f := range fragments {
-		all = append(all, p.Feed(f)...)
-	}
-	return append(all, p.Close()...)
-}
-
-var _ = Describe("Gemma4Parser", func() {
-	DescribeTable("parses streamed gemma4 output into ChatDeltas",
-		func(c parseGemma4Case) {
-			content, reasoning, tools := flattenGemma4Deltas(parseGemma4Fragments(c.startInThought, c.fragments))
-			Expect(content).To(Equal(c.wantContent))
-			Expect(reasoning).To(Equal(c.wantReasoning))
-			Expect(tools).To(HaveLen(len(c.wantTools)))
-			seenIDs := map[string]bool{}
-			for i, want := range c.wantTools {
-				Expect(tools[i].name).To(Equal(want.name), "tool %d name", i)
-				Expect(tools[i].args).To(MatchJSON(want.argsJSON), "tool %d arguments", i)
-				Expect(tools[i].id).ToNot(BeEmpty(), "tool %d id", i)
-				Expect(seenIDs).ToNot(HaveKey(tools[i].id), "tool %d id must be unique", i)
-				seenIDs[tools[i].id] = true
-			}
-		},
-
-		// --- (1) pure content -------------------------------------------------
-		// vLLM: test_no_tool_calls
-		Entry("pure content, single fragment", parseGemma4Case{
-			fragments:   []string{"Hello, how can I help you today?"},
-			wantContent: "Hello, how can I help you today?",
-		}),
-
-		// --- (2) thought -> final transition ----------------------------------
-		// enable_thinking render: prompt ends at <|turn>model\n and the model
-		// opens/closes its own thought channel in the OUTPUT (vLLM
-		// Gemma4ReasoningParser docstring; tpl L356-L362). The "thought\n"
-		// role label after <|channel> is structural and must be stripped
-		// (vLLM _THOUGHT_PREFIX handling).
-		Entry("thought channel then final content", parseGemma4Case{
-			fragments:     []string{"<|channel>thought\nLet me think about this.\n<channel|>The answer is 42."},
-			wantReasoning: "Let me think about this.\n",
-			wantContent:   "The answer is 42.",
-		}),
-
-		// --- (3) startInThought both ways -------------------------------------
-		Entry("startInThought=true routes initial text to reasoning until <channel|>", parseGemma4Case{
-			startInThought: true,
-			fragments:      []string{"I am thinking hard.<channel|>Done."},
-			wantReasoning:  "I am thinking hard.",
-			wantContent:    "Done.",
-		}),
-		// A stray <channel|> with no open channel is swallowed, matching the
-		// template's strip_thinking (tpl L148-L158: the marker is dropped,
-		// text on both sides is kept).
-		Entry("startInThought=false keeps the same text as content, stray <channel|> swallowed", parseGemma4Case{
-			startInThought: false,
-			fragments:      []string{"I am thinking hard.<channel|>Done."},
-			wantContent:    "I am thinking hard.Done.",
-		}),
-
-		// --- (4) one tool call, full payload type zoo --------------------------
-		Entry("single tool call: strings, numbers, bools, null, nested object and array", parseGemma4Case{
-			fragments: []string{`<|tool_call>call:complex_function{text:<|"|>with, comma and {braces}<|"|>,count:42,score:3.14,yes:true,no:false,nothing:null,obj:{inner:<|"|>v<|"|>,k:1},arr:[<|"|>a<|"|>,2,true]}<tool_call|>`},
-			wantTools: []wantGemma4Tool{{
-				name:     "complex_function",
-				argsJSON: `{"text":"with, comma and {braces}","count":42,"score":3.14,"yes":true,"no":false,"nothing":null,"obj":{"inner":"v","k":1},"arr":["a",2,true]}`,
-			}},
-		}),
-
-		// --- (5) payload split across 3 fragments ------------------------------
-		Entry("tool-call payload split across three fragments", parseGemma4Case{
-			fragments: []string{
-				"<|tool_call>call:get_weather{loc",
-				`ation:<|"|>Paris, Fra`,
-				`nce<|"|>}<tool_call|>`,
-			},
-			wantTools: []wantGemma4Tool{{name: "get_weather", argsJSON: `{"location":"Paris, France"}`}},
-		}),
-
-		// --- (6) marker split across fragments ----------------------------------
-		Entry("tool-call open marker split across fragments", parseGemma4Case{
-			fragments: []string{
-				"<|tool_ca",
-				`ll>call:get_weather{location:<|"|>London<|"|>}<tool_call|>`,
-			},
-			wantTools: []wantGemma4Tool{{name: "get_weather", argsJSON: `{"location":"London"}`}},
-		}),
-		Entry("channel open marker split across fragments", parseGemma4Case{
-			fragments: []string{
-				"<|chan",
-				"nel>thought\ndeep thought<channel|>final",
-			},
-			wantReasoning: "deep thought",
-			wantContent:   "final",
-		}),
-
-		// --- (7) trailing partial marker held, flushed by Close -----------------
-		Entry("trailing partial marker is held back and flushed by Close", parseGemma4Case{
-			fragments:   []string{"Hello <|tool"},
-			wantContent: "Hello <|tool",
-		}),
-
-		// --- (8) malformed/incomplete payload -> content fallback ---------------
-		// vLLM: test_incomplete_tool_call (no end marker: the whole text stays
-		// content, never silently dropped).
-		Entry("incomplete tool payload at Close is emitted as raw content", parseGemma4Case{
-			fragments:   []string{`<|tool_call>call:get_weather{location:<|"|>London`},
-			wantContent: `<|tool_call>call:get_weather{location:<|"|>London`,
-		}),
-		Entry("malformed complete payload is emitted as raw content, parsing continues", parseGemma4Case{
-			fragments:   []string{"<|tool_call>oops no call syntax<tool_call|> done"},
-			wantContent: "<|tool_call>oops no call syntax<tool_call|> done",
-		}),
-
-		// --- (9) <turn|> ends the turn -------------------------------------------
-		Entry("text after <turn|> is ignored, including later fragments", parseGemma4Case{
-			fragments: []string{
-				"before<turn|>after",
-				`more <|tool_call>call:f{}<tool_call|>`,
-			},
-			wantContent: "before",
-		}),
-		Entry("<turn|> inside a thought channel ends the turn", parseGemma4Case{
-			startInThought: true,
-			fragments:      []string{"thinking<turn|>ignored"},
-			wantReasoning:  "thinking",
-		}),
-
-		// --- (10) ported vLLM non-streaming cases ---------------------------------
-		// vLLM: test_single_tool_call
-		Entry("vLLM: test_single_tool_call", parseGemma4Case{
-			fragments: []string{`<|tool_call>call:get_weather{location:<|"|>London<|"|>}<tool_call|>`},
-			wantTools: []wantGemma4Tool{{name: "get_weather", argsJSON: `{"location":"London"}`}},
-		}),
-		// vLLM: test_multiple_arguments
-		Entry("vLLM: test_multiple_arguments", parseGemma4Case{
-			fragments: []string{`<|tool_call>call:get_weather{location:<|"|>San Francisco<|"|>,unit:<|"|>celsius<|"|>}<tool_call|>`},
-			wantTools: []wantGemma4Tool{{name: "get_weather", argsJSON: `{"location":"San Francisco","unit":"celsius"}`}},
-		}),
-		// vLLM: test_text_before_tool_call. DIVERGENCE: vLLM's non-streaming
-		// extractor trims the content ("...you."); a streaming parser cannot
-		// retroactively trim already-emitted text, so the trailing space is
-		// kept (vLLM's own streaming path keeps it too, see
-		// test_streaming_text_before_tool_call which only checks a prefix).
-		Entry("vLLM: test_text_before_tool_call (streaming semantics: no trim)", parseGemma4Case{
-			fragments:   []string{`Let me check the weather for you. <|tool_call>call:get_weather{location:<|"|>Paris<|"|>}<tool_call|>`},
-			wantContent: "Let me check the weather for you. ",
-			wantTools:   []wantGemma4Tool{{name: "get_weather", argsJSON: `{"location":"Paris"}`}},
-		}),
-		// vLLM: test_multiple_tool_calls (also covers case 11: multi-tool sequence)
-		Entry("vLLM: test_multiple_tool_calls", parseGemma4Case{
-			fragments: []string{`<|tool_call>call:get_weather{location:<|"|>London<|"|>}<tool_call|><|tool_call>call:get_time{location:<|"|>London<|"|>}<tool_call|>`},
-			wantTools: []wantGemma4Tool{
-				{name: "get_weather", argsJSON: `{"location":"London"}`},
-				{name: "get_time", argsJSON: `{"location":"London"}`},
-			},
-		}),
-		// vLLM: test_nested_arguments
-		Entry("vLLM: test_nested_arguments", parseGemma4Case{
-			fragments: []string{`<|tool_call>call:complex_function{nested:{inner:<|"|>value<|"|>},list:[<|"|>a<|"|>,<|"|>b<|"|>]}<tool_call|>`},
-			wantTools: []wantGemma4Tool{{name: "complex_function", argsJSON: `{"nested":{"inner":"value"},"list":["a","b"]}`}},
-		}),
-		// vLLM: test_tool_call_with_number_and_boolean
-		Entry("vLLM: test_tool_call_with_number_and_boolean", parseGemma4Case{
-			fragments: []string{`<|tool_call>call:set_status{is_active:true,count:42,score:3.14}<tool_call|>`},
-			wantTools: []wantGemma4Tool{{name: "set_status", argsJSON: `{"is_active":true,"count":42,"score":3.14}`}},
-		}),
-		// vLLM: test_hyphenated_function_name
-		Entry("vLLM: test_hyphenated_function_name", parseGemma4Case{
-			fragments: []string{`<|tool_call>call:get-weather{location:<|"|>London<|"|>}<tool_call|>`},
-			wantTools: []wantGemma4Tool{{name: "get-weather", argsJSON: `{"location":"London"}`}},
-		}),
-		// vLLM: test_dotted_function_name
-		Entry("vLLM: test_dotted_function_name", parseGemma4Case{
-			fragments: []string{`<|tool_call>call:weather.get{location:<|"|>London<|"|>}<tool_call|>`},
-			wantTools: []wantGemma4Tool{{name: "weather.get", argsJSON: `{"location":"London"}`}},
-		}),
-		// vLLM: test_no_arguments
-		Entry("vLLM: test_no_arguments", parseGemma4Case{
-			fragments: []string{"<|tool_call>call:get_status{}<tool_call|>"},
-			wantTools: []wantGemma4Tool{{name: "get_status", argsJSON: `{}`}},
-		}),
-
-		// --- ported vLLM streaming cases (chunk lists reused as fragments) --------
-		// vLLM: test_basic_streaming_single_tool
-		Entry("vLLM: test_basic_streaming_single_tool", parseGemma4Case{
-			fragments: []string{
-				"<|tool_call>",
-				"call:get_weather{",
-				`location:<|"|>Paris`,
-				", France",
-				`<|"|>}`,
-				"<tool_call|>",
-			},
-			wantTools: []wantGemma4Tool{{name: "get_weather", argsJSON: `{"location":"Paris, France"}`}},
-		}),
-		// vLLM: test_streaming_multi_arg
-		Entry("vLLM: test_streaming_multi_arg", parseGemma4Case{
-			fragments: []string{
-				"<|tool_call>",
-				"call:get_weather{",
-				`location:<|"|>Tokyo<|"|>,`,
-				`unit:<|"|>celsius<|"|>}`,
-				"<tool_call|>",
-			},
-			wantTools: []wantGemma4Tool{{name: "get_weather", argsJSON: `{"location":"Tokyo","unit":"celsius"}`}},
-		}),
-		// vLLM: test_streaming_text_before_tool_call
-		Entry("vLLM: test_streaming_text_before_tool_call", parseGemma4Case{
-			fragments: []string{
-				"Let me check ",
-				"the weather. ",
-				"<|tool_call>",
-				"call:get_weather{",
-				`location:<|"|>London<|"|>}`,
-				"<tool_call|>",
-			},
-			wantContent: "Let me check the weather. ",
-			wantTools:   []wantGemma4Tool{{name: "get_weather", argsJSON: `{"location":"London"}`}},
-		}),
-		// vLLM: test_streaming_numeric_args
-		Entry("vLLM: test_streaming_numeric_args", parseGemma4Case{
-			fragments: []string{
-				"<|tool_call>",
-				"call:set_config{",
-				"count:42,",
-				"active:true}",
-				"<tool_call|>",
-			},
-			wantTools: []wantGemma4Tool{{name: "set_config", argsJSON: `{"count":42,"active":true}`}},
-		}),
-		// vLLM: test_streaming_boolean_split_across_chunks
-		Entry("vLLM: test_streaming_boolean_split_across_chunks", parseGemma4Case{
-			fragments: []string{
-				"<|tool_call>",
-				"call:search{input:{all:tru",
-				"e}}",
-				"<tool_call|>",
-			},
-			wantTools: []wantGemma4Tool{{name: "search", argsJSON: `{"input":{"all":true}}`}},
-		}),
-		// vLLM: test_streaming_false_split_across_chunks
-		Entry("vLLM: test_streaming_false_split_across_chunks", parseGemma4Case{
-			fragments: []string{
-				"<|tool_call>",
-				"call:set{flag:fals",
-				"e}",
-				"<tool_call|>",
-			},
-			wantTools: []wantGemma4Tool{{name: "set", argsJSON: `{"flag":false}`}},
-		}),
-		// vLLM: test_streaming_number_split_across_chunks
-		Entry("vLLM: test_streaming_number_split_across_chunks", parseGemma4Case{
-			fragments: []string{
-				"<|tool_call>",
-				"call:set{count:4",
-				"2}",
-				"<tool_call|>",
-			},
-			wantTools: []wantGemma4Tool{{name: "set", argsJSON: `{"count":42}`}},
-		}),
-		// vLLM: test_streaming_empty_args
-		Entry("vLLM: test_streaming_empty_args", parseGemma4Case{
-			fragments: []string{
-				"<|tool_call>",
-				"call:get_status{}",
-				"<tool_call|>",
-			},
-			wantTools: []wantGemma4Tool{{name: "get_status", argsJSON: `{}`}},
-		}),
-		// vLLM: test_streaming_split_delimiter_no_invalid_json (string
-		// delimiter <|"|> split across fragments must not leak fragments).
-		Entry("vLLM: test_streaming_split_delimiter_no_invalid_json", parseGemma4Case{
-			fragments: []string{
-				"<|tool_call>",
-				"call:todowrite{",
-				`content:<|"|>Buy milk<|`,
-				`"|>}`,
-				"<tool_call|>",
-			},
-			wantTools: []wantGemma4Tool{{name: "todowrite", argsJSON: `{"content":"Buy milk"}`}},
-		}),
-		// vLLM: test_streaming_does_not_duplicate_plain_text_after_tool_call
-		Entry("vLLM: test_streaming_does_not_duplicate_plain_text_after_tool_call", parseGemma4Case{
-			fragments: []string{
-				"<|tool_call>",
-				"call:get_weather{",
-				`location:<|"|>Paris<|"|>}`,
-				"<tool_call|><",
-				"div>",
-			},
-			wantContent: "<div>",
-			wantTools:   []wantGemma4Tool{{name: "get_weather", argsJSON: `{"location":"Paris"}`}},
-		}),
-		// vLLM: test_streaming_html_argument_does_not_duplicate_tag_prefixes
-		Entry("vLLM: test_streaming_html_argument_does_not_duplicate_tag_prefixes", parseGemma4Case{
-			fragments: []string{
-				"<|tool_call>",
-				"call:write_file{",
-				`path:<|"|>index.html<|"|>,`,
-				`content:<|"|><!DOCTYPE html>` + "\n<",
-				`html lang="zh-CN">` + "\n<",
-				"head>\n    <",
-				`meta charset="UTF-8">` + "\n    <",
-				`meta name="viewport" content="width=device-width">` + "\n",
-				`<|"|>}`,
-				"<tool_call|>",
-			},
-			wantTools: []wantGemma4Tool{{
-				name:     "write_file",
-				argsJSON: `{"path":"index.html","content":"<!DOCTYPE html>\n<html lang=\"zh-CN\">\n<head>\n    <meta charset=\"UTF-8\">\n    <meta name=\"viewport\" content=\"width=device-width\">\n"}`,
-			}},
-		}),
-		// vLLM: test_streaming_single_chunk_complete_tool_call
-		Entry("vLLM: test_streaming_single_chunk_complete_tool_call", parseGemma4Case{
-			fragments: []string{`<|tool_call>call:name_a_color{color_hex:<|"|>00ff11<|"|>}<tool_call|>`},
-			wantTools: []wantGemma4Tool{{name: "name_a_color", argsJSON: `{"color_hex":"00ff11"}`}},
-		}),
-		// vLLM: test_streaming_multi_chunk_batched_tool_calls (two complete
-		// calls in ONE fragment; both must come out with distinct indices)
-		Entry("vLLM: test_streaming_multi_chunk_batched_tool_calls", parseGemma4Case{
-			fragments: []string{
-				`<|tool_call>call:get_weather{location:<|"|>London<|"|>}<tool_call|>` +
-					`<|tool_call>call:get_time{timezone:<|"|>GMT<|"|>}<tool_call|>`,
-			},
-			wantTools: []wantGemma4Tool{
-				{name: "get_weather", argsJSON: `{"location":"London"}`},
-				{name: "get_time", argsJSON: `{"timezone":"GMT"}`},
-			},
-		}),
-		// vLLM: test_streaming_trailing_bare_bool_not_duplicated
-		Entry("vLLM: test_streaming_trailing_bare_bool_not_duplicated", parseGemma4Case{
-			fragments: []string{
-				"<|tool_call>",
-				"call:Edit{",
-				`file_path:<|"|>src/env.py<|"|>,`,
-				`old_string:<|"|>old_val<|"|>,`,
-				`new_string:<|"|>new_val<|"|>,`,
-				"replace_all:",
-				"false}",
-				"<tool_call|>",
-			},
-			wantTools: []wantGemma4Tool{{
-				name:     "Edit",
-				argsJSON: `{"file_path":"src/env.py","old_string":"old_val","new_string":"new_val","replace_all":false}`,
-			}},
-		}),
-
-		// --- implicit reasoning end on <|tool_call> (vLLM is_reasoning_end:
-		// a tool_call token means reasoning is over) -----------------------------
-		Entry("tool call inside an open thought channel ends the reasoning", parseGemma4Case{
-			startInThought: true,
-			fragments:      []string{`need the weather<|tool_call>call:get_weather{location:<|"|>Rome<|"|>}<tool_call|>`},
-			wantReasoning:  "need the weather",
-			wantTools:      []wantGemma4Tool{{name: "get_weather", argsJSON: `{"location":"Rome"}`}},
-		}),
-
-		// --- (12) empty fragments are no-ops --------------------------------------
-		Entry("empty fragments are no-ops", parseGemma4Case{
-			fragments:   []string{"", "Hello", "", "", " world", ""},
-			wantContent: "Hello world",
-		}),
-	)
-
-	It("returns no deltas for an empty fragment and after Close", func() {
-		p := NewGemma4Parser(false)
-		Expect(p.Feed("")).To(BeEmpty())
-		Expect(p.Feed("hi")).ToNot(BeEmpty())
-		Expect(p.Close()).To(BeEmpty()) // nothing held back
-		// The parser is finished after Close: further input is dropped.
-		Expect(p.Feed("more")).To(BeEmpty())
-		Expect(p.Close()).To(BeEmpty())
-	})
-
-	It("generates index-based tool call ids (call_<index>)", func() {
-		// Mirrors the index-based id convention of pkg/grpc/rich_test.go and
-		// keeps ids deterministic for the split-invariance property below.
-		deltas := parseGemma4Fragments(false, []string{
-			`<|tool_call>call:a{}<tool_call|><|tool_call>call:b{}<tool_call|>`,
-		})
-		_, _, tools := flattenGemma4Deltas(deltas)
-		Expect(tools).To(HaveLen(2))
-		Expect(tools[0].id).To(Equal("call_0"))
-		Expect(tools[1].id).To(Equal("call_1"))
-	})
-
-	// Property: for a fixed full output, EVERY 2-split position must yield
-	// exactly the same flattened result as the unsplit parse. This kills
-	// fragment-boundary bugs (mid-marker, mid-delimiter, mid-payload splits).
-	DescribeTable("2-split fragment invariance",
-		func(startInThought bool, full string) {
-			refContent, refReasoning, refTools := flattenGemma4Deltas(
-				parseGemma4Fragments(startInThought, []string{full}))
-			for i := 0; i <= len(full); i++ {
-				content, reasoning, tools := flattenGemma4Deltas(
-					parseGemma4Fragments(startInThought, []string{full[:i], full[i:]}))
-				Expect(content).To(Equal(refContent), fmt.Sprintf("content diverged at split %d", i))
-				Expect(reasoning).To(Equal(refReasoning), fmt.Sprintf("reasoning diverged at split %d", i))
-				Expect(tools).To(Equal(refTools), fmt.Sprintf("tool calls diverged at split %d", i))
-			}
-		},
-		Entry("thought + content + two tool calls + turn end", false,
-			"<|channel>thought\nPondering the request...\n<channel|>Sure - calling tools now. "+
-				`<|tool_call>call:get_weather{location:<|"|>Paris, France<|"|>,unit:<|"|>celsius<|"|>,days:3,detailed:true}<tool_call|>`+
-				`<|tool_call>call:get_time{timezone:<|"|>Europe/Lisbon<|"|>,nested:{flag:false,vals:[1,2.5,<|"|>x<|"|>]}}<tool_call|>`+
-				"Done.<turn|>ignored tail"),
-		Entry("startInThought + tool call + trailing partial marker", true,
-			`Deep thought<channel|>final answer <|tool_call>call:noop{}<tool_call|> trailing <|tool`),
-		Entry("malformed payload fallback", false,
-			`pre <|tool_call>not a call<tool_call|> post`),
-	)
-})
-
-// Decoder-level ports of vLLM's TestParseGemma4Args / TestParseGemma4Array
-// (non-partial mode; the partial-withholding tests do not apply because this
-// parser only ever decodes COMPLETE payloads, see gemma4_parser.go).
-var _ = Describe("decodeGemma4Args", func() {
-	DescribeTable("decodes the gemma4 call syntax into JSON arguments",
-		func(in, wantJSON string) {
-			Expect(decodeGemma4Args(in, 0)).To(MatchJSON(wantJSON))
-		},
-		// vLLM: test_empty_string / test_whitespace_only
-		Entry("empty string", "", `{}`),
-		Entry("whitespace only", "   ", `{}`),
-		// vLLM: test_single_string_value
-		Entry("single string value", `location:<|"|>Paris<|"|>`, `{"location":"Paris"}`),
-		// vLLM: test_string_value_with_comma
-		Entry("string value with comma", `location:<|"|>Paris, France<|"|>`, `{"location":"Paris, France"}`),
-		// vLLM: test_multiple_string_values
-		Entry("multiple string values", `location:<|"|>San Francisco<|"|>,unit:<|"|>celsius<|"|>`, `{"location":"San Francisco","unit":"celsius"}`),
-		// vLLM: test_integer_value / test_float_value
-		Entry("integer value", "count:42", `{"count":42}`),
-		Entry("float value", "score:3.14", `{"score":3.14}`),
-		// vLLM: test_boolean_true / test_boolean_false
-		Entry("boolean true", "flag:true", `{"flag":true}`),
-		Entry("boolean false", "flag:false", `{"flag":false}`),
-		// vLLM: test_null_value (bare null must become JSON null, not "null")
-		Entry("null value", "param:null", `{"param":null}`),
-		// vLLM: test_mixed_types
-		Entry("mixed types", `name:<|"|>test<|"|>,count:42,active:true,score:3.14`,
-			`{"name":"test","count":42,"active":true,"score":3.14}`),
-		// vLLM: test_nested_object
-		Entry("nested object", `nested:{inner:<|"|>value<|"|>}`, `{"nested":{"inner":"value"}}`),
-		// vLLM: test_array_of_strings
-		Entry("array of strings", `items:[<|"|>a<|"|>,<|"|>b<|"|>]`, `{"items":["a","b"]}`),
-		// vLLM: test_unterminated_string (take everything after the delimiter)
-		Entry("unterminated string", `key:<|"|>unterminated`, `{"key":"unterminated"}`),
-		// vLLM: test_empty_value (key with no value after colon)
-		Entry("empty value", "key:", `{"key":""}`),
-		// vLLM: test_trailing_dot_float_partial_withheld, non-partial branch
-		// (trailing-dot floats parse normally outside streaming).
-		Entry("trailing dot float, complete payload", "left:108.,right:22.8", `{"left":108.0,"right":22.8}`),
-	)
-
-	It("terminates and yields valid JSON on malformed input", func() {
-		// vLLM: test_malformed_partial_array (the assertion there is only
-		// "returns a dict without hanging"; ours is "valid JSON object").
-		out := decodeGemma4Args(":[t:[]", 0)
-		var v map[string]any
-		Expect(json.Unmarshal([]byte(out), &v)).To(Succeed())
-	})
-
-	It("degrades nesting beyond the recursion cap to a string value", func() {
-		// 200 levels of a:{a:{...a:1...}}. Without the depth cap the mutual
-		// recursion would grow the stack with the model's output; a Go stack
-		// overflow is a fatal process kill, so levels past gemma4MaxArgsDepth
-		// must gracefully fall back to the raw inner text as a JSON string.
-		const depth = 200
-		body := strings.Repeat("a:{", depth-1) + "a:1" + strings.Repeat("}", depth-1)
-		out := decodeGemma4Args(body, 0)
-		var v map[string]any
-		Expect(json.Unmarshal([]byte(out), &v)).To(Succeed())
-		levels := 0
-		var cur any = v
-		for {
-			m, ok := cur.(map[string]any)
-			if !ok {
-				break
-			}
-			Expect(m).To(HaveKey("a"))
-			cur = m["a"]
-			levels++
-		}
-		Expect(levels).To(Equal(gemma4MaxArgsDepth + 1))
-		Expect(cur).To(BeAssignableToTypeOf(""))
-		Expect(cur).To(ContainSubstring("a:{"))
-	})
-})
-
-var _ = Describe("decodeGemma4Array", func() {
-	DescribeTable("decodes gemma4 array bodies into JSON arrays",
-		func(in, wantJSON string) {
-			Expect(decodeGemma4Array(in, 0)).To(MatchJSON(wantJSON))
-		},
-		// vLLM: test_string_array / test_empty_array / test_bare_values
-		Entry("string array", `<|"|>a<|"|>,<|"|>b<|"|>`, `["a","b"]`),
-		Entry("empty array", "", `[]`),
-		Entry("bare values", "42,true,3.14", `[42,true,3.14]`),
-		// vLLM: test_string_element_with_closing_bracket (a ']' inside a
-		// delimited string must not close the array)
-		Entry("string element with closing bracket", `[<|"|>a]b<|"|>,<|"|>c<|"|>],<|"|>tail<|"|>`, `[["a]b","c"],"tail"]`),
-		// vLLM: test_stray_closing_bracket (no-progress abort, keep prefix)
-		Entry("stray closing bracket", "42,]trailing", `[42]`),
-	)
-})

backend/go/dllm/gemma4_renderer_test.go

@@ -1,406 +0,0 @@
-package main
-
-// Renderer specs for RenderGemma4 against the canonical gemma4 chat template
-// (see the normative template comment in gemma4_renderer.go).
-//
-// Fixture provenance:
-//   - "single user message" and "enable_thinking" are the EXACT expected
-//     decodes from transformers tests/models/diffusion_gemma/
-//     test_modeling_diffusion_gemma.py (test_diffusion_gemma_chat_template
-//     and ..._with_thinking) with ONE difference: the transformers fixtures
-//     start with "<bos>" because apply_chat_template tokenizes the rendered
-//     text with add_bos. Our prompt goes through dllm_capi_generate, whose
-//     run_generate already tokenizes with prepend_bos = vocab.add_bos
-//     (dllm.cpp src/capi.cpp:230-231, true for gemma4), so the renderer must
-//     NOT emit a literal <bos> (it would double) and every expected string
-//     here drops that leading token.
-//   - All other expected strings were produced by rendering the verbatim
-//     GGUF template with jinja2 3.1.2 (bos_token="<bos>") and dropping the
-//     leading "<bos>" for the same reason.
-
-import (
-	. "github.com/onsi/ginkgo/v2"
-	. "github.com/onsi/gomega"
-
-	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
-)
-
-// Two-function tools array used by the tool fixtures (OpenAI wire shape, as
-// LocalAI passes it through PredictOptions.Tools).
-const testToolsJSON = `[{"type":"function","function":{"name":"get_weather","description":"Get the current weather in a location.","parameters":{"type":"object","properties":{"location":{"type":"string","description":"The city name."},"unit":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location"]}}},{"type":"function","function":{"name":"get_time","description":"Get the current time in a timezone.","parameters":{"type":"object","properties":{"timezone":{"type":"string","description":"IANA timezone name."}},"required":["timezone"]}}}]`
-
-// The <|tool>...<tool|> block the template renders for testToolsJSON inside
-// the system turn (jinja2-verified).
-const testToolsBlock = `<|tool>declaration:get_weather{description:<|"|>Get the current weather in a location.<|"|>,parameters:{properties:{location:{description:<|"|>The city name.<|"|>,type:<|"|>STRING<|"|>},unit:{enum:[<|"|>celsius<|"|>,<|"|>fahrenheit<|"|>],type:<|"|>STRING<|"|>}},required:[<|"|>location<|"|>],type:<|"|>OBJECT<|"|>}}<tool|><|tool>declaration:get_time{description:<|"|>Get the current time in a timezone.<|"|>,parameters:{properties:{timezone:{description:<|"|>IANA timezone name.<|"|>,type:<|"|>STRING<|"|>}},required:[<|"|>timezone<|"|>],type:<|"|>OBJECT<|"|>}}<tool|>`
-
-// A single tool exercising the deep format_parameters branches: array items
-// (string-typed and nested-array), nullable, enum+nullable, nested object
-// properties/required, and a response declaration.
-const complexToolsJSON = `[{"type":"function","function":{"name":"complex_tool","description":"A complex tool.","parameters":{"type":"object","properties":{"tags":{"type":"array","description":"Tags.","items":{"type":"string"}},"matrix":{"type":"array","items":{"type":"array","items":{"type":"number"}}},"opts":{"type":"object","description":"Options.","properties":{"depth":{"type":"integer","nullable":true}},"required":["depth"]},"mode":{"type":"string","enum":["a","b"],"nullable":true}},"required":["tags","opts"]},"response":{"description":"The result.","type":"object"}}}]`
-
-// jinja2-verified render of complexToolsJSON. Notable template quirks pinned
-// here: nested array items go through format_argument with ESCAPED keys and
-// an un-uppercased type (<|"|>type<|"|>:<|"|>number<|"|>), while direct item
-// types are uppercased; properties dictsort case-insensitively.
-const complexToolsBlock = `<|tool>declaration:complex_tool{description:<|"|>A complex tool.<|"|>,parameters:{properties:{matrix:{items:{items:{<|"|>type<|"|>:<|"|>number<|"|>},type:<|"|>ARRAY<|"|>},type:<|"|>ARRAY<|"|>},mode:{enum:[<|"|>a<|"|>,<|"|>b<|"|>],nullable:true,type:<|"|>STRING<|"|>},opts:{description:<|"|>Options.<|"|>,properties:{depth:{nullable:true,type:<|"|>INTEGER<|"|>}},required:[<|"|>depth<|"|>],type:<|"|>OBJECT<|"|>},tags:{description:<|"|>Tags.<|"|>,items:{type:<|"|>STRING<|"|>},type:<|"|>ARRAY<|"|>}},required:[<|"|>tags<|"|>,<|"|>opts<|"|>],type:<|"|>OBJECT<|"|>},response:{description:<|"|>The result.<|"|>,type:<|"|>OBJECT<|"|>}}<tool|>`
-
-type renderGemma4Case struct {
-	msgs      []*pb.Message
-	toolsJSON string
-	// nImages mirrors len(PredictOptions.Images): the OpenAI layer strips
-	// image content parts out of the messages, so the renderer re-injects
-	// one engine marker per image on the last user message (see the IMAGE
-	// NOTE on RenderGemma4).
-	nImages            int
-	enableThinking     bool
-	noGenerationPrompt bool // inverted so the zero value is the common case
-	expected           string
-}
-
-var _ = Describe("RenderGemma4", func() {
-	DescribeTable("renders the canonical gemma4 prompt",
-		func(c renderGemma4Case) {
-			out, err := RenderGemma4(c.msgs, c.toolsJSON, c.nImages, c.enableThinking, !c.noGenerationPrompt)
-			Expect(err).ToNot(HaveOccurred())
-			Expect(out).To(Equal(c.expected))
-			// The C-ABI generate prepends BOS itself: a literal <bos>
-			// anywhere in the rendered prompt would double-encode it.
-			Expect(out).ToNot(ContainSubstring("<bos>"))
-		},
-
-		// transformers fixture (test_diffusion_gemma_chat_template), sans <bos>:
-		// default thinking pre-opens an EMPTY thought channel in the
-		// generation prompt.
-		Entry("single user message, default (no thinking)", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: "Write a long essay about Portugal."},
-			},
-			expected: "<|turn>user\nWrite a long essay about Portugal.<turn|>\n<|turn>model\n<|channel>thought\n<channel|>",
-		}),
-
-		// transformers fixture (test_diffusion_gemma_chat_template_with_thinking),
-		// sans <bos>: a system turn carrying <|think|> and NO auto-opened
-		// thought channel.
-		Entry("enable_thinking=true", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: "Write a long essay about Portugal."},
-			},
-			enableThinking: true,
-			expected:       "<|turn>system\n<|think|>\n<turn|>\n<|turn>user\nWrite a long essay about Portugal.<turn|>\n<|turn>model\n",
-		}),
-
-		Entry("multi-turn user/assistant/user", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: "Hello, who are you?"},
-				{Role: "assistant", Content: "I am Gemma, a helpful assistant."},
-				{Role: "user", Content: "Tell me a joke."},
-			},
-			expected: "<|turn>user\nHello, who are you?<turn|>\n<|turn>model\nI am Gemma, a helpful assistant.<turn|>\n<|turn>user\nTell me a joke.<turn|>\n<|turn>model\n<|channel>thought\n<channel|>",
-		}),
-
-		// tpl L178-L195: a leading system message is folded into the system
-		// turn (trimmed) and consumed from the loop.
-		Entry("system message folds into the system turn", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "system", Content: "You are a pirate."},
-				{Role: "user", Content: "Hello!"},
-			},
-			expected: "<|turn>system\nYou are a pirate.<turn|>\n<|turn>user\nHello!<turn|>\n<|turn>model\n<|channel>thought\n<channel|>",
-		}),
-
-		// tpl L182-L185: <|think|> goes at the very top of the SAME system
-		// turn, before the system prompt text.
-		Entry("system message with enable_thinking shares the turn", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "system", Content: "You are a pirate."},
-				{Role: "user", Content: "Hello!"},
-			},
-			enableThinking: true,
-			expected:       "<|turn>system\n<|think|>\nYou are a pirate.<turn|>\n<|turn>user\nHello!<turn|>\n<|turn>model\n",
-		}),
-
-		// tpl L196-L203: tool declarations render in the system turn, one
-		// <|tool>declaration:...<tool|> block per tool, no separators.
-		Entry("tools array (two functions)", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: "What is the weather in Tokyo?"},
-			},
-			toolsJSON: testToolsJSON,
-			expected:  "<|turn>system\n" + testToolsBlock + "<turn|>\n<|turn>user\nWhat is the weather in Tokyo?<turn|>\n<|turn>model\n<|channel>thought\n<channel|>",
-		}),
-
-		// format_parameters deep branches (tpl L1-L85) + response declaration
-		// (tpl L106-L116).
-		Entry("complex tool schema (array items, nullable, nested object, response)", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: "go"},
-			},
-			toolsJSON: complexToolsJSON,
-			expected:  "<|turn>system\n" + complexToolsBlock + "<turn|>\n<|turn>user\ngo<turn|>\n<|turn>model\n<|channel>thought\n<channel|>",
-		}),
-
-		// tpl L243-L313: assistant tool_calls render as
-		// <|tool_call>call:name{args}<tool_call|>; the following role=tool
-		// message renders inline as <|tool_response>response:name{value:..}
-		// <tool_response|>; the model turn stays OPEN (no <turn|>, no new
-		// generation prompt) so the model continues after the response.
-		Entry("assistant tool_calls + role=tool result", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: "What is the weather in Tokyo?"},
-				{Role: "assistant", Content: "", ToolCalls: `[{"index":0,"id":"call_1","type":"function","function":{"name":"get_weather","arguments":"{\"location\":\"Tokyo\",\"unit\":\"celsius\"}"}}]`},
-				{Role: "tool", ToolCallId: "call_1", Content: "Sunny, 22 degrees celsius."},
-			},
-			toolsJSON: testToolsJSON,
-			expected:  "<|turn>system\n" + testToolsBlock + "<turn|>\n<|turn>user\nWhat is the weather in Tokyo?<turn|>\n<|turn>model\n" + `<|tool_call>call:get_weather{location:<|"|>Tokyo<|"|>,unit:<|"|>celsius<|"|>}<tool_call|><|tool_response>response:get_weather{value:<|"|>Sunny, 22 degrees celsius.<|"|>}<tool_response|>`,
-		}),
-
-		// tpl L348-L349: a tool_calls turn with no rendered responses ends
-		// on an OPEN <|tool_response> marker for the runtime to fill, and
-		// add_generation_prompt adds nothing (tpl L357).
-		Entry("assistant tool_calls without a result leaves <|tool_response> open", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: "What is the weather in Tokyo?"},
-				{Role: "assistant", Content: "", ToolCalls: `[{"index":0,"id":"call_1","type":"function","function":{"name":"get_weather","arguments":"{\"location\":\"Tokyo\",\"unit\":\"celsius\"}"}}]`},
-			},
-			toolsJSON: testToolsJSON,
-			expected:  "<|turn>system\n" + testToolsBlock + "<turn|>\n<|turn>user\nWhat is the weather in Tokyo?<turn|>\n<|turn>model\n" + `<|tool_call>call:get_weather{location:<|"|>Tokyo<|"|>,unit:<|"|>celsius<|"|>}<tool_call|><|tool_response>`,
-		}),
-
-		// tpl L237-L241: reasoning_content renders as a thought channel only
-		// on a tool-calling turn after the last user message.
-		Entry("reasoning_content with tool_calls renders the thought channel", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: "weather?"},
-				{Role: "assistant", Content: "", ReasoningContent: "I should call the tool", ToolCalls: `[{"index":0,"id":"c1","type":"function","function":{"name":"get_weather","arguments":"{\"location\":\"Tokyo\"}"}}]`},
-				{Role: "tool", ToolCallId: "c1", Content: "Sunny"},
-			},
-			expected: "<|turn>user\nweather?<turn|>\n<|turn>model\n<|channel>thought\nI should call the tool\n<channel|>" + `<|tool_call>call:get_weather{location:<|"|>Tokyo<|"|>}<tool_call|><|tool_response>response:get_weather{value:<|"|>Sunny<|"|>}<tool_response|>`,
-		}),
-
-		// tpl L220-L235: the assistant answer following its own tool round
-		// continues the SAME model turn (no second <|turn>model).
-		Entry("tool round then final assistant answer then user", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: "weather?"},
-				{Role: "assistant", Content: "", ToolCalls: `[{"index":0,"id":"c1","type":"function","function":{"name":"get_weather","arguments":"{\"location\":\"Tokyo\"}"}}]`},
-				{Role: "tool", ToolCallId: "c1", Content: "Sunny"},
-				{Role: "assistant", Content: "It is sunny."},
-				{Role: "user", Content: "thanks"},
-			},
-			expected: "<|turn>user\nweather?<turn|>\n<|turn>model\n" + `<|tool_call>call:get_weather{location:<|"|>Tokyo<|"|>}<tool_call|><|tool_response>response:get_weather{value:<|"|>Sunny<|"|>}<tool_response|>` + "It is sunny.<turn|>\n<|turn>user\nthanks<turn|>\n<|turn>model\n<|channel>thought\n<channel|>",
-		}),
-
-		// format_argument (tpl L118-L147): numbers keep their JSON literal,
-		// booleans lower-case, nested maps have unquoted dictsorted keys,
-		// arrays bracketed; top-level args are dictsorted case-insensitively.
-		Entry("tool_call argument types (number/bool/nested/array)", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: "go"},
-				{Role: "assistant", Content: "", ToolCalls: `[{"index":0,"id":"c1","type":"function","function":{"name":"f","arguments":"{\"count\":42,\"ratio\":3.5,\"flag\":true,\"off\":false,\"nested\":{\"x\":\"y\",\"n\":7},\"list\":[\"a\",1,true]}"}}]`},
-			},
-			expected: "<|turn>user\ngo<turn|>\n<|turn>model\n" + `<|tool_call>call:f{count:42,flag:true,list:[<|"|>a<|"|>,1,true],nested:{n:7,x:<|"|>y<|"|>},off:false,ratio:3.5}<tool_call|><|tool_response>`,
-		}),
-
-		// jinja dictsort is case-insensitive: alpha sorts before Beta.
-		Entry("tool_call argument dictsort is case-insensitive", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: "go"},
-				{Role: "assistant", Content: "", ToolCalls: `[{"index":0,"id":"c1","type":"function","function":{"name":"f","arguments":"{\"Beta\":1,\"alpha\":2}"}}]`},
-			},
-			expected: "<|turn>user\ngo<turn|>\n<|turn>model\n<|tool_call>call:f{alpha:2,Beta:1}<tool_call|><|tool_response>",
-		}),
-
-		// jinja renders Python None as "None" (round-trips through vLLM's
-		// parser, which lowers "none" back to null).
-		Entry("tool_call null argument renders as None", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: "go"},
-				{Role: "assistant", Content: "", ToolCalls: `[{"index":0,"id":"c1","type":"function","function":{"name":"f","arguments":"{\"maybe\":null}"}}]`},
-			},
-			expected: "<|turn>user\ngo<turn|>\n<|turn>model\n<|tool_call>call:f{maybe:None}<tool_call|><|tool_response>",
-		}),
-
-		Entry("tool_call empty arguments render empty braces", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: "go"},
-				{Role: "assistant", Content: "", ToolCalls: `[{"index":0,"id":"c1","type":"function","function":{"name":"f","arguments":"{}"}}]`},
-			},
-			expected: "<|turn>user\ngo<turn|>\n<|turn>model\n<|tool_call>call:f{}<tool_call|><|tool_response>",
-		}),
-
-		// tpl L253-L254: a non-object arguments string renders verbatim.
-		Entry("tool_call non-object string arguments render verbatim", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: "go"},
-				{Role: "assistant", Content: "", ToolCalls: `[{"index":0,"id":"c1","type":"function","function":{"name":"f","arguments":"just text"}}]`},
-			},
-			expected: "<|turn>user\ngo<turn|>\n<|turn>model\n<|tool_call>call:f{just text}<tool_call|><|tool_response>",
-		}),
-
-		// tpl L278-L285: unmatched tool_call_id falls back to the tool
-		// message's own name.
-		Entry("tool result name falls back when tool_call_id does not match", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: "go"},
-				{Role: "assistant", Content: "", ToolCalls: `[{"index":0,"id":"c1","type":"function","function":{"name":"f","arguments":"{}"}}]`},
-				{Role: "tool", ToolCallId: "OTHER", Name: "named_tool", Content: "out"},
-			},
-			expected: "<|turn>user\ngo<turn|>\n<|turn>model\n" + `<|tool_call>call:f{}<tool_call|><|tool_response>response:named_tool{value:<|"|>out<|"|>}<tool_response|>`,
-		}),
-
-		// strip_thinking (tpl L148-L158): historical assistant content loses
-		// its <|channel>...<channel|> spans.
-		Entry("assistant content thinking channels are stripped", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: "hi"},
-				{Role: "assistant", Content: "<|channel>thought\nsecret\n<channel|>visible answer"},
-				{Role: "user", Content: "more"},
-			},
-			expected: "<|turn>user\nhi<turn|>\n<|turn>model\nvisible answer<turn|>\n<|turn>user\nmore<turn|>\n<|turn>model\n<|channel>thought\n<channel|>",
-		}),
-
-		// tpl L220-L235: consecutive assistant messages suppress the second
-		// <|turn>model (continuation), but each still closes with <turn|>.
-		Entry("consecutive assistant messages continue the model turn", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: "hi"},
-				{Role: "assistant", Content: "part one"},
-				{Role: "assistant", Content: "part two"},
-				{Role: "user", Content: "ok"},
-			},
-			expected: "<|turn>user\nhi<turn|>\n<|turn>model\npart one<turn|>\npart two<turn|>\n<|turn>user\nok<turn|>\n<|turn>model\n<|channel>thought\n<channel|>",
-		}),
-
-		Entry("add_generation_prompt=false renders no model turn", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: "hi"},
-			},
-			noGenerationPrompt: true,
-			expected:           "<|turn>user\nhi<turn|>\n",
-		}),
-
-		// One engine marker per image, appended directly after the user
-		// text with no separator (tpl L323-L341 emits parts back-to-back;
-		// "<image>" is dllm_capi.h's splice marker, not the template's
-		// <|image|> text token - see the IMAGE NOTE on RenderGemma4).
-		Entry("one image appends one engine marker to the user message", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: "What is in this picture?"},
-			},
-			nImages:  1,
-			expected: "<|turn>user\nWhat is in this picture?<image><turn|>\n<|turn>model\n<|channel>thought\n<channel|>",
-		}),
-
-		Entry("multiple images append markers in image order", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: "Compare these."},
-			},
-			nImages:  3,
-			expected: "<|turn>user\nCompare these.<image><image><image><turn|>\n<|turn>model\n<|channel>thought\n<channel|>",
-		}),
-
-		// Flattened delivery loses per-message attribution, so all images
-		// attach to the LAST user message (llama.cpp grpc-server convention).
-		Entry("images attach to the last user message in multi-turn", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: "hi"},
-				{Role: "assistant", Content: "hello"},
-				{Role: "user", Content: "and this?"},
-			},
-			nImages:  1,
-			expected: "<|turn>user\nhi<turn|>\n<|turn>model\nhello<turn|>\n<|turn>user\nand this?<image><turn|>\n<|turn>model\n<|channel>thought\n<channel|>",
-		}),
-
-		// tpl L346: the markers count as captured_content, so an image-only
-		// user message still has content and closes its turn normally.
-		Entry("image with empty user text still closes the turn", renderGemma4Case{
-			msgs: []*pb.Message{
-				{Role: "user", Content: ""},
-			},
-			nImages:  1,
-			expected: "<|turn>user\n<image><turn|>\n<|turn>model\n<|channel>thought\n<channel|>",
-		}),
-	)
-
-	Describe("error handling", func() {
-		It("fails loud on an unknown role", func() {
-			_, err := RenderGemma4([]*pb.Message{
-				{Role: "narrator", Content: "Meanwhile..."},
-			}, "", 0, false, true)
-			Expect(err).To(HaveOccurred())
-			Expect(err.Error()).To(ContainSubstring(`unknown role "narrator"`))
-		})
-
-		It("fails on invalid tools JSON", func() {
-			_, err := RenderGemma4([]*pb.Message{
-				{Role: "user", Content: "hi"},
-			}, "{not json", 0, false, true)
-			Expect(err).To(HaveOccurred())
-			Expect(err.Error()).To(ContainSubstring("tools JSON"))
-		})
-
-		It("fails on invalid tool_calls JSON", func() {
-			_, err := RenderGemma4([]*pb.Message{
-				{Role: "user", Content: "hi"},
-				{Role: "assistant", Content: "", ToolCalls: "{not json"},
-			}, "", 0, false, true)
-			Expect(err).To(HaveOccurred())
-			Expect(err.Error()).To(ContainSubstring("tool_calls JSON"))
-		})
-
-		It("fails on an orphan tool message, naming its index", func() {
-			// A role:tool message with no preceding assistant tool_calls turn
-			// would be silently dropped by the jinja; we fail loud instead.
-			_, err := RenderGemma4([]*pb.Message{
-				{Role: "user", Content: "hi"},
-				{Role: "tool", Content: `{"temp": 20}`, ToolCallId: "call_1"},
-			}, "", 0, false, true)
-			Expect(err).To(HaveOccurred())
-			Expect(err.Error()).To(ContainSubstring("orphan tool message 1"))
-		})
-
-		It("fails on trailing garbage after the tools JSON array", func() {
-			_, err := RenderGemma4([]*pb.Message{
-				{Role: "user", Content: "hi"},
-			}, "[] junk", 0, false, true)
-			Expect(err).To(HaveOccurred())
-			Expect(err.Error()).To(ContainSubstring("tools JSON"))
-		})
-
-		It("fails when the tools JSON is not an array", func() {
-			_, err := RenderGemma4([]*pb.Message{
-				{Role: "user", Content: "hi"},
-			}, `{"type":"function"}`, 0, false, true)
-			Expect(err).To(HaveOccurred())
-			Expect(err.Error()).To(ContainSubstring("tools JSON is not an array"))
-		})
-
-		It("fails when a tools array element is not an object", func() {
-			_, err := RenderGemma4([]*pb.Message{
-				{Role: "user", Content: "hi"},
-			}, `[42]`, 0, false, true)
-			Expect(err).To(HaveOccurred())
-			Expect(err.Error()).To(ContainSubstring("tools[0] is not an object"))
-		})
-
-		It("rejects a nil message via the unknown-role check", func() {
-			// Pins current behavior: pb getters are nil-safe, so a nil message
-			// reads as role "" and trips the fail-loud unknown-role guard.
-			_, err := RenderGemma4([]*pb.Message{nil}, "", 0, false, true)
-			Expect(err).To(HaveOccurred())
-			Expect(err.Error()).To(ContainSubstring(`unknown role "" in message 0`))
-		})
-
-		It("fails loud on images with no user message to attach them to", func() {
-			// The engine would reject the markerless prompt anyway
-			// (marker/image count mismatch); the renderer surfaces the bad
-			// request with a usable message instead.
-			_, err := RenderGemma4([]*pb.Message{
-				{Role: "system", Content: "sys"},
-				{Role: "assistant", Content: "hi"},
-			}, "", 1, false, true)
-			Expect(err).To(HaveOccurred())
-			Expect(err.Error()).To(ContainSubstring("no user message"))
-		})
-	})
-})

backend/go/dllm/main.go

@@ -1,98 +0,0 @@
-package main
-
-// Started internally by LocalAI - one gRPC server per loaded model.
-//
-// Loads libdllm.so via purego and registers the flat C-ABI declared in
-// dllm.cpp's include/dllm_capi.h (ABI v1): 9 mandatory symbols plus the
-// Dlsym-probed optional multimodal pair. The library name can
-// be overridden with DLLM_LIBRARY (mirrors the PARAKEET_LIBRARY /
-// WHISPER_LIBRARY convention in the sibling backends); the default looks
-// for the .so next to this binary (run.sh puts the package dir on
-// LD_LIBRARY_PATH).
-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
-}
-
-// loadCAPI dlopens libName and binds the 9 dllm_capi_* entry points 1:1 to
-// dllm_capi.h, so an `nm libdllm.so | grep dllm_capi` is enough to spot
-// drift. Shared with the test suite (ensureLibLoaded), which drives the
-// bridge without the gRPC server.
-//
-// The C-ABI returns malloc'd char* buffers from tokenize_json/generate; we
-// register those as uintptr so we get the raw pointer back and can call
-// dllm_capi_free_string on it (purego's string return would copy and forget
-// the original pointer, leaking it on every call). last_error returns a
-// BORROWED pointer instead, so it is registered as a plain string: purego
-// copies it and nothing must be freed.
-func loadCAPI(libName string) error {
-	lib, err := purego.Dlopen(libName, purego.RTLD_NOW|purego.RTLD_GLOBAL)
-	if err != nil {
-		return fmt.Errorf("dllm: dlopen %q: %w", libName, err)
-	}
-
-	libFuncs := []LibFuncs{
-		{&cppAbiVersion, "dllm_capi_abi_version"},
-		{&cppLoad, "dllm_capi_load"},
-		{&cppFree, "dllm_capi_free"},
-		{&cppLastError, "dllm_capi_last_error"},
-		{&cppFreeString, "dllm_capi_free_string"},
-		{&cppTokenizeJSON, "dllm_capi_tokenize_json"},
-		{&cppGenerate, "dllm_capi_generate"},
-		{&cppGenerateStream, "dllm_capi_generate_stream"},
-		{&cppCancel, "dllm_capi_cancel"},
-	}
-	for _, lf := range libFuncs {
-		purego.RegisterLibFunc(lf.FuncPtr, lib, lf.Name)
-	}
-
-	// Multimodal entry points (dllm_capi.h's P4 surface). Additive: the ABI
-	// version stays 1 and consumers detect the surface by probing the symbols
-	// (the parakeet-cpp optional-symbol pattern), so the backend still loads
-	// against an older text-only libdllm.so - image requests then fail with
-	// errMMUnsupported instead of a boot failure.
-	if sym, err := purego.Dlsym(lib, "dllm_capi_generate_mm"); err == nil && sym != 0 {
-		purego.RegisterLibFunc(&cppGenerateMM, lib, "dllm_capi_generate_mm")
-	}
-	if sym, err := purego.Dlsym(lib, "dllm_capi_generate_stream_mm"); err == nil && sym != 0 {
-		purego.RegisterLibFunc(&cppGenerateStreamMM, lib, "dllm_capi_generate_stream_mm")
-	}
-	return nil
-}
-
-func main() {
-	libName := os.Getenv("DLLM_LIBRARY")
-	if libName == "" {
-		libName = "libdllm.so"
-	}
-
-	if err := loadCAPI(libName); err != nil {
-		panic(err)
-	}
-
-	// Hard-fail on an ABI mismatch: the flat-pointer bindings above would
-	// otherwise misbehave silently against a future libdllm.so.
-	if v := cAbiVersion(); v != dllmABIVersion {
-		panic(fmt.Errorf("dllm: libdllm.so ABI=%d, this backend speaks ABI=%d", v, dllmABIVersion))
-	}
-	fmt.Fprintf(os.Stderr, "[dllm] ABI=%d multimodal=%t\n", cAbiVersion(), cMMSupported())
-
-	flag.Parse()
-
-	if err := grpc.StartServer(*addr, &Dllm{}); err != nil {
-		panic(err)
-	}
-}

backend/go/dllm/package.sh

@@ -1,24 +0,0 @@
-#!/bin/bash
-#
-# T1 packaging stub: copy the binary, run.sh and libdllm.so into package/.
-# The full ldd walk (libc, libstdc++, libgomp, GPU runtimes, arch
-# detection) lands with the registration task, mirroring
-# backend/go/whisper/package.sh.
-
-set -e
-
-CURDIR=$(dirname "$(realpath "$0")")
-
-mkdir -p "$CURDIR/package/lib"
-
-cp -avf "$CURDIR/dllm-grpc" "$CURDIR/package/"
-cp -avf "$CURDIR/run.sh" "$CURDIR/package/"
-
-# libdllm.so + any soname symlinks, should upstream ever add them.
-cp -avf "$CURDIR"/libdllm.so* "$CURDIR/package/lib/" 2>/dev/null || {
-	echo "ERROR: libdllm.so not found in $CURDIR, run 'make' first" >&2
-	exit 1
-}
-
-echo "T1 package layout (full ldd walk lands with registration):"
-ls -liah "$CURDIR/package/" "$CURDIR/package/lib/"

backend/go/dllm/run.sh

@@ -1,16 +0,0 @@
-#!/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
-# whisper / parakeet-cpp backends' runtime layout).
-if [ -f "$CURDIR/lib/ld.so" ]; then
-	echo "Using lib/ld.so"
-	exec "$CURDIR/lib/ld.so" "$CURDIR/dllm-grpc" "$@"
-fi
-
-exec "$CURDIR/dllm-grpc" "$@"

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

@@ -0,0 +1,7 @@
+sources/
+build*/
+package/
+liblocateanythingcpp*.so
+locate-anything-cpp
+test-models/
+test-data/

backend/go/locate-anything-cpp/CMakeLists.txt

@@ -0,0 +1,57 @@
+cmake_minimum_required(VERSION 3.18)
+project(liblocateanythingcpp LANGUAGES C CXX)
+
+set(CMAKE_POSITION_INDEPENDENT_CODE ON)
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+
+# Static-link ggml + locate_anything so the resulting .so has no runtime
+# dependency on extra ggml/locate_anything shared libraries — only on
+# libc/libstdc++/libgomp, which the LocalAI package step bundles into the
+# docker image.
+set(BUILD_SHARED_LIBS OFF CACHE BOOL "Build static libraries" FORCE)
+
+# locate-anything.cpp build switches: skip CLI/tests, keep static lib.
+set(LA_BUILD_CLI OFF CACHE BOOL "Disable locate-anything CLI" FORCE)
+set(LA_BUILD_TESTS OFF CACHE BOOL "Disable locate-anything tests" FORCE)
+set(LA_SHARED OFF CACHE BOOL "Build locate_anything as static lib" FORCE)
+
+# Unlike rt-detr.cpp, locate-anything.cpp ships no in-tree ggml patches, so
+# there is no apply_ggml_patches.sh hook to shim here.
+add_subdirectory(./sources/locate-anything.cpp)
+
+# locate-anything.cpp's top-level CMakeLists points its own target's include
+# dirs at ${CMAKE_SOURCE_DIR}/{include,src,third_party,...}. CMAKE_SOURCE_DIR
+# is the *top-level* source dir of the whole CMake tree, so when we pull it in
+# via add_subdirectory it resolves to OUR directory, not theirs, and the
+# locate_anything target fails to find its own headers (la_capi.h, stb_image.h,
+# la_gguf_keys.h). Re-add the correct, subdir-relative include paths to the
+# already-defined target so it compiles regardless of where it's nested.
+set(LA_SRC ${CMAKE_CURRENT_SOURCE_DIR}/sources/locate-anything.cpp)
+target_include_directories(locate_anything PRIVATE
+    ${LA_SRC}/include
+    ${LA_SRC}/src
+    ${LA_SRC}/third_party
+    ${LA_SRC}/third_party/stb)
+
+# locate-anything.cpp's C-API symbols already live inside liblocate_anything
+# (src/la_capi.cpp is compiled into the lib). We re-export them via a MODULE
+# library that links locate_anything so the symbols are visible at dlopen time.
+add_library(locateanythingcpp MODULE
+    sources/locate-anything.cpp/src/la_capi.cpp)
+
+target_include_directories(locateanythingcpp PRIVATE
+    sources/locate-anything.cpp/include
+    sources/locate-anything.cpp/src
+    sources/locate-anything.cpp/third_party
+    sources/locate-anything.cpp/third_party/stb
+)
+
+target_link_libraries(locateanythingcpp PRIVATE locate_anything ggml)
+
+if(CMAKE_CXX_COMPILER_ID MATCHES "GNU" AND CMAKE_CXX_COMPILER_VERSION VERSION_LESS 9.0)
+    target_link_libraries(locateanythingcpp PRIVATE stdc++fs)
+endif()
+
+set_property(TARGET locateanythingcpp PROPERTY CXX_STANDARD 17)
+set_target_properties(locateanythingcpp PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR})

backend/go/locate-anything-cpp/Makefile

@@ -0,0 +1,134 @@
+CMAKE_ARGS?=
+BUILD_TYPE?=
+NATIVE?=false
+
+GOCMD?=go
+GO_TAGS?=
+JOBS?=$(shell nproc --ignore=1)
+
+# locate-anything.cpp. Pin to a specific commit for a stable build; leaving
+# this on `master` always picks up the latest C-API surface (incl. the
+# per-detection accessor functions used by golocateanythingcpp.go).
+LOCATEANYTHING_REPO?=https://github.com/mudler/locate-anything.cpp.git
+LOCATEANYTHING_VERSION?=92c1682da792c1e8a5dec91acc2be4b02c742ded
+
+ifeq ($(NATIVE),false)
+	CMAKE_ARGS+=-DGGML_NATIVE=OFF
+endif
+
+# Forward LocalAI's BUILD_TYPE to the matching ggml backend switch.
+ifeq ($(BUILD_TYPE),cublas)
+	CMAKE_ARGS+=-DGGML_CUDA=ON -DLA_GGML_CUDA=ON
+else ifeq ($(BUILD_TYPE),openblas)
+	CMAKE_ARGS+=-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
+else ifeq ($(BUILD_TYPE),clblas)
+	CMAKE_ARGS+=-DGGML_CLBLAST=ON
+else ifeq ($(BUILD_TYPE),hipblas)
+	ROCM_HOME ?= /opt/rocm
+	ROCM_PATH ?= /opt/rocm
+	export CXX=$(ROCM_HOME)/llvm/bin/clang++
+	export CC=$(ROCM_HOME)/llvm/bin/clang
+	AMDGPU_TARGETS?=gfx908,gfx90a,gfx942,gfx950,gfx1030,gfx1100,gfx1101,gfx1102,gfx1200,gfx1201
+	CMAKE_ARGS+=-DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=$(AMDGPU_TARGETS)
+else ifeq ($(BUILD_TYPE),vulkan)
+	CMAKE_ARGS+=-DGGML_VULKAN=ON -DLA_GGML_VULKAN=ON
+else ifeq ($(OS),Darwin)
+	ifneq ($(BUILD_TYPE),metal)
+		CMAKE_ARGS+=-DGGML_METAL=OFF
+	else
+		CMAKE_ARGS+=-DGGML_METAL=ON
+		CMAKE_ARGS+=-DGGML_METAL_EMBED_LIBRARY=ON
+		CMAKE_ARGS+=-DLA_GGML_METAL=ON
+	endif
+endif
+
+ifeq ($(BUILD_TYPE),sycl_f16)
+	CMAKE_ARGS+=-DGGML_SYCL=ON \
+		-DCMAKE_C_COMPILER=icx \
+		-DCMAKE_CXX_COMPILER=icpx \
+		-DGGML_SYCL_F16=ON
+endif
+
+ifeq ($(BUILD_TYPE),sycl_f32)
+	CMAKE_ARGS+=-DGGML_SYCL=ON \
+		-DCMAKE_C_COMPILER=icx \
+		-DCMAKE_CXX_COMPILER=icpx
+endif
+
+sources/locate-anything.cpp:
+	mkdir -p sources && \
+	git clone --recursive $(LOCATEANYTHING_REPO) sources/locate-anything.cpp && \
+	cd sources/locate-anything.cpp && \
+	git checkout $(LOCATEANYTHING_VERSION) && \
+	git submodule update --init --recursive --depth 1 --single-branch
+
+# Detect OS
+UNAME_S := $(shell uname -s)
+
+# Only build CPU variants on Linux
+ifeq ($(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
+endif
+
+locate-anything-cpp: main.go golocateanythingcpp.go $(VARIANT_TARGETS)
+	CGO_ENABLED=0 $(GOCMD) build -tags "$(GO_TAGS)" -o locate-anything-cpp ./
+
+package: locate-anything-cpp
+	bash package.sh
+
+build: package
+
+clean: purge
+	rm -rf liblocateanythingcpp*.so locate-anything-cpp package sources
+
+purge:
+	rm -rf build*
+
+# Build all variants (Linux only)
+ifeq ($(UNAME_S),Linux)
+liblocateanythingcpp-avx.so: sources/locate-anything.cpp
+	rm -rfv build-$@
+	$(info ${GREEN}I locate-anything-cpp build info:avx${RESET})
+	SO_TARGET=$@ CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off" $(MAKE) liblocateanythingcpp-custom
+	rm -rfv build-$@
+
+liblocateanythingcpp-avx2.so: sources/locate-anything.cpp
+	rm -rfv build-$@
+	$(info ${GREEN}I locate-anything-cpp build info:avx2${RESET})
+	SO_TARGET=$@ CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on -DGGML_BMI2=on" $(MAKE) liblocateanythingcpp-custom
+	rm -rfv build-$@
+
+liblocateanythingcpp-avx512.so: sources/locate-anything.cpp
+	rm -rfv build-$@
+	$(info ${GREEN}I locate-anything-cpp build info:avx512${RESET})
+	SO_TARGET=$@ CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=on -DGGML_FMA=on -DGGML_F16C=on -DGGML_BMI2=on" $(MAKE) liblocateanythingcpp-custom
+	rm -rfv build-$@
+endif
+
+# Build fallback variant (all platforms)
+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" $(MAKE) liblocateanythingcpp-custom
+	rm -rfv build-$@
+
+liblocateanythingcpp-custom: CMakeLists.txt
+	mkdir -p build-$(SO_TARGET) && \
+	cd build-$(SO_TARGET) && \
+	cmake .. $(CMAKE_ARGS) && \
+	cmake --build . --config Release -j$(JOBS) && \
+	cd .. && \
+	mv build-$(SO_TARGET)/liblocateanythingcpp.so ./$(SO_TARGET)
+
+all: locate-anything-cpp package
+
+# `test` is invoked by the top-level Makefile's `test-extra` target. It builds
+# the backend binary + the fallback shared library (needed for dlopen at
+# runtime), then runs test.sh which downloads the q8_0 GGUF + COCO image and
+# exercises the gRPC Load/Detect wire path via the Go smoke test in
+# main_test.go.
+test: locate-anything-cpp liblocateanythingcpp-fallback.so
+	bash test.sh

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

@@ -0,0 +1,174 @@
+package main
+
+// golocateanythingcpp.go - gRPC handlers (Load, Detect) for the
+// locate-anything-cpp backend.
+//
+// Embeds base.SingleThread to default unimplemented RPCs to "not supported"
+// while we only implement open-vocabulary object detection (Detect).
+
+import (
+	"encoding/base64"
+	"fmt"
+	"os"
+	"path/filepath"
+	"unsafe"
+
+	"github.com/mudler/LocalAI/pkg/grpc/base"
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+)
+
+// la_ctx* is an opaque handle. la_capi_load returns it directly (0 == failure),
+// unlike rfdetr's out-parameter convention.
+var (
+	// la_capi_load(const char* gguf_path, int n_threads) -> la_ctx* (0 = fail)
+	CapiLoad func(gguf string, nThreads int32) uintptr
+	// la_capi_free(la_ctx* ctx)
+	CapiFree func(handle uintptr)
+	// la_capi_locate_path(ctx, image_path, prompt, mode) -> char* json (0 = err)
+	CapiLocatePath func(handle uintptr, imagePath string, prompt string, mode int32) uintptr
+	// la_capi_locate_buffer(ctx, bytes, len, prompt, mode) -> char* json (0 = err)
+	CapiLocateBuffer func(handle uintptr, bytes uintptr, length uintptr, prompt string, mode int32) uintptr
+	// la_capi_get_n_detections(ctx) -> int
+	CapiGetNDetections func(handle uintptr) int32
+	// la_capi_get_detection_box(ctx, i, out_xyxy[4]) -> int (0 on success)
+	CapiGetDetectionBox func(handle uintptr, i int32, outXYXY uintptr) int32
+	// la_capi_get_detection_label(ctx, i, buf, buf_size) -> int (required size incl NUL; two-call sizing)
+	CapiGetDetectionLabel func(handle uintptr, i int32, buf uintptr, bufSize int32) int32
+	// la_capi_free_string(char* s)
+	CapiFreeString func(s uintptr)
+	// la_capi_last_error(ctx) -> const char* (owned by ctx, "" if none / null ctx).
+	// purego marshals the returned C string into a Go string (a copy), so we
+	// never free it and avoid raw pointer arithmetic.
+	CapiLastError func(handle uintptr) string
+)
+
+type LocateAnythingCpp struct {
+	base.SingleThread
+	handle uintptr
+}
+
+// Load loads the GGUF model at opts.ModelFile (joined with opts.ModelPath if
+// relative) and stores the la_ctx handle for later Detect calls.
+func (r *LocateAnythingCpp) Load(opts *pb.ModelOptions) error {
+	modelFile := opts.ModelFile
+	if modelFile == "" {
+		modelFile = opts.Model
+	}
+	if modelFile == "" {
+		return fmt.Errorf("locate-anything-cpp: ModelFile is empty")
+	}
+
+	var modelPath string
+	if filepath.IsAbs(modelFile) {
+		modelPath = modelFile
+	} else {
+		modelPath = filepath.Join(opts.ModelPath, modelFile)
+	}
+
+	if _, err := os.Stat(modelPath); err != nil {
+		return fmt.Errorf("locate-anything-cpp: model file not found: %s: %w", modelPath, err)
+	}
+
+	threads := opts.Threads
+	if threads <= 0 {
+		threads = 4
+	}
+
+	// Release previous model if any (re-Load).
+	if r.handle != 0 {
+		CapiFree(r.handle)
+		r.handle = 0
+	}
+
+	h := CapiLoad(modelPath, threads)
+	if h == 0 {
+		// la_capi_last_error needs a ctx; on a failed load we have none (it
+		// returns "" for a null ctx), so the text is best-effort. Surface it
+		// when present.
+		if msg := CapiLastError(0); msg != "" {
+			return fmt.Errorf("locate-anything-cpp: la_capi_load failed for %s: %s", modelPath, msg)
+		}
+		return fmt.Errorf("locate-anything-cpp: la_capi_load failed for %s", modelPath)
+	}
+	r.handle = h
+	return nil
+}
+
+// Detect runs open-vocabulary detection on the base64-encoded image in opts.Src
+// using the required text prompt in opts.Prompt, returning one pb.Detection per
+// located object with its predicted label as ClassName.
+func (r *LocateAnythingCpp) Detect(opts *pb.DetectOptions) (pb.DetectResponse, error) {
+	if r.handle == 0 {
+		return pb.DetectResponse{}, fmt.Errorf("locate-anything-cpp: model not loaded")
+	}
+
+	// Open-vocabulary detection is prompt-driven; without a prompt there is
+	// nothing to locate.
+	prompt := opts.Prompt
+	if prompt == "" {
+		return pb.DetectResponse{}, fmt.Errorf("locate-anything-cpp: a text prompt is required (open-vocabulary detection)")
+	}
+
+	// Decode base64 image and write to temp file.
+	imgData, err := base64.StdEncoding.DecodeString(opts.Src)
+	if err != nil {
+		return pb.DetectResponse{}, fmt.Errorf("locate-anything-cpp: failed to decode base64 image: %w", err)
+	}
+
+	tmpFile, err := os.CreateTemp("", "locate-anything-*.img")
+	if err != nil {
+		return pb.DetectResponse{}, fmt.Errorf("locate-anything-cpp: failed to create temp file: %w", err)
+	}
+	defer func() { _ = os.Remove(tmpFile.Name()) }()
+
+	if _, err := tmpFile.Write(imgData); err != nil {
+		_ = tmpFile.Close()
+		return pb.DetectResponse{}, fmt.Errorf("locate-anything-cpp: failed to write temp file: %w", err)
+	}
+	if err := tmpFile.Close(); err != nil {
+		return pb.DetectResponse{}, fmt.Errorf("locate-anything-cpp: failed to close temp file: %w", err)
+	}
+
+	// mode 0 = hybrid (Parallel Box Decoding). The JSON return value is unused:
+	// structured detections are read via the accessor functions. Still must
+	// free the returned string.
+	jsonPtr := CapiLocatePath(r.handle, tmpFile.Name(), prompt, 0)
+	if jsonPtr != 0 {
+		CapiFreeString(jsonPtr)
+	}
+
+	n := CapiGetNDetections(r.handle)
+	if n < 0 {
+		return pb.DetectResponse{}, fmt.Errorf("locate-anything-cpp: invalid n_detections=%d", n)
+	}
+
+	detections := make([]*pb.Detection, 0, n)
+	for i := int32(0); i < n; i++ {
+		var xyxy [4]float32 // x1, y1, x2, y2
+		if CapiGetDetectionBox(r.handle, i, uintptr(unsafe.Pointer(&xyxy[0]))) != 0 {
+			continue
+		}
+
+		// Two-call sizing for the label string.
+		label := ""
+		need := CapiGetDetectionLabel(r.handle, i, 0, 0)
+		if need > 0 {
+			buf := make([]byte, need)
+			CapiGetDetectionLabel(r.handle, i, uintptr(unsafe.Pointer(&buf[0])), need)
+			label = string(buf[:need-1])
+		}
+
+		detections = append(detections, &pb.Detection{
+			X:          xyxy[0],
+			Y:          xyxy[1],
+			Width:      xyxy[2] - xyxy[0],
+			Height:     xyxy[3] - xyxy[1],
+			Confidence: 1.0,
+			ClassName:  label,
+		})
+	}
+
+	return pb.DetectResponse{
+		Detections: detections,
+	}, nil
+}

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

@@ -0,0 +1,59 @@
+package main
+
+// main.go - entry point for the locate-anything-cpp gRPC backend.
+//
+// Dlopens liblocateanythingcpp-<variant>.so via purego at the path in
+// LOCATEANYTHING_LIBRARY (set by run.sh based on /proc/cpuinfo), registers
+// the la_capi_* C ABI symbols, then starts the gRPC server.
+
+import (
+	"flag"
+	"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() {
+	// Get library name from environment variable, default to fallback
+	libName := os.Getenv("LOCATEANYTHING_LIBRARY")
+	if libName == "" {
+		libName = "./liblocateanythingcpp-fallback.so"
+	}
+
+	lib, err := purego.Dlopen(libName, purego.RTLD_NOW|purego.RTLD_GLOBAL)
+	if err != nil {
+		panic(err)
+	}
+
+	libFuncs := []LibFuncs{
+		{&CapiLoad, "la_capi_load"},
+		{&CapiFree, "la_capi_free"},
+		{&CapiLocatePath, "la_capi_locate_path"},
+		{&CapiLocateBuffer, "la_capi_locate_buffer"},
+		{&CapiGetNDetections, "la_capi_get_n_detections"},
+		{&CapiGetDetectionBox, "la_capi_get_detection_box"},
+		{&CapiGetDetectionLabel, "la_capi_get_detection_label"},
+		{&CapiFreeString, "la_capi_free_string"},
+		{&CapiLastError, "la_capi_last_error"},
+	}
+
+	for _, lf := range libFuncs {
+		purego.RegisterLibFunc(lf.FuncPtr, lib, lf.Name)
+	}
+
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &LocateAnythingCpp{}); err != nil {
+		panic(err)
+	}
+}

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

@@ -0,0 +1,176 @@
+package main
+
+// main_test.go - end-to-end smoke test for the locate-anything-cpp gRPC backend.
+//
+// Spawns the compiled locate-anything-cpp binary on a free local port, dials it
+// via gRPC, and exercises LoadModel + Detect against the test fixtures
+// downloaded by test.sh: the q8_0 GGUF of nvidia/LocateAnything-3B and a real
+// COCO image with people + cars. Asserts that open-vocabulary detection driven
+// by a text prompt returns at least one detection, each carrying a non-empty
+// class name and a bounding box of non-zero size.
+//
+// The spec Skip()s cleanly if its fixtures (the ~6.3 GB model, the test image,
+// the built binary, or the fallback .so) are missing, so the test target stays
+// usable on a fresh checkout / on CI runners where the large model hasn't been
+// downloaded.
+
+import (
+	"context"
+	"encoding/base64"
+	"fmt"
+	"net"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"testing"
+	"time"
+
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+	"google.golang.org/grpc"
+	"google.golang.org/grpc/credentials/insecure"
+)
+
+func TestDetect(t *testing.T) {
+	RegisterFailHandler(Fail)
+	RunSpecs(t, "locate-anything-cpp backend smoke suite")
+}
+
+// freePort grabs an ephemeral TCP port and immediately releases it so the
+// spawned backend can bind to it. There is a tiny TOCTOU window here but in
+// practice it's adequate for a smoke test on a quiet runner.
+func freePort() int {
+	l, err := net.Listen("tcp", "127.0.0.1:0")
+	Expect(err).ToNot(HaveOccurred(), "freePort listen")
+	port := l.Addr().(*net.TCPAddr).Port
+	Expect(l.Close()).To(Succeed())
+	return port
+}
+
+// startBackend spawns the locate-anything-cpp binary on the given port and
+// waits until it accepts TCP connections (up to 10s). It mirrors how main.go
+// resolves the purego library: the LOCATEANYTHING_LIBRARY env var points the
+// dlopen at the freshly built fallback .so, and the la_capi_* symbols are
+// registered there. The returned cleanup func kills the process and reaps it.
+func startBackend(port int) func() {
+	binary, err := filepath.Abs("./locate-anything-cpp")
+	Expect(err).ToNot(HaveOccurred())
+	if _, err := os.Stat(binary); err != nil {
+		Skip(fmt.Sprintf("backend binary not built: %s (run `make locate-anything-cpp` first)", binary))
+	}
+
+	libPath, err := filepath.Abs("./liblocateanythingcpp-fallback.so")
+	Expect(err).ToNot(HaveOccurred())
+	if _, err := os.Stat(libPath); err != nil {
+		Skip(fmt.Sprintf("fallback library not built: %s (run `make liblocateanythingcpp-fallback.so` first)", libPath))
+	}
+
+	addr := fmt.Sprintf("127.0.0.1:%d", port)
+	cmd := exec.Command(binary, "--addr", addr)
+	cmd.Env = append(os.Environ(), "LOCATEANYTHING_LIBRARY="+libPath)
+	cmd.Stdout = os.Stderr
+	cmd.Stderr = os.Stderr
+	Expect(cmd.Start()).To(Succeed())
+
+	cleanup := func() {
+		if cmd.Process != nil {
+			_ = cmd.Process.Kill()
+			_, _ = cmd.Process.Wait()
+		}
+	}
+
+	deadline := time.Now().Add(10 * time.Second)
+	for time.Now().Before(deadline) {
+		c, err := net.DialTimeout("tcp", addr, 200*time.Millisecond)
+		if err == nil {
+			_ = c.Close()
+			return cleanup
+		}
+		time.Sleep(200 * time.Millisecond)
+	}
+
+	cleanup()
+	Fail(fmt.Sprintf("backend did not become ready on %s within 10s", addr))
+	return func() {}
+}
+
+// loadTestImage reads the COCO test image downloaded by test.sh and returns its
+// base64-encoded content (the wire format accepted by the Detect RPC).
+func loadTestImage() string {
+	imgPath, err := filepath.Abs("test-data/test.jpg")
+	Expect(err).ToNot(HaveOccurred())
+	imgBytes, err := os.ReadFile(imgPath)
+	if err != nil {
+		Skip(fmt.Sprintf("test image not present: %s (run test.sh first)", imgPath))
+	}
+	return base64.StdEncoding.EncodeToString(imgBytes)
+}
+
+// dialBackend opens a gRPC client connection to the spawned backend.
+func dialBackend(port int) (pb.BackendClient, func()) {
+	addr := fmt.Sprintf("127.0.0.1:%d", port)
+	conn, err := grpc.NewClient(addr, grpc.WithTransportCredentials(insecure.NewCredentials()))
+	Expect(err).ToNot(HaveOccurred())
+	return pb.NewBackendClient(conn), func() { _ = conn.Close() }
+}
+
+// modelPathOrSkip resolves the model file under ./test-models/ and Skip()s the
+// current spec if it's missing (the ~6.3 GB GGUF is not present on a fresh
+// checkout / on CI runners without the download).
+func modelPathOrSkip(name string) string {
+	modelDir, err := filepath.Abs("test-models")
+	Expect(err).ToNot(HaveOccurred())
+	modelPath := filepath.Join(modelDir, name)
+	if _, err := os.Stat(modelPath); err != nil {
+		Skip(fmt.Sprintf("model not present: %s (run test.sh first)", modelPath))
+	}
+	return modelPath
+}
+
+var _ = Describe("locate-anything-cpp backend", func() {
+	It("runs open-vocabulary detection against a known-good COCO image", func() {
+		modelPath := modelPathOrSkip("locate-anything-q8_0.gguf")
+		imgB64 := loadTestImage()
+
+		port := freePort()
+		cleanup := startBackend(port)
+		defer cleanup()
+
+		client, closeConn := dialBackend(port)
+		defer closeConn()
+
+		// The q8_0 model is ~6.3 GB and hybrid Parallel Box Decoding on CPU is
+		// not cheap, so give LoadModel + Detect a generous deadline.
+		ctx, cancel := context.WithTimeout(context.Background(), 20*time.Minute)
+		defer cancel()
+
+		loadResp, err := client.LoadModel(ctx, &pb.ModelOptions{
+			Model:     "locate-anything-q8_0.gguf",
+			ModelFile: modelPath,
+			Threads:   4,
+		})
+		Expect(err).ToNot(HaveOccurred(), "LoadModel")
+		Expect(loadResp.GetSuccess()).To(BeTrue(), "LoadModel reported failure: %s", loadResp.GetMessage())
+
+		// Open-vocabulary detection is prompt-driven; the prompt names the
+		// classes to locate (people + cars), separated by the </c> control token.
+		detResp, err := client.Detect(ctx, &pb.DetectOptions{
+			Src:    imgB64,
+			Prompt: "Locate all the instances that matches the following description: person</c>car.",
+		})
+		Expect(err).ToNot(HaveOccurred(), "Detect")
+		Expect(detResp.GetDetections()).ToNot(BeEmpty(), "no detections returned on a known-good COCO image")
+
+		_, _ = fmt.Fprintf(GinkgoWriter, "detection OK: %d detections\n", len(detResp.GetDetections()))
+		for i, d := range detResp.GetDetections() {
+			Expect(d.GetClassName()).ToNot(BeEmpty(), "detection %d has empty class_name", i)
+			Expect(d.GetWidth()).To(BeNumerically(">", float32(0)),
+				"detection %d has non-positive width", i)
+			Expect(d.GetHeight()).To(BeNumerically(">", float32(0)),
+				"detection %d has non-positive height", i)
+			_, _ = fmt.Fprintf(GinkgoWriter, "  [%d] %s box=(%.1f,%.1f,%.1fx%.1f)\n",
+				i, d.GetClassName(), d.GetX(), d.GetY(), d.GetWidth(), d.GetHeight())
+		}
+	})
+})

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

@@ -0,0 +1,59 @@
+#!/bin/bash
+
+# Script to copy the appropriate libraries based on architecture
+
+set -e
+
+CURDIR=$(dirname "$(realpath $0)")
+REPO_ROOT="${CURDIR}/../../.."
+
+# Create lib directory
+mkdir -p $CURDIR/package/lib
+
+cp -avf $CURDIR/liblocateanythingcpp-*.so $CURDIR/package/
+cp -avf $CURDIR/locate-anything-cpp $CURDIR/package/
+cp -fv $CURDIR/run.sh $CURDIR/package/
+
+# Detect architecture and copy appropriate libraries
+if [ -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
+    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
+    # ARM64 architecture
+    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 [ $(uname -s) = "Darwin" ]; then
+    echo "Detected Darwin"
+else
+    echo "Error: Could not detect architecture"
+    exit 1
+fi
+
+# Package GPU libraries based on BUILD_TYPE
+GPU_LIB_SCRIPT="${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/
+ls -liah $CURDIR/package/lib/

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

@@ -0,0 +1,52 @@
+#!/bin/bash
+set -ex
+
+# Get the absolute current dir where the script is located
+CURDIR=$(dirname "$(realpath $0)")
+
+cd /
+
+echo "CPU info:"
+if [ "$(uname)" != "Darwin" ]; then
+	grep -e "model\sname" /proc/cpuinfo | head -1
+	grep -e "flags" /proc/cpuinfo | head -1
+fi
+
+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 $CURDIR/liblocateanythingcpp-avx.so ]; then
+			LIBRARY="$CURDIR/liblocateanythingcpp-avx.so"
+		fi
+	fi
+
+	if grep -q -e "\savx2\s" /proc/cpuinfo ; then
+		echo "CPU:    AVX2   found OK"
+		if [ -e $CURDIR/liblocateanythingcpp-avx2.so ]; then
+			LIBRARY="$CURDIR/liblocateanythingcpp-avx2.so"
+		fi
+	fi
+
+	# Check avx 512
+	if grep -q -e "\savx512f\s" /proc/cpuinfo ; then
+		echo "CPU:    AVX512F found OK"
+		if [ -e $CURDIR/liblocateanythingcpp-avx512.so ]; then
+			LIBRARY="$CURDIR/liblocateanythingcpp-avx512.so"
+		fi
+	fi
+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
+	echo "Using lib/ld.so"
+	echo "Using library: $LIBRARY"
+	exec $CURDIR/lib/ld.so $CURDIR/locate-anything-cpp "$@"
+fi
+
+echo "Using library: $LIBRARY"
+exec $CURDIR/locate-anything-cpp "$@"

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

@@ -0,0 +1,47 @@
+#!/bin/bash
+set -e
+
+CURDIR=$(dirname "$(realpath $0)")
+
+echo "Running locate-anything-cpp backend tests..."
+
+# Test model from the mudler/locate-anything.cpp-gguf HuggingFace repo. This is
+# the q8_0 quantization of nvidia/LocateAnything-3B (~6.3 GB), so the download
+# is the slow step. It is resumed with `curl -C -` and skipped entirely if the
+# file is already present.
+LOCATEANYTHING_MODEL_DIR="${LOCATEANYTHING_MODEL_DIR:-$CURDIR/test-models}"
+
+LOCATEANYTHING_MODEL_FILE="${LOCATEANYTHING_MODEL_FILE:-locate-anything-q8_0.gguf}"
+LOCATEANYTHING_MODEL_URL="${LOCATEANYTHING_MODEL_URL:-https://huggingface.co/mudler/locate-anything.cpp-gguf/resolve/main/locate-anything-q8_0.gguf}"
+
+mkdir -p "$LOCATEANYTHING_MODEL_DIR"
+
+if [ ! -f "$LOCATEANYTHING_MODEL_DIR/$LOCATEANYTHING_MODEL_FILE" ]; then
+    echo "Downloading locate-anything q8_0 model (~6.3 GB, this is slow)..."
+    # -C - resumes a partial download so an interrupted run doesn't restart from 0.
+    curl -L -C - -o "$LOCATEANYTHING_MODEL_DIR/$LOCATEANYTHING_MODEL_FILE" "$LOCATEANYTHING_MODEL_URL" --progress-bar
+fi
+
+# Use a real COCO test image (people + cars) from the upstream rf-detr.cpp repo
+# (~46 KB). Open-vocabulary detection needs real content to locate, so a
+# synthetic image would trivially yield zero detections.
+TEST_IMAGE_DIR="$CURDIR/test-data"
+TEST_IMAGE_FILE="$TEST_IMAGE_DIR/test.jpg"
+TEST_IMAGE_URL="${TEST_IMAGE_URL:-https://raw.githubusercontent.com/mudler/rf-detr.cpp/main/tests/fixtures/ci/test_image.jpg}"
+
+mkdir -p "$TEST_IMAGE_DIR"
+if [ ! -f "$TEST_IMAGE_FILE" ]; then
+    echo "Downloading COCO test image..."
+    curl -L -o "$TEST_IMAGE_FILE" "$TEST_IMAGE_URL" --progress-bar
+fi
+
+echo "locate-anything-cpp test setup complete."
+echo "  model:      $LOCATEANYTHING_MODEL_DIR/$LOCATEANYTHING_MODEL_FILE"
+echo "  test image: $TEST_IMAGE_FILE"
+
+# Run the Go smoke test: spawns the backend binary on a free port, calls
+# LoadModel + Detect via gRPC against the downloaded GGUF + COCO image.
+echo ""
+echo "Running Go smoke test..."
+cd "$CURDIR"
+go test -v -timeout 30m ./...

backend/go/omnivoice-cpp/.gitignore

@@ -0,0 +1,17 @@
+# Fetched upstream sources
+sources/
+
+# CMake build directories
+build*/
+
+# Compiled shared libraries
+*.so
+
+# Compiled backend binary
+omnivoice-cpp
+
+# Packaging output
+package/
+
+# Downloaded e2e models
+omnivoice-models/

backend/go/omnivoice-cpp/CMakeLists.txt

@@ -0,0 +1,53 @@
+cmake_minimum_required(VERSION 3.14)
+project(gomnivoicecpp LANGUAGES C CXX)
+set(CMAKE_POSITION_INDEPENDENT_CODE ON)
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+
+set(OMNIVOICE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/sources/omnivoice.cpp)
+
+# Override upstream's CMAKE_CUDA_ARCHITECTURES before add_subdirectory.
+if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+    set(CMAKE_CUDA_ARCHITECTURES "75-virtual;80-virtual;86-real;89-real")
+endif()
+
+# Add the upstream project. Its own CMakeLists adds ggml + builds
+# omnivoice-core (STATIC, contains src/omnivoice.cpp i.e. the ov_* impl).
+# EXCLUDE_FROM_ALL keeps its CLI tools/tests from building unless referenced.
+add_subdirectory(${OMNIVOICE_DIR} omnivoice EXCLUDE_FROM_ALL)
+
+# Upstream generates version.h into its own CMAKE_CURRENT_BINARY_DIR and adds
+# the top-level ${CMAKE_BINARY_DIR} to omnivoice-core's include path. When the
+# project is nested under add_subdirectory those two directories differ
+# (<build>/omnivoice vs <build>), so omnivoice.cpp cannot find version.h. Point
+# omnivoice-core at the subproject binary dir where version.h is actually
+# generated. (Fix lives here, never in the fetched upstream checkout.)
+target_include_directories(omnivoice-core PRIVATE ${CMAKE_BINARY_DIR}/omnivoice)
+
+add_library(gomnivoicecpp MODULE cpp/gomnivoicecpp.cpp)
+target_link_libraries(gomnivoicecpp PRIVATE omnivoice-core)
+
+target_include_directories(gomnivoicecpp PRIVATE ${OMNIVOICE_DIR}/src)
+target_include_directories(gomnivoicecpp SYSTEM PRIVATE ${OMNIVOICE_DIR}/ggml/include)
+
+# Link GPU backends if the upstream ggml created them.
+foreach(backend blas cuda metal vulkan sycl)
+    if(TARGET ggml-${backend})
+        target_link_libraries(gomnivoicecpp PRIVATE ggml-${backend})
+        if(backend STREQUAL "cuda")
+            find_package(CUDAToolkit QUIET)
+            if(CUDAToolkit_FOUND)
+                target_link_libraries(gomnivoicecpp PRIVATE CUDA::cudart)
+            endif()
+        endif()
+    endif()
+endforeach()
+
+if(MSVC)
+    target_compile_options(gomnivoicecpp PRIVATE /W4 /wd4100 /wd4505)
+else()
+    target_compile_options(gomnivoicecpp PRIVATE -Wall -Wextra
+                          -Wno-unused-parameter -Wno-unused-function)
+endif()
+
+set_property(TARGET gomnivoicecpp PROPERTY CXX_STANDARD 17)
+set_target_properties(gomnivoicecpp PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR})

backend/go/omnivoice-cpp/Makefile

@@ -0,0 +1,122 @@
+CMAKE_ARGS?=
+BUILD_TYPE?=
+NATIVE?=false
+
+GOCMD?=go
+GO_TAGS?=
+JOBS?=$(shell nproc --ignore=1)
+
+# omnivoice.cpp version
+OMNIVOICE_REPO?=https://github.com/ServeurpersoCom/omnivoice.cpp
+OMNIVOICE_VERSION?=2603355a5dfacae5cfc33531d5d0933221843509
+SO_TARGET?=libgomnivoicecpp.so
+
+CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF
+
+ifeq ($(NATIVE),false)
+	CMAKE_ARGS+=-DGGML_NATIVE=OFF
+endif
+
+ifeq ($(BUILD_TYPE),cublas)
+	CMAKE_ARGS+=-DGGML_CUDA=ON
+else ifeq ($(BUILD_TYPE),openblas)
+	CMAKE_ARGS+=-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
+else ifeq ($(BUILD_TYPE),clblas)
+	CMAKE_ARGS+=-DGGML_CLBLAST=ON -DCLBlast_DIR=/some/path
+else ifeq ($(BUILD_TYPE),hipblas)
+	CMAKE_ARGS+=-DGGML_HIPBLAS=ON
+else ifeq ($(BUILD_TYPE),vulkan)
+	CMAKE_ARGS+=-DGGML_VULKAN=ON
+else ifeq ($(OS),Darwin)
+	ifneq ($(BUILD_TYPE),metal)
+		CMAKE_ARGS+=-DGGML_METAL=OFF
+	else
+		CMAKE_ARGS+=-DGGML_METAL=ON
+		CMAKE_ARGS+=-DGGML_METAL_EMBED_LIBRARY=ON
+	endif
+endif
+
+ifeq ($(BUILD_TYPE),sycl_f16)
+	CMAKE_ARGS+=-DGGML_SYCL=ON \
+		-DCMAKE_C_COMPILER=icx \
+		-DCMAKE_CXX_COMPILER=icpx \
+		-DGGML_SYCL_F16=ON
+endif
+
+ifeq ($(BUILD_TYPE),sycl_f32)
+	CMAKE_ARGS+=-DGGML_SYCL=ON \
+		-DCMAKE_C_COMPILER=icx \
+		-DCMAKE_CXX_COMPILER=icpx
+endif
+
+sources/omnivoice.cpp:
+	mkdir -p sources/omnivoice.cpp
+	cd sources/omnivoice.cpp && \
+	git init && \
+	git remote add origin $(OMNIVOICE_REPO) && \
+	git fetch origin && \
+	git checkout $(OMNIVOICE_VERSION) && \
+	git submodule update --init --recursive --depth 1 --single-branch
+
+# Detect OS
+UNAME_S := $(shell uname -s)
+
+# Only build CPU variants on Linux
+ifeq ($(UNAME_S),Linux)
+	VARIANT_TARGETS = libgomnivoicecpp-avx.so libgomnivoicecpp-avx2.so libgomnivoicecpp-avx512.so libgomnivoicecpp-fallback.so
+else
+	VARIANT_TARGETS = libgomnivoicecpp-fallback.so
+endif
+
+omnivoice-cpp: main.go gomnivoicecpp.go $(VARIANT_TARGETS)
+	CGO_ENABLED=0 $(GOCMD) build -tags "$(GO_TAGS)" -o omnivoice-cpp ./
+
+package: omnivoice-cpp
+	bash package.sh
+
+build: package
+
+clean: purge
+	rm -rf libgomnivoicecpp*.so package sources/omnivoice.cpp omnivoice-cpp
+
+purge:
+	rm -rf build*
+
+.NOTPARALLEL:
+
+ifeq ($(UNAME_S),Linux)
+libgomnivoicecpp-avx.so: sources/omnivoice.cpp
+	$(info ${GREEN}I omnivoice-cpp build info:avx${RESET})
+	SO_TARGET=libgomnivoicecpp-avx.so CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off" $(MAKE) libgomnivoicecpp-custom
+	rm -rf build-libgomnivoicecpp-avx.so
+
+libgomnivoicecpp-avx2.so: sources/omnivoice.cpp
+	$(info ${GREEN}I omnivoice-cpp build info:avx2${RESET})
+	SO_TARGET=libgomnivoicecpp-avx2.so CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on -DGGML_BMI2=on" $(MAKE) libgomnivoicecpp-custom
+	rm -rf build-libgomnivoicecpp-avx2.so
+
+libgomnivoicecpp-avx512.so: sources/omnivoice.cpp
+	$(info ${GREEN}I omnivoice-cpp build info:avx512${RESET})
+	SO_TARGET=libgomnivoicecpp-avx512.so CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=on -DGGML_FMA=on -DGGML_F16C=on -DGGML_BMI2=on" $(MAKE) libgomnivoicecpp-custom
+	rm -rf build-libgomnivoicecpp-avx512.so
+endif
+
+libgomnivoicecpp-fallback.so: sources/omnivoice.cpp
+	$(info ${GREEN}I omnivoice-cpp build info:fallback${RESET})
+	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" $(MAKE) libgomnivoicecpp-custom
+	rm -rf build-libgomnivoicecpp-fallback.so
+
+libgomnivoicecpp-custom: CMakeLists.txt cpp/gomnivoicecpp.cpp cpp/gomnivoicecpp.h
+	mkdir -p build-$(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)
+
+test: omnivoice-cpp
+	@echo "Running omnivoice-cpp tests..."
+	bash test.sh
+	@echo "omnivoice-cpp tests completed."
+
+all: omnivoice-cpp package

backend/go/omnivoice-cpp/audio.go

@@ -0,0 +1,129 @@
+package main
+
+import (
+	"bytes"
+	"encoding/binary"
+	"fmt"
+	"os"
+	"runtime"
+
+	"github.com/go-audio/audio"
+	"github.com/go-audio/wav"
+)
+
+const omnivoiceSampleRate = 24000
+
+// wavHeader24k returns a 44-byte WAV header for a streaming 24 kHz mono 16-bit
+// PCM stream, with placeholder (0xFFFFFFFF) sizes since the total length is
+// unknown up front. Emitted as the first chunk of TTSStream so the HTTP layer
+// receives a self-describing WAV (the gRPC TTSStream path never sets Message,
+// so the backend owns the header - see core/backend/tts.go:ModelTTSStream).
+func wavHeader24k() []byte {
+	var buf bytes.Buffer
+	w := func(v any) { _ = binary.Write(&buf, binary.LittleEndian, v) }
+	buf.WriteString("RIFF")
+	w(uint32(0xFFFFFFFF))
+	buf.WriteString("WAVE")
+	buf.WriteString("fmt ")
+	w(uint32(16))                      // Subchunk1Size
+	w(uint16(1))                       // PCM
+	w(uint16(1))                       // mono
+	w(uint32(omnivoiceSampleRate))     // sample rate
+	w(uint32(omnivoiceSampleRate * 2)) // byte rate = SR * blockAlign
+	w(uint16(2))                       // block align (16-bit mono)
+	w(uint16(16))                      // bits per sample
+	buf.WriteString("data")
+	w(uint32(0xFFFFFFFF))
+	return buf.Bytes()
+}
+
+// floatToPCM16LE clamps each sample to [-1,1] and encodes it as little-endian
+// signed 16-bit PCM.
+func floatToPCM16LE(samples []float32) []byte {
+	out := make([]byte, len(samples)*2)
+	for i, s := range samples {
+		if s > 1 {
+			s = 1
+		} else if s < -1 {
+			s = -1
+		}
+		v := int16(s * 32767)
+		out[i*2] = byte(v)
+		out[i*2+1] = byte(v >> 8)
+	}
+	return out
+}
+
+// writeWAV24k writes samples as a finalized 24 kHz mono 16-bit WAV at dst.
+func writeWAV24k(dst string, samples []float32) error {
+	f, err := os.Create(dst)
+	if err != nil {
+		return fmt.Errorf("omnivoice: create %q: %w", dst, err)
+	}
+	enc := wav.NewEncoder(f, omnivoiceSampleRate, 16, 1, 1)
+	ints := make([]int, len(samples))
+	for i, s := range samples {
+		if s > 1 {
+			s = 1
+		} else if s < -1 {
+			s = -1
+		}
+		ints[i] = int(s * 32767)
+	}
+	b := &audio.IntBuffer{
+		Format:         &audio.Format{NumChannels: 1, SampleRate: omnivoiceSampleRate},
+		Data:           ints,
+		SourceBitDepth: 16,
+	}
+	if err := enc.Write(b); err != nil {
+		_ = enc.Close()
+		_ = f.Close()
+		return fmt.Errorf("omnivoice: encode WAV: %w", err)
+	}
+	if err := enc.Close(); err != nil {
+		_ = f.Close()
+		return fmt.Errorf("omnivoice: finalize WAV: %w", err)
+	}
+	return f.Close()
+}
+
+// readWAVAsFloat decodes a WAV file (any sample rate/channels) to a mono
+// float32 slice in [-1,1] for use as reference audio. OmniVoice expects 24 kHz;
+// callers should supply 24 kHz reference clips.
+func readWAVAsFloat(path string) ([]float32, error) {
+	f, err := os.Open(path)
+	if err != nil {
+		return nil, fmt.Errorf("omnivoice: open ref %q: %w", path, err)
+	}
+	defer func() { _ = f.Close() }()
+
+	dec := wav.NewDecoder(f)
+	buf, err := dec.FullPCMBuffer()
+	if err != nil {
+		return nil, fmt.Errorf("omnivoice: decode ref %q: %w", path, err)
+	}
+	ch := int(buf.Format.NumChannels)
+	if ch < 1 {
+		ch = 1
+	}
+	bitDepth := int(buf.SourceBitDepth)
+	if bitDepth == 0 {
+		bitDepth = 16
+	}
+	scale := float32(int64(1) << uint(bitDepth-1))
+	n := len(buf.Data) / ch
+	out := make([]float32, n)
+	for i := 0; i < n; i++ {
+		// Downmix to mono by averaging channels.
+		var acc int
+		for c := 0; c < ch; c++ {
+			acc += buf.Data[i*ch+c]
+		}
+		out[i] = float32(acc) / float32(ch) / scale
+	}
+	return out, nil
+}
+
+// runtimeKeepAlive prevents the GC from reclaiming the reference-audio slice
+// while its backing pointer is in use across the C call.
+func runtimeKeepAlive(v any) { runtime.KeepAlive(v) }

backend/go/omnivoice-cpp/cpp/gomnivoicecpp.cpp

@@ -0,0 +1,166 @@
+#include "gomnivoicecpp.h"
+#include "ggml-backend.h"
+#include "omnivoice.h"
+
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+
+static ov_context *g_ctx = nullptr;
+
+static void ggml_log_cb(enum ggml_log_level level, const char *log,
+                        void * /*data*/) {
+    if (!log)
+        return;
+    const char *lvl = "?????";
+    switch (level) {
+    case GGML_LOG_LEVEL_DEBUG: lvl = "DEBUG"; break;
+    case GGML_LOG_LEVEL_INFO:  lvl = "INFO";  break;
+    case GGML_LOG_LEVEL_WARN:  lvl = "WARN";  break;
+    case GGML_LOG_LEVEL_ERROR: lvl = "ERROR"; break;
+    default: break;
+    }
+    fprintf(stderr, "[%-5s] %s", lvl, log);
+    fflush(stderr);
+}
+
+int omni_load(const char *model_path, const char *codec_path, int use_fa,
+              int clamp_fp16) {
+    ggml_log_set(ggml_log_cb, nullptr);
+    ggml_backend_load_all();
+
+    if (!model_path || model_path[0] == '\0') {
+        fprintf(stderr, "[omnivoice-cpp] ERROR: model_path is required\n");
+        return 1;
+    }
+    if (!codec_path || codec_path[0] == '\0') {
+        fprintf(stderr, "[omnivoice-cpp] ERROR: codec_path is required\n");
+        return 2;
+    }
+
+    ov_init_params p;
+    ov_init_default_params(&p);
+    p.model_path = model_path;
+    p.codec_path = codec_path;
+    p.use_fa = use_fa != 0;
+    p.clamp_fp16 = clamp_fp16 != 0;
+
+    fprintf(stderr, "[omnivoice-cpp] Loading model=%s codec=%s\n", model_path,
+            codec_path);
+
+    g_ctx = ov_init(&p);
+    if (!g_ctx) {
+        fprintf(stderr, "[omnivoice-cpp] FATAL: ov_init failed: %s\n",
+                ov_last_error());
+        return 3;
+    }
+    fprintf(stderr, "[omnivoice-cpp] Model loaded (%s)\n", ov_version());
+    return 0;
+}
+
+// Fill an ov_tts_params from the flat wrapper arguments.
+static void fill_params(ov_tts_params *tp, const char *text, const char *lang,
+                        const char *instruct, const float *ref_samples,
+                        int ref_n, const char *ref_text, long long seed,
+                        int denoise) {
+    ov_tts_default_params(tp);
+    tp->text = text ? text : "";
+    tp->lang = lang ? lang : "";
+    if (instruct && instruct[0] != '\0')
+        tp->instruct = instruct;
+    if (ref_samples && ref_n > 0) {
+        tp->ref_audio_24k = ref_samples;
+        tp->ref_n_samples = ref_n;
+        if (ref_text && ref_text[0] != '\0')
+            tp->ref_text = ref_text;
+        tp->denoise = denoise != 0;
+    }
+    if (seed >= 0)
+        tp->mg_seed = (uint64_t)seed;
+}
+
+float *omni_tts(const char *text, const char *lang, const char *instruct,
+                const float *ref_samples, int ref_n, const char *ref_text,
+                long long seed, int denoise, int *out_n) {
+    if (out_n)
+        *out_n = 0;
+    if (!g_ctx) {
+        fprintf(stderr, "[omnivoice-cpp] ERROR: model not loaded\n");
+        return nullptr;
+    }
+    if (!text || text[0] == '\0') {
+        fprintf(stderr, "[omnivoice-cpp] ERROR: text is required\n");
+        return nullptr; // omni_tts: out_n already 0
+    }
+    ov_tts_params tp;
+    fill_params(&tp, text, lang, instruct, ref_samples, ref_n, ref_text, seed,
+                denoise);
+
+    ov_audio out = {0};
+    enum ov_status rc = ov_synthesize(g_ctx, &tp, &out);
+    if (rc != OV_STATUS_OK || out.n_samples <= 0 || !out.samples) {
+        fprintf(stderr, "[omnivoice-cpp] ERROR: synthesize failed (rc=%d): %s\n",
+                (int)rc, ov_last_error());
+        ov_audio_free(&out);
+        return nullptr;
+    }
+
+    // Copy into a plain malloc buffer the Go side can free symmetrically via
+    // omni_pcm_free; then release the ov_audio-owned buffer.
+    size_t bytes = (size_t)out.n_samples * sizeof(float);
+    float *buf = (float *)malloc(bytes);
+    if (!buf) {
+        fprintf(stderr, "[omnivoice-cpp] ERROR: malloc(%zu) failed\n", bytes);
+        ov_audio_free(&out);
+        return nullptr;
+    }
+    memcpy(buf, out.samples, bytes);
+    if (out_n)
+        *out_n = out.n_samples;
+    ov_audio_free(&out);
+    return buf;
+}
+
+int omni_tts_stream(const char *text, const char *lang, const char *instruct,
+                    const float *ref_samples, int ref_n, const char *ref_text,
+                    long long seed, int denoise, omni_pcm_chunk_cb cb,
+                    void *user_data) {
+    if (!g_ctx) {
+        fprintf(stderr, "[omnivoice-cpp] ERROR: model not loaded\n");
+        return 1;
+    }
+    if (!cb) {
+        fprintf(stderr, "[omnivoice-cpp] ERROR: stream callback is null\n");
+        return 2;
+    }
+    if (!text || text[0] == '\0') {
+        fprintf(stderr, "[omnivoice-cpp] ERROR: text is required\n");
+        return 4;
+    }
+    ov_tts_params tp;
+    fill_params(&tp, text, lang, instruct, ref_samples, ref_n, ref_text, seed,
+                denoise);
+    // ov_audio_chunk_cb has the identical signature to omni_pcm_chunk_cb
+    // (bool vs int return are ABI-compatible; non-zero == true).
+    tp.on_chunk = (ov_audio_chunk_cb)cb;
+    tp.on_chunk_user_data = user_data;
+
+    ov_audio out = {0}; // stays empty in streaming mode
+    enum ov_status rc = ov_synthesize(g_ctx, &tp, &out);
+    ov_audio_free(&out);
+    if (rc != OV_STATUS_OK && rc != OV_STATUS_CANCELLED) {
+        fprintf(stderr, "[omnivoice-cpp] ERROR: stream synth failed (rc=%d): %s\n",
+                (int)rc, ov_last_error());
+        return 3;
+    }
+    return 0;
+}
+
+void omni_pcm_free(float *p) { free(p); }
+
+void omni_unload(void) {
+    if (g_ctx) {
+        ov_free(g_ctx);
+        g_ctx = nullptr;
+    }
+}

backend/go/omnivoice-cpp/cpp/gomnivoicecpp.h

@@ -0,0 +1,38 @@
+#pragma once
+
+#include <cstdint>
+
+extern "C" {
+
+// Streaming PCM chunk callback. samples is mono float PCM at 24 kHz, valid
+// only for the duration of the call. Return non-zero to continue, 0 to abort.
+typedef int (*omni_pcm_chunk_cb)(const float *samples, int n_samples,
+                                 void *user_data);
+
+// Load the LM (model_path) + codec (codec_path) GGUFs. use_fa / clamp_fp16
+// map to ov_init_params. Returns 0 on success, non-zero on failure.
+int omni_load(const char *model_path, const char *codec_path, int use_fa,
+              int clamp_fp16);
+
+// Synthesize to a malloc'd float PCM buffer (caller frees via omni_pcm_free).
+// ref_samples != null && ref_n > 0 => voice cloning (ref_text optional).
+// instruct != null && non-empty => voice design. seed < 0 keeps the default
+// MaskGIT seed. denoise toggles the <|denoise|> marker (only with a reference).
+// Writes the sample count to *out_n. Returns NULL on failure (out_n set to 0).
+float *omni_tts(const char *text, const char *lang, const char *instruct,
+                const float *ref_samples, int ref_n, const char *ref_text,
+                long long seed, int denoise, int *out_n);
+
+// Streaming synthesis: cb is invoked per PCM chunk as audio is produced.
+// Same reference/design/seed semantics as omni_tts. Returns 0 on success.
+int omni_tts_stream(const char *text, const char *lang, const char *instruct,
+                    const float *ref_samples, int ref_n, const char *ref_text,
+                    long long seed, int denoise, omni_pcm_chunk_cb cb,
+                    void *user_data);
+
+// Free a buffer returned by omni_tts.
+void omni_pcm_free(float *p);
+
+// Release the OmniVoice context.
+void omni_unload(void);
+}

backend/go/omnivoice-cpp/e2e_test.go

@@ -0,0 +1,74 @@
+package main
+
+import (
+	"os"
+	"strings"
+
+	"github.com/ebitengine/purego"
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+func ttsReq(text, voice string, lang *string, dst string) *pb.TTSRequest {
+	return &pb.TTSRequest{Text: text, Voice: voice, Language: lang, Dst: dst}
+}
+
+var _ = Describe("OmniVoice e2e", Label("e2e"), func() {
+	var loaded bool
+
+	BeforeEach(func() {
+		modelPath := os.Getenv("OMNIVOICE_MODEL")
+		codecPath := os.Getenv("OMNIVOICE_CODEC")
+		if modelPath == "" || codecPath == "" {
+			Skip("OMNIVOICE_MODEL / OMNIVOICE_CODEC not set; skipping e2e")
+		}
+		if !loaded {
+			lib := os.Getenv("OMNIVOICE_LIBRARY")
+			if lib == "" {
+				lib = "./libgomnivoicecpp-fallback.so"
+			}
+			h, err := purego.Dlopen(lib, purego.RTLD_NOW|purego.RTLD_GLOBAL)
+			Expect(err).ToNot(HaveOccurred())
+			purego.RegisterLibFunc(&CppLoad, h, "omni_load")
+			purego.RegisterLibFunc(&CppTTS, h, "omni_tts")
+			purego.RegisterLibFunc(&CppTTSStream, h, "omni_tts_stream")
+			purego.RegisterLibFunc(&CppPCMFree, h, "omni_pcm_free")
+			purego.RegisterLibFunc(&CppUnload, h, "omni_unload")
+			Expect(CppLoad(modelPath, codecPath, 0, 0)).To(Equal(0))
+			loaded = true
+		}
+	})
+
+	It("synthesizes a WAV file via TTS", func() {
+		b := &OmnivoiceCpp{opts: loadOptions{seed: 42, denoise: true}}
+		dst := GinkgoT().TempDir() + "/out.wav"
+		lang := "en"
+		err := b.TTS(ttsReq("Hello world.", "", &lang, dst))
+		Expect(err).ToNot(HaveOccurred())
+		fi, err := os.Stat(dst)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(fi.Size()).To(BeNumerically(">", int64(44)))
+	})
+
+	It("streams audio chunks via TTSStream", func() {
+		b := &OmnivoiceCpp{opts: loadOptions{seed: 42, denoise: true}}
+		results := make(chan []byte, 1024)
+		lang := "en"
+		done := make(chan error, 1)
+		go func() { done <- b.TTSStream(ttsReq("Hello there, streaming test.", "", &lang, ""), results) }()
+
+		var chunks int
+		var first []byte
+		for c := range results {
+			if chunks == 0 {
+				first = c
+			}
+			chunks++
+		}
+		Expect(<-done).ToNot(HaveOccurred())
+		Expect(chunks).To(BeNumerically(">=", 2))
+		Expect(string(first[0:4])).To(Equal("RIFF"))
+		Expect(strings.HasPrefix(string(first[8:12]), "WAVE")).To(BeTrue())
+	})
+})

backend/go/omnivoice-cpp/gomnivoicecpp.go

@@ -0,0 +1,246 @@
+package main
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"unsafe"
+
+	"github.com/ebitengine/purego"
+	"github.com/mudler/LocalAI/pkg/grpc/base"
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+)
+
+var (
+	// omni_load(model_path, codec_path, use_fa, clamp_fp16) int
+	CppLoad func(modelPath, codecPath string, useFA, clampFP16 int) int
+	// omni_tts(text, lang, instruct, ref_samples, ref_n, ref_text, seed, denoise, out_n) -> float* (uintptr)
+	CppTTS func(text, lang, instruct string, refSamples unsafe.Pointer, refN int,
+		refText string, seed int64, denoise int, outN unsafe.Pointer) uintptr
+	// omni_tts_stream(text, lang, instruct, ref_samples, ref_n, ref_text, seed, denoise, cb, user) int
+	CppTTSStream func(text, lang, instruct string, refSamples unsafe.Pointer, refN int,
+		refText string, seed int64, denoise int, cb uintptr, user uintptr) int
+	CppPCMFree func(ptr uintptr)
+	CppUnload  func()
+)
+
+type OmnivoiceCpp struct {
+	base.SingleThread
+	opts loadOptions
+	// audioPath is the model-config reference voice (tts.audio_path), used as
+	// the default voice-cloning reference when a request does not set Voice.
+	audioPath string
+}
+
+func (o *OmnivoiceCpp) 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)
+	}
+
+	o.opts = parseOptions(opts.Options)
+
+	// Resolve the codec/tokenizer GGUF: explicit option, else auto-discover a
+	// *tokenizer*.gguf sibling of the base model.
+	codec := o.opts.codecPath
+	if codec != "" && !filepath.IsAbs(codec) {
+		codec = filepath.Join(filepath.Dir(model), codec)
+	}
+	if codec == "" {
+		codec = discoverTokenizer(filepath.Dir(model))
+	}
+	if codec == "" {
+		return fmt.Errorf("omnivoice: no codec/tokenizer GGUF found; set option 'tokenizer:<file>'")
+	}
+	o.opts.codecPath = codec
+
+	// tts.audio_path (ModelOptions.AudioPath) is the config-level voice-cloning
+	// reference: a default reference WAV used when a request omits Voice.
+	// Resolved relative to the model directory like the codec.
+	o.audioPath = opts.AudioPath
+	if o.audioPath != "" && !filepath.IsAbs(o.audioPath) {
+		o.audioPath = filepath.Join(filepath.Dir(model), o.audioPath)
+	}
+
+	useFA := boolToInt(o.opts.useFA)
+	clamp := boolToInt(o.opts.clampFP16)
+
+	fmt.Fprintf(os.Stderr, "[omnivoice-cpp] Load model=%s codec=%s use_fa=%d clamp_fp16=%d\n",
+		model, codec, useFA, clamp)
+
+	if rc := CppLoad(model, codec, useFA, clamp); rc != 0 {
+		return fmt.Errorf("omnivoice: failed to load model (rc=%d)", rc)
+	}
+	return nil
+}
+
+// discoverTokenizer returns the first *tokenizer*.gguf in dir, or "".
+func discoverTokenizer(dir string) string {
+	entries, err := os.ReadDir(dir)
+	if err != nil {
+		return ""
+	}
+	for _, e := range entries {
+		name := strings.ToLower(e.Name())
+		if strings.Contains(name, "tokenizer") && strings.HasSuffix(name, ".gguf") {
+			return filepath.Join(dir, e.Name())
+		}
+	}
+	return ""
+}
+
+func boolToInt(b bool) int {
+	if b {
+		return 1
+	}
+	return 0
+}
+
+// refAudio loads the reference WAV (voice cloning) if voice points to a file.
+// Returns nil if no cloning (empty or non-path - voice design uses Instructions).
+func (o *OmnivoiceCpp) refAudio(voice string) ([]float32, error) {
+	v := strings.TrimSpace(voice)
+	if v == "" {
+		return nil, nil
+	}
+	if _, err := os.Stat(v); err != nil {
+		return nil, nil
+	}
+	return readWAVAsFloat(v)
+}
+
+// refAudioFor resolves the cloning reference for a request: the per-request
+// Voice takes precedence, falling back to the model-config audio_path. Empty
+// result means no cloning (voice design via Instructions still applies).
+func (o *OmnivoiceCpp) refAudioFor(req *pb.TTSRequest) ([]float32, error) {
+	voice := strings.TrimSpace(req.Voice)
+	if voice == "" {
+		voice = o.audioPath
+	}
+	return o.refAudio(voice)
+}
+
+func reqParam(req *pb.TTSRequest, key string) string {
+	if req.Params == nil {
+		return ""
+	}
+	return req.Params[key]
+}
+
+func (o *OmnivoiceCpp) seedFor(req *pb.TTSRequest) int64 {
+	if s := reqParam(req, "seed"); s != "" {
+		var n int64
+		if _, err := fmt.Sscan(s, &n); err == nil {
+			return n
+		}
+	}
+	return o.opts.seed
+}
+
+func optStr(p *string) string {
+	if p == nil {
+		return ""
+	}
+	return *p
+}
+
+func (o *OmnivoiceCpp) TTS(req *pb.TTSRequest) error {
+	if req.Dst == "" {
+		return fmt.Errorf("omnivoice: TTS requires a destination path")
+	}
+	lang := normalizeLanguage(optStr(req.Language))
+	instruct := optStr(req.Instructions)
+	refText := reqParam(req, "ref_text")
+	seed := o.seedFor(req)
+
+	ref, err := o.refAudioFor(req)
+	if err != nil {
+		return err
+	}
+	var refPtr unsafe.Pointer
+	if len(ref) > 0 {
+		refPtr = unsafe.Pointer(&ref[0])
+	}
+
+	var n int32
+	ptr := CppTTS(req.Text, lang, instruct, refPtr, len(ref), refText, seed,
+		boolToInt(o.opts.denoise), unsafe.Pointer(&n))
+	runtimeKeepAlive(ref)
+	if ptr == 0 || n <= 0 {
+		return fmt.Errorf("omnivoice: synthesis failed")
+	}
+	defer CppPCMFree(ptr)
+	src := unsafe.Slice((*float32)(unsafe.Pointer(ptr)), int(n)) //nolint:govet // C-allocated PCM, copied out before free
+	out := make([]float32, int(n))
+	copy(out, src)
+	return writeWAV24k(req.Dst, out)
+}
+
+// streamState carries the active TTSStream channel to the single shared C
+// callback. base.SingleThread serializes TTS/TTSStream, so one global slot is
+// safe and avoids leaking a purego callback per request (purego callbacks
+// cannot be freed and are capped).
+var (
+	streamMu     sync.Mutex
+	streamChan   chan []byte
+	streamCbOnce sync.Once
+	streamCbPtr  uintptr
+)
+
+// streamCallback is registered once and forwards each PCM chunk to streamChan.
+func streamCallback(samples *float32, nSamples int32, _ uintptr) uintptr {
+	if nSamples <= 0 || samples == nil || streamChan == nil {
+		return 1 // continue
+	}
+	src := unsafe.Slice(samples, int(nSamples))
+	cp := make([]float32, int(nSamples)) // copy out of C memory before returning
+	copy(cp, src)
+	streamChan <- floatToPCM16LE(cp)
+	return 1 // continue
+}
+
+func (o *OmnivoiceCpp) TTSStream(req *pb.TTSRequest, results chan []byte) error {
+	defer close(results)
+	if req.Text == "" {
+		return fmt.Errorf("omnivoice: TTSStream requires text")
+	}
+
+	streamCbOnce.Do(func() {
+		streamCbPtr = purego.NewCallback(streamCallback)
+	})
+
+	lang := normalizeLanguage(optStr(req.Language))
+	instruct := optStr(req.Instructions)
+	refText := reqParam(req, "ref_text")
+	seed := o.seedFor(req)
+
+	ref, err := o.refAudioFor(req)
+	if err != nil {
+		return err
+	}
+	var refPtr unsafe.Pointer
+	if len(ref) > 0 {
+		refPtr = unsafe.Pointer(&ref[0])
+	}
+
+	// Emit the WAV header first so the HTTP layer gets a self-describing stream.
+	results <- wavHeader24k()
+
+	streamMu.Lock()
+	streamChan = results
+	rc := CppTTSStream(req.Text, lang, instruct, refPtr, len(ref), refText, seed,
+		boolToInt(o.opts.denoise), streamCbPtr, 0)
+	streamChan = nil
+	streamMu.Unlock()
+	runtimeKeepAlive(ref)
+
+	if rc != 0 {
+		return fmt.Errorf("omnivoice: streaming synthesis failed (rc=%d)", rc)
+	}
+	return nil
+}

backend/go/omnivoice-cpp/gomnivoicecpp_test.go

@@ -0,0 +1,90 @@
+package main
+
+import (
+	"bytes"
+	"encoding/binary"
+	"testing"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+func TestOmnivoiceCpp(t *testing.T) {
+	RegisterFailHandler(Fail)
+	RunSpecs(t, "omnivoice-cpp suite")
+}
+
+var _ = Describe("normalizeLanguage", func() {
+	DescribeTable("maps caller language to OmniVoice codes",
+		func(in, want string) {
+			Expect(normalizeLanguage(in)).To(Equal(want))
+		},
+		Entry("empty stays empty", "", ""),
+		Entry("english full name", "English", "en"),
+		Entry("chinese full name", "Chinese", "zh"),
+		Entry("locale suffix stripped", "en-US", "en"),
+		Entry("underscore locale", "zh_CN", "zh"),
+		Entry("already a code", "en", "en"),
+		Entry("unknown passes through normalized", "xx", "xx"),
+	)
+})
+
+var _ = Describe("parseOptions", func() {
+	It("extracts codec, use_fa, clamp_fp16, seed, denoise", func() {
+		o := parseOptions([]string{
+			"tokenizer:tok.gguf",
+			"use_fa:true",
+			"clamp_fp16:true",
+			"seed:7",
+			"denoise:false",
+			"unknown:ignored",
+		})
+		Expect(o.codecPath).To(Equal("tok.gguf"))
+		Expect(o.useFA).To(BeTrue())
+		Expect(o.clampFP16).To(BeTrue())
+		Expect(o.seed).To(Equal(int64(7)))
+		Expect(o.denoise).To(BeFalse())
+	})
+
+	It("accepts codec: as an alias for tokenizer:", func() {
+		o := parseOptions([]string{"codec:c.gguf"})
+		Expect(o.codecPath).To(Equal("c.gguf"))
+	})
+
+	It("defaults seed to -1 and denoise to true", func() {
+		o := parseOptions(nil)
+		Expect(o.seed).To(Equal(int64(-1)))
+		Expect(o.denoise).To(BeTrue())
+	})
+})
+
+var _ = Describe("wavHeader24k", func() {
+	It("emits a 44-byte streaming WAV header at 24 kHz mono 16-bit", func() {
+		h := wavHeader24k()
+		Expect(h).To(HaveLen(44))
+		Expect(string(h[0:4])).To(Equal("RIFF"))
+		Expect(string(h[8:12])).To(Equal("WAVE"))
+		Expect(string(h[12:16])).To(Equal("fmt "))
+		Expect(string(h[36:40])).To(Equal("data"))
+		var sampleRate uint32
+		Expect(binary.Read(bytes.NewReader(h[24:28]), binary.LittleEndian, &sampleRate)).To(Succeed())
+		Expect(sampleRate).To(Equal(uint32(24000)))
+	})
+})
+
+var _ = Describe("floatToPCM16LE", func() {
+	It("clamps and converts float PCM to little-endian int16 bytes", func() {
+		b := floatToPCM16LE([]float32{0, 1.0, -1.0, 2.0, -2.0})
+		Expect(b).To(HaveLen(10)) // 5 samples * 2 bytes
+		read := func(off int) int16 {
+			var v int16
+			_ = binary.Read(bytes.NewReader(b[off:off+2]), binary.LittleEndian, &v)
+			return v
+		}
+		Expect(read(0)).To(Equal(int16(0)))
+		Expect(read(2)).To(Equal(int16(32767)))
+		Expect(read(4)).To(Equal(int16(-32767)))
+		Expect(read(6)).To(Equal(int16(32767)))  // clamped from 2.0
+		Expect(read(8)).To(Equal(int16(-32767))) // clamped from -2.0
+	})
+})

backend/go/omnivoice-cpp/main.go

@@ -0,0 +1,48 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+import (
+	"flag"
+	"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("OMNIVOICE_LIBRARY")
+	if libName == "" {
+		libName = "./libgomnivoicecpp-fallback.so"
+	}
+
+	lib, err := purego.Dlopen(libName, purego.RTLD_NOW|purego.RTLD_GLOBAL)
+	if err != nil {
+		panic(err)
+	}
+
+	libFuncs := []LibFuncs{
+		{&CppLoad, "omni_load"},
+		{&CppTTS, "omni_tts"},
+		{&CppTTSStream, "omni_tts_stream"},
+		{&CppPCMFree, "omni_pcm_free"},
+		{&CppUnload, "omni_unload"},
+	}
+	for _, lf := range libFuncs {
+		purego.RegisterLibFunc(lf.FuncPtr, lib, lf.Name)
+	}
+
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &OmnivoiceCpp{}); err != nil {
+		panic(err)
+	}
+}

backend/go/omnivoice-cpp/options.go

@@ -0,0 +1,74 @@
+package main
+
+import (
+	"strconv"
+	"strings"
+)
+
+// loadOptions holds the parsed model-level options for OmniVoice.
+type loadOptions struct {
+	codecPath string
+	useFA     bool
+	clampFP16 bool
+	seed      int64
+	denoise   bool
+}
+
+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: seed -1 (engine default), denoise true.
+func parseOptions(opts []string) loadOptions {
+	o := loadOptions{seed: -1, denoise: true}
+	for _, oo := range opts {
+		key, value, ok := splitOption(oo)
+		if !ok {
+			continue
+		}
+		switch key {
+		case "tokenizer", "codec":
+			o.codecPath = value
+		case "use_fa":
+			o.useFA = value == "true" || value == "1"
+		case "clamp_fp16":
+			o.clampFP16 = value == "true" || value == "1"
+		case "seed":
+			if n, err := strconv.ParseInt(value, 10, 64); err == nil {
+				o.seed = n
+			}
+		case "denoise":
+			o.denoise = value == "true" || value == "1"
+		}
+	}
+	return o
+}
+
+// languageNameAliases maps full language names to OmniVoice codes. OmniVoice's
+// lang hint accepts "" (auto), "en", "zh" per the upstream convention; other
+// codes pass through and the engine treats unknown hints as auto.
+var languageNameAliases = map[string]string{
+	"english": "en",
+	"chinese": "zh",
+}
+
+// normalizeLanguage lowercases, trims, strips a region/locale suffix, and
+// resolves common full names. Empty stays empty so the engine auto-detects.
+func normalizeLanguage(lang string) string {
+	lang = strings.ToLower(strings.TrimSpace(lang))
+	if lang == "" {
+		return ""
+	}
+	if i := strings.IndexAny(lang, "-_."); i >= 0 {
+		lang = lang[:i]
+	}
+	if code, ok := languageNameAliases[lang]; ok {
+		return code
+	}
+	return lang
+}

backend/go/omnivoice-cpp/package.sh

@@ -0,0 +1,64 @@
+#!/bin/bash
+
+# Script to copy the appropriate libraries based on architecture
+# This script is used in the final stage of the Dockerfile
+
+set -e
+
+CURDIR=$(dirname "$(realpath $0)")
+REPO_ROOT="${CURDIR}/../../.."
+
+# Create lib directory
+mkdir -p $CURDIR/package/lib
+
+cp -avf $CURDIR/omnivoice-cpp $CURDIR/package/
+cp -fv $CURDIR/libgomnivoicecpp-*.so $CURDIR/package/
+cp -fv $CURDIR/run.sh $CURDIR/package/
+
+# Detect architecture and copy appropriate libraries
+if [ -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
+    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/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/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
+    # ARM64 architecture
+    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/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/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 [ $(uname -s) = "Darwin" ]; then
+    echo "Detected Darwin"
+else
+    echo "Error: Could not detect architecture"
+    exit 1
+fi
+
+# Package GPU libraries based on BUILD_TYPE
+GPU_LIB_SCRIPT="${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/
+ls -liah $CURDIR/package/lib/

backend/go/omnivoice-cpp/run.sh

@@ -0,0 +1,52 @@
+#!/bin/bash
+set -ex
+
+# Get the absolute current dir where the script is located
+CURDIR=$(dirname "$(realpath $0)")
+
+cd /
+
+echo "CPU info:"
+if [ "$(uname)" != "Darwin" ]; then
+	grep -e "model\sname" /proc/cpuinfo | head -1
+	grep -e "flags" /proc/cpuinfo | head -1
+fi
+
+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 $CURDIR/libgomnivoicecpp-avx.so ]; then
+			LIBRARY="$CURDIR/libgomnivoicecpp-avx.so"
+		fi
+	fi
+
+	if grep -q -e "\savx2\s" /proc/cpuinfo ; then
+		echo "CPU:    AVX2   found OK"
+		if [ -e $CURDIR/libgomnivoicecpp-avx2.so ]; then
+			LIBRARY="$CURDIR/libgomnivoicecpp-avx2.so"
+		fi
+	fi
+
+	# Check avx 512
+	if grep -q -e "\savx512f\s" /proc/cpuinfo ; then
+		echo "CPU:    AVX512F found OK"
+		if [ -e $CURDIR/libgomnivoicecpp-avx512.so ]; then
+			LIBRARY="$CURDIR/libgomnivoicecpp-avx512.so"
+		fi
+	fi
+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
+	echo "Using lib/ld.so"
+	echo "Using library: $LIBRARY"
+	exec $CURDIR/lib/ld.so $CURDIR/omnivoice-cpp "$@"
+fi
+
+echo "Using library: $LIBRARY"
+exec $CURDIR/omnivoice-cpp "$@"

backend/go/omnivoice-cpp/test.sh

@@ -0,0 +1,30 @@
+#!/bin/bash
+set -e
+
+CURDIR=$(dirname "$(realpath $0)")
+cd "$CURDIR"
+
+echo "Running omnivoice-cpp backend tests..."
+
+if [ -z "$OMNIVOICE_MODEL" ]; then
+    MODEL_DIR="./omnivoice-models"
+    mkdir -p "$MODEL_DIR"
+    REPO_ID="Serveurperso/OmniVoice-GGUF"
+    BASE_URL="https://huggingface.co/${REPO_ID}/resolve/main"
+    FILES=( "omnivoice-base-Q4_K_M.gguf" "omnivoice-tokenizer-Q4_K_M.gguf" )
+    for file in "${FILES[@]}"; do
+        dest="${MODEL_DIR}/${file}"
+        if [ -f "${dest}" ]; then
+            echo "  [skip] ${file}"
+        else
+            echo "  [download] ${file}..."
+            curl -L -o "${dest}" "${BASE_URL}/${file}" --progress-bar
+        fi
+    done
+    export OMNIVOICE_MODEL="${MODEL_DIR}/omnivoice-base-Q4_K_M.gguf"
+    export OMNIVOICE_CODEC="${MODEL_DIR}/omnivoice-tokenizer-Q4_K_M.gguf"
+fi
+
+go test -v -timeout 1200s .
+
+echo "All omnivoice-cpp e2e tests passed."

backend/go/parakeet-cpp/Makefile

@@ -1,6 +1,6 @@
 # parakeet-cpp backend Makefile.
 #
-# Upstream pin lives below as PARAKEET_VERSION?=e270af73b94c9a5c37ec516230219ed4580e1db6
+# Upstream pin lives below as PARAKEET_VERSION?=92a5f0306be354c109150fe58ae4cc4f8a21ca45
 # (.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?=e270af73b94c9a5c37ec516230219ed4580e1db6
+PARAKEET_VERSION?=92a5f0306be354c109150fe58ae4cc4f8a21ca45
 PARAKEET_REPO?=https://github.com/mudler/parakeet.cpp
 
 GOCMD?=go
@@ -39,7 +39,10 @@ endif
 # is overwritten back to OFF and the build silently falls back to CPU. Forward the
 # PARAKEET_GGML_* options instead. (openblas is not gated, so -DGGML_BLAS passes through.)
 ifeq ($(BUILD_TYPE),cublas)
-	CMAKE_ARGS+=-DPARAKEET_GGML_CUDA=ON
+	# GGML_CUDA_GRAPHS is OFF by ggml default; enabling it gives a small free
+	# speedup (~1% measured on GB10, never negative) by capturing/replaying the
+	# CUDA graph. Not gated by parakeet.cpp, so it passes straight through to ggml.
+	CMAKE_ARGS+=-DPARAKEET_GGML_CUDA=ON -DGGML_CUDA_GRAPHS=ON
 else ifeq ($(BUILD_TYPE),openblas)
 	CMAKE_ARGS+=-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
 else ifeq ($(BUILD_TYPE),hipblas)

backend/go/parakeet-cpp/goparakeetcpp.go

@@ -98,17 +98,21 @@ type transcriptJSON struct {
 }
 
 // streamFeedJSON mirrors the document returned by
-// parakeet_capi_stream_feed_json / parakeet_capi_stream_finalize_json (ABI v4):
+// parakeet_capi_stream_feed_json / parakeet_capi_stream_finalize_json (ABI v5):
 //
-//	{"text":"...","eou":0,"frame_sec":0.080000,
+//	{"text":"...","eou":0,"eob":0,"frame_sec":0.080000,
 //	 "words":[{"w":"...","start":0.480,"end":0.640,"conf":0.9100}, ...]}
 //
 // "text" is the newly-finalized text since the last call; "eou" is 1 when an
-// <EOU>/<EOB> fired this feed; "words" are the words finalized this call with
-// absolute (stream-relative) start/end seconds.
+// <EOU> (end of utterance) fired this feed and "eob" is 1 when an <EOB>
+// (backchannel) fired. ABI v4 conflated the two into "eou"; v5 split them, so
+// we read both and treat either as an utterance boundary for segmentation.
+// "words" are the words finalized this call with absolute (stream-relative)
+// start/end seconds.
 type streamFeedJSON struct {
 	Text     string           `json:"text"`
 	Eou      int              `json:"eou"`
+	Eob      int              `json:"eob"`
 	FrameSec float64          `json:"frame_sec"`
 	Words    []transcriptWord `json:"words"`
 }
@@ -483,7 +487,10 @@ type streamSegmenter struct {
 
 func (s *streamSegmenter) add(doc streamFeedJSON) {
 	s.cur = append(s.cur, doc.Words...)
-	if doc.Eou != 0 {
+	// Close the segment on either turn signal: <EOU> (end of utterance) or
+	// <EOB> (backchannel). ABI v4 reported both via "eou"; v5 split them, so we
+	// OR them here to keep the v4 segmentation boundaries.
+	if doc.Eou != 0 || doc.Eob != 0 {
 		s.flush()
 	}
 }
@@ -671,11 +678,12 @@ func (p *ParakeetCpp) AudioTranscriptionStream(ctx context.Context, opts *pb.Tra
 	return nil
 }
 
-// streamJSON drives the ABI v4 streaming JSON entry points: each feed/finalize
-// returns a {text,eou,frame_sec,words} document. The newly-finalized text is
-// emitted as a delta (unchanged streaming contract) while words are accumulated
-// into per-utterance segments (closed on EOU) so the closing FinalResult carries
-// timestamped segments. Runs under engineMu (already held by the caller).
+// streamJSON drives the streaming JSON entry points (present since ABI v4): each
+// feed/finalize returns a {text,eou,eob,frame_sec,words} document. The
+// newly-finalized text is emitted as a delta (unchanged streaming contract)
+// while words are accumulated into per-utterance segments (closed on <EOU> or
+// <EOB>) so the closing FinalResult carries timestamped segments. Runs under
+// engineMu (already held by the caller).
 func (p *ParakeetCpp) streamJSON(ctx context.Context, stream uintptr, data []float32,
 	duration float32, results chan *pb.TranscriptStreamResponse) error {
 	var (

backend/go/parakeet-cpp/segments_test.go

@@ -124,4 +124,17 @@ var _ = Describe("streaming segment assembly", func() {
 		Expect(acc.segments()).To(HaveLen(1))
 		Expect(acc.segments()[0].Text).To(Equal("hi there"))
 	})
+
+	// ABI v5 split <EOB> (backchannel) out of the "eou" flag into its own "eob"
+	// field; a backchannel must still close the segment as it did in v4.
+	It("closes a segment on EOB (backchannel) too", func() {
+		acc := &streamSegmenter{}
+		acc.add(streamFeedJSON{Text: "uh huh", Eou: 0, Eob: 1, Words: []transcriptWord{
+			{W: "uh", Start: 0.0, End: 0.2}, {W: "huh", Start: 0.2, End: 0.5},
+		}})
+		segs := acc.segments()
+		Expect(segs).To(HaveLen(1))
+		Expect(segs[0].Text).To(Equal("uh huh"))
+		Expect(segs[0].End).To(Equal(secondsToNanos(0.5)))
+	})
 })

backend/go/qwen3-tts-cpp/CMakeLists.txt

@@ -3,35 +3,36 @@ project(goqwen3ttscpp LANGUAGES C CXX)
 set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
-set(QWEN3TTS_DIR ${CMAKE_CURRENT_SOURCE_DIR}/sources/qwen3-tts.cpp)
+set(QWENTTS_DIR ${CMAKE_CURRENT_SOURCE_DIR}/sources/qwentts.cpp)
 
 # Override upstream's CMAKE_CUDA_ARCHITECTURES before add_subdirectory.
 if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
     set(CMAKE_CUDA_ARCHITECTURES "75-virtual;80-virtual;86-real;89-real")
 endif()
 
-# Build ggml from the upstream's submodule FIRST, so that ggml/ggml-base/ggml-cpu
-# CMake targets exist when the upstream project references them by name.
-# The upstream CMakeLists.txt uses target_link_libraries(... ggml ggml-base ggml-cpu)
-# with target_link_directories pointing at a pre-built ggml/build/. By adding ggml
-# as a subdirectory here, CMake resolves those names as targets instead.
-add_subdirectory(${QWEN3TTS_DIR}/ggml ggml EXCLUDE_FROM_ALL)
+# Add the upstream project. Its own CMakeLists adds ggml + cpp-httplib + yyjson
+# and builds qwen-core (STATIC, the qt_* impl). EXCLUDE_FROM_ALL keeps its CLI
+# tools / tts-server / tests from building unless referenced.
+add_subdirectory(${QWENTTS_DIR} qwentts EXCLUDE_FROM_ALL)
 
-# Now add the upstream project
-add_subdirectory(${QWEN3TTS_DIR} qwen3tts EXCLUDE_FROM_ALL)
+# Upstream generates version.h into its own CMAKE_CURRENT_BINARY_DIR and adds
+# the top-level ${CMAKE_BINARY_DIR} to qwen-core's include path. Under
+# add_subdirectory those two dirs differ (<build>/qwentts vs <build>), so
+# qwen.cpp cannot find version.h. Point qwen-core at the subproject binary dir
+# where version.h is actually generated. (Fix lives here, never in the fetched
+# upstream checkout.)
+target_include_directories(qwen-core PRIVATE ${CMAKE_BINARY_DIR}/qwentts)
 
 add_library(goqwen3ttscpp MODULE cpp/goqwen3ttscpp.cpp)
-target_link_libraries(goqwen3ttscpp PRIVATE qwen3_tts)
+target_link_libraries(goqwen3ttscpp PRIVATE qwen-core)
 
-target_include_directories(goqwen3ttscpp PRIVATE ${QWEN3TTS_DIR}/src)
-target_include_directories(goqwen3ttscpp SYSTEM PRIVATE ${QWEN3TTS_DIR}/ggml/include)
+target_include_directories(goqwen3ttscpp PRIVATE ${QWENTTS_DIR}/src)
+target_include_directories(goqwen3ttscpp SYSTEM PRIVATE ${QWENTTS_DIR}/ggml/include)
 
-# Link GPU backends if available
-foreach(backend blas cuda metal vulkan)
+# Link GPU backends if the upstream ggml created them.
+foreach(backend blas cuda metal vulkan sycl)
     if(TARGET ggml-${backend})
         target_link_libraries(goqwen3ttscpp PRIVATE ggml-${backend})
-        string(TOUPPER ${backend} BACKEND_UPPER)
-        target_compile_definitions(goqwen3ttscpp PRIVATE QWEN3TTS_HAVE_${BACKEND_UPPER})
         if(backend STREQUAL "cuda")
             find_package(CUDAToolkit QUIET)
             if(CUDAToolkit_FOUND)
@@ -44,12 +45,8 @@ endforeach()
 if(MSVC)
     target_compile_options(goqwen3ttscpp PRIVATE /W4 /wd4100 /wd4505)
 else()
-    target_compile_options(goqwen3ttscpp PRIVATE -Wall -Wextra -Wshadow -Wconversion
-                          -Wno-unused-parameter -Wno-unused-function -Wno-sign-conversion)
-endif()
-
-if(CMAKE_CXX_COMPILER_ID MATCHES "GNU" AND CMAKE_CXX_COMPILER_VERSION VERSION_LESS 9.0)
-    target_link_libraries(goqwen3ttscpp PRIVATE stdc++fs)
+    target_compile_options(goqwen3ttscpp PRIVATE -Wall -Wextra
+                          -Wno-unused-parameter -Wno-unused-function)
 endif()
 
 set_property(TARGET goqwen3ttscpp PROPERTY CXX_STANDARD 17)

backend/go/qwen3-tts-cpp/Makefile

@@ -6,9 +6,9 @@ GOCMD?=go
 GO_TAGS?=
 JOBS?=$(shell nproc --ignore=1)
 
-# qwen3-tts.cpp version
-QWEN3TTS_REPO?=https://github.com/predict-woo/qwen3-tts.cpp
-QWEN3TTS_CPP_VERSION?=136e5d36c17083da0321fd96512dc7b263f94a44
+# qwentts.cpp version
+QWEN3TTS_REPO?=https://github.com/ServeurpersoCom/qwentts.cpp
+QWEN3TTS_CPP_VERSION?=0bf4a18b22e8bb8718d95294e9f7f45c0d4270a4
 SO_TARGET?=libgoqwen3ttscpp.so
 
 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF
@@ -49,9 +49,9 @@ ifeq ($(BUILD_TYPE),sycl_f32)
 		-DCMAKE_CXX_COMPILER=icpx
 endif
 
-sources/qwen3-tts.cpp:
-	mkdir -p sources/qwen3-tts.cpp
-	cd sources/qwen3-tts.cpp && \
+sources/qwentts.cpp:
+	mkdir -p sources/qwentts.cpp
+	cd sources/qwentts.cpp && \
 	git init && \
 	git remote add origin $(QWEN3TTS_REPO) && \
 	git fetch origin && \
@@ -78,7 +78,7 @@ package: qwen3-tts-cpp
 build: package
 
 clean: purge
-	rm -rf libgoqwen3ttscpp*.so package sources/qwen3-tts.cpp qwen3-tts-cpp
+	rm -rf libgoqwen3ttscpp*.so package sources/qwentts.cpp qwen3-tts-cpp
 
 purge:
 	rm -rf build*
@@ -88,24 +88,24 @@ purge:
 
 # Build all variants (Linux only)
 ifeq ($(UNAME_S),Linux)
-libgoqwen3ttscpp-avx.so: sources/qwen3-tts.cpp
+libgoqwen3ttscpp-avx.so: sources/qwentts.cpp
 	$(info ${GREEN}I qwen3-tts-cpp build info:avx${RESET})
 	SO_TARGET=libgoqwen3ttscpp-avx.so CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_BMI2=off" $(MAKE) libgoqwen3ttscpp-custom
 	rm -rf build-libgoqwen3ttscpp-avx.so
 
-libgoqwen3ttscpp-avx2.so: sources/qwen3-tts.cpp
+libgoqwen3ttscpp-avx2.so: sources/qwentts.cpp
 	$(info ${GREEN}I qwen3-tts-cpp build info:avx2${RESET})
 	SO_TARGET=libgoqwen3ttscpp-avx2.so CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on -DGGML_BMI2=on" $(MAKE) libgoqwen3ttscpp-custom
 	rm -rf build-libgoqwen3ttscpp-avx2.so
 
-libgoqwen3ttscpp-avx512.so: sources/qwen3-tts.cpp
+libgoqwen3ttscpp-avx512.so: sources/qwentts.cpp
 	$(info ${GREEN}I qwen3-tts-cpp build info:avx512${RESET})
 	SO_TARGET=libgoqwen3ttscpp-avx512.so CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=on -DGGML_FMA=on -DGGML_F16C=on -DGGML_BMI2=on" $(MAKE) libgoqwen3ttscpp-custom
 	rm -rf build-libgoqwen3ttscpp-avx512.so
 endif
 
 # Build fallback variant (all platforms)
-libgoqwen3ttscpp-fallback.so: sources/qwen3-tts.cpp
+libgoqwen3ttscpp-fallback.so: sources/qwentts.cpp
 	$(info ${GREEN}I qwen3-tts-cpp build info:fallback${RESET})
 	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

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

@@ -0,0 +1,128 @@
+package main
+
+import (
+	"bytes"
+	"encoding/binary"
+	"fmt"
+	"os"
+	"runtime"
+
+	"github.com/go-audio/audio"
+	"github.com/go-audio/wav"
+)
+
+const qwen3ttsSampleRate = 24000
+
+// wavHeader24k returns a 44-byte WAV header for a streaming 24 kHz mono 16-bit
+// PCM stream, with placeholder (0xFFFFFFFF) sizes since the total length is
+// unknown up front. Emitted as the first chunk of TTSStream so the HTTP layer
+// receives a self-describing WAV (the gRPC TTSStream path never sets Message,
+// so the backend owns the header - see core/backend/tts.go:ModelTTSStream).
+func wavHeader24k() []byte {
+	var buf bytes.Buffer
+	w := func(v any) { _ = binary.Write(&buf, binary.LittleEndian, v) }
+	buf.WriteString("RIFF")
+	w(uint32(0xFFFFFFFF))
+	buf.WriteString("WAVE")
+	buf.WriteString("fmt ")
+	w(uint32(16))                     // Subchunk1Size
+	w(uint16(1))                      // PCM
+	w(uint16(1))                      // mono
+	w(uint32(qwen3ttsSampleRate))     // sample rate
+	w(uint32(qwen3ttsSampleRate * 2)) // byte rate = SR * blockAlign
+	w(uint16(2))                      // block align (16-bit mono)
+	w(uint16(16))                     // bits per sample
+	buf.WriteString("data")
+	w(uint32(0xFFFFFFFF))
+	return buf.Bytes()
+}
+
+// floatToPCM16LE clamps each sample to [-1,1] and encodes it as little-endian
+// signed 16-bit PCM.
+func floatToPCM16LE(samples []float32) []byte {
+	out := make([]byte, len(samples)*2)
+	for i, s := range samples {
+		if s > 1 {
+			s = 1
+		} else if s < -1 {
+			s = -1
+		}
+		v := int16(s * 32767)
+		out[i*2] = byte(v)
+		out[i*2+1] = byte(v >> 8)
+	}
+	return out
+}
+
+// writeWAV24k writes samples as a finalized 24 kHz mono 16-bit WAV at dst.
+func writeWAV24k(dst string, samples []float32) error {
+	f, err := os.Create(dst)
+	if err != nil {
+		return fmt.Errorf("qwen3-tts: create %q: %w", dst, err)
+	}
+	enc := wav.NewEncoder(f, qwen3ttsSampleRate, 16, 1, 1)
+	ints := make([]int, len(samples))
+	for i, s := range samples {
+		if s > 1 {
+			s = 1
+		} else if s < -1 {
+			s = -1
+		}
+		ints[i] = int(s * 32767)
+	}
+	b := &audio.IntBuffer{
+		Format:         &audio.Format{NumChannels: 1, SampleRate: qwen3ttsSampleRate},
+		Data:           ints,
+		SourceBitDepth: 16,
+	}
+	if err := enc.Write(b); err != nil {
+		_ = enc.Close()
+		_ = f.Close()
+		return fmt.Errorf("qwen3-tts: encode WAV: %w", err)
+	}
+	if err := enc.Close(); err != nil {
+		_ = f.Close()
+		return fmt.Errorf("qwen3-tts: finalize WAV: %w", err)
+	}
+	return f.Close()
+}
+
+// readWAVAsFloat decodes a WAV file (any sample rate/channels) to a mono
+// float32 slice in [-1,1] for use as cloning reference audio. qwentts expects
+// 24 kHz; callers should supply 24 kHz reference clips.
+func readWAVAsFloat(path string) ([]float32, error) {
+	f, err := os.Open(path)
+	if err != nil {
+		return nil, fmt.Errorf("qwen3-tts: open ref %q: %w", path, err)
+	}
+	defer func() { _ = f.Close() }()
+
+	dec := wav.NewDecoder(f)
+	buf, err := dec.FullPCMBuffer()
+	if err != nil {
+		return nil, fmt.Errorf("qwen3-tts: decode ref %q: %w", path, err)
+	}
+	ch := int(buf.Format.NumChannels)
+	if ch < 1 {
+		ch = 1
+	}
+	bitDepth := int(buf.SourceBitDepth)
+	if bitDepth == 0 {
+		bitDepth = 16
+	}
+	scale := float32(int64(1) << uint(bitDepth-1))
+	n := len(buf.Data) / ch
+	out := make([]float32, n)
+	for i := 0; i < n; i++ {
+		var acc int
+		for c := 0; c < ch; c++ {
+			acc += buf.Data[i*ch+c]
+		}
+		out[i] = float32(acc) / float32(ch) / scale
+	}
+	return out, nil
+}
+
+// runtimeKeepAlive prevents the GC from reclaiming the reference-audio slice
+// while its backing pointer is in use across the C call.
+func runtimeKeepAlive(v any) { runtime.KeepAlive(v) }

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

@@ -0,0 +1,54 @@
+package main
+
+import (
+	"path/filepath"
+
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+// These specs pin the voice-selection logic in resolveRequest, in particular
+// the config-level audio_path (tts.audio_path -> ModelOptions.AudioPath) being
+// used as the default voice-cloning reference. No model/C library is needed:
+// resolveRequest only reads the reference WAV via readWAVAsFloat (pure Go).
+var _ = Describe("resolveRequest voice/clone selection", func() {
+	var dir, refWav string
+
+	BeforeEach(func() {
+		dir = GinkgoT().TempDir()
+		refWav = filepath.Join(dir, "ref.wav")
+		// 0.5s of non-silent 24kHz mono audio as a clone reference.
+		samples := make([]float32, qwen3ttsSampleRate/2)
+		for i := range samples {
+			samples[i] = 0.1
+		}
+		Expect(writeWAV24k(refWav, samples)).To(Succeed())
+	})
+
+	It("uses the config audio_path as the clone reference when Voice is empty", func() {
+		q := &Qwen3TtsCpp{audioPath: refWav}
+		_, _, speaker, _, ref, _, err := q.resolveRequest(&pb.TTSRequest{Text: "hi"})
+		Expect(err).ToNot(HaveOccurred())
+		Expect(speaker).To(BeEmpty())
+		Expect(len(ref)).To(Equal(qwen3ttsSampleRate / 2))
+	})
+
+	It("lets a per-request audio Voice override audio_path", func() {
+		other := filepath.Join(dir, "other.wav")
+		Expect(writeWAV24k(other, make([]float32, 100))).To(Succeed())
+		q := &Qwen3TtsCpp{audioPath: refWav}
+		_, _, speaker, _, ref, _, err := q.resolveRequest(&pb.TTSRequest{Text: "hi", Voice: other})
+		Expect(err).ToNot(HaveOccurred())
+		Expect(speaker).To(BeEmpty())
+		Expect(len(ref)).To(Equal(100))
+	})
+
+	It("does not trigger audio_path cloning for a named-speaker Voice", func() {
+		q := &Qwen3TtsCpp{audioPath: refWav}
+		_, _, speaker, _, ref, _, err := q.resolveRequest(&pb.TTSRequest{Text: "hi", Voice: "serena"})
+		Expect(err).ToNot(HaveOccurred())
+		Expect(speaker).To(Equal("serena"))
+		Expect(ref).To(BeNil())
+	})
+})

backend/go/qwen3-tts-cpp/cpp/goqwen3ttscpp.cpp

@@ -1,161 +1,191 @@
 #include "goqwen3ttscpp.h"
 #include "ggml-backend.h"
-#include "qwen3_tts.h"
+#include "qwen.h"
 
-#include <cmath>
 #include <cstdio>
 #include <cstdlib>
 #include <cstring>
-#include <string>
 
-using namespace qwen3_tts;
+static qt_context *g_ctx = nullptr;
 
-// Global engine (loaded once, reused across requests)
-static Qwen3TTS *g_engine = nullptr;
-static bool g_loaded = false;
-static int g_threads = 4;
-
-static void ggml_log_cb(enum ggml_log_level level, const char *log, void *data) {
-    const char *level_str;
+static void ggml_log_cb(enum ggml_log_level level, const char *log,
+                        void * /*data*/) {
     if (!log)
         return;
+    const char *lvl = "?????";
     switch (level) {
-    case GGML_LOG_LEVEL_DEBUG:
-        level_str = "DEBUG";
-        break;
-    case GGML_LOG_LEVEL_INFO:
-        level_str = "INFO";
-        break;
-    case GGML_LOG_LEVEL_WARN:
-        level_str = "WARN";
-        break;
-    case GGML_LOG_LEVEL_ERROR:
-        level_str = "ERROR";
-        break;
-    default:
-        level_str = "?????";
-        break;
+    case GGML_LOG_LEVEL_DEBUG: lvl = "DEBUG"; break;
+    case GGML_LOG_LEVEL_INFO:  lvl = "INFO";  break;
+    case GGML_LOG_LEVEL_WARN:  lvl = "WARN";  break;
+    case GGML_LOG_LEVEL_ERROR: lvl = "ERROR"; break;
+    default: break;
     }
-    fprintf(stderr, "[%-5s] ", level_str);
-    fputs(log, stderr);
+    fprintf(stderr, "[%-5s] %s", lvl, log);
     fflush(stderr);
 }
 
-// Map language string to language_id token used by the model
-static int language_to_id(const char *lang) {
-    if (!lang || lang[0] == '\0')
-        return 2050; // default: English
-    std::string l(lang);
-    if (l == "en")
-        return 2050;
-    if (l == "ru")
-        return 2069;
-    if (l == "zh")
-        return 2055;
-    if (l == "ja")
-        return 2058;
-    if (l == "ko")
-        return 2064;
-    if (l == "de")
-        return 2053;
-    if (l == "fr")
-        return 2061;
-    if (l == "es")
-        return 2054;
-    if (l == "it")
-        return 2056;
-    if (l == "pt")
-        return 2057;
-    fprintf(stderr, "[qwen3-tts-cpp] Unknown language '%s', defaulting to English\n",
-            lang);
-    return 2050;
-}
-
-int load_model(const char *model_dir, int n_threads) {
+int qt3_load(const char *talker_path, const char *codec_path, int use_fa,
+             int clamp_fp16) {
     ggml_log_set(ggml_log_cb, nullptr);
     ggml_backend_load_all();
 
-    if (n_threads <= 0)
-        n_threads = 4;
-    g_threads = n_threads;
-
-    fprintf(stderr, "[qwen3-tts-cpp] Loading models from %s (threads=%d)\n",
-            model_dir, n_threads);
-
-    g_engine = new Qwen3TTS();
-    if (!g_engine->load_models(model_dir)) {
-        fprintf(stderr, "[qwen3-tts-cpp] FATAL: failed to load models from %s\n",
-                model_dir);
-        delete g_engine;
-        g_engine = nullptr;
+    if (!talker_path || talker_path[0] == '\0') {
+        fprintf(stderr, "[qwen3-tts-cpp] ERROR: talker_path is required\n");
         return 1;
     }
-
-    g_loaded = true;
-    fprintf(stderr, "[qwen3-tts-cpp] Models loaded successfully\n");
-    return 0;
-}
-
-int synthesize(const char *text, const char *ref_audio_path, const char *dst,
-               const char *language, float temperature, float top_p,
-               int top_k, float repetition_penalty, int max_audio_tokens,
-               int n_threads) {
-    if (!g_loaded || !g_engine) {
-        fprintf(stderr, "[qwen3-tts-cpp] ERROR: models not loaded\n");
-        return 1;
-    }
-
-    if (!text || !dst) {
-        fprintf(stderr, "[qwen3-tts-cpp] ERROR: text and dst are required\n");
+    if (!codec_path || codec_path[0] == '\0') {
+        fprintf(stderr, "[qwen3-tts-cpp] ERROR: codec_path is required\n");
         return 2;
     }
 
-    tts_params params;
-    params.max_audio_tokens = max_audio_tokens > 0 ? max_audio_tokens : 4096;
-    params.temperature = temperature;
-    params.top_p = top_p;
-    params.top_k = top_k;
-    params.repetition_penalty = repetition_penalty;
-    params.n_threads = n_threads > 0 ? n_threads : g_threads;
-    params.language_id = language_to_id(language);
+    qt_init_params p;
+    qt_init_default_params(&p);
+    p.talker_path = talker_path;
+    p.codec_path = codec_path;
+    p.use_fa = use_fa != 0;
+    p.clamp_fp16 = clamp_fp16 != 0;
 
-    fprintf(stderr, "[qwen3-tts-cpp] Synthesizing: text='%.50s%s', lang_id=%d, "
-                    "temp=%.2f, threads=%d\n",
-            text, (strlen(text) > 50 ? "..." : ""), params.language_id,
-            temperature, params.n_threads);
+    fprintf(stderr, "[qwen3-tts-cpp] Loading talker=%s codec=%s\n", talker_path,
+            codec_path);
 
-    tts_result result;
-    bool has_ref = ref_audio_path && ref_audio_path[0] != '\0';
-
-    if (has_ref) {
-        fprintf(stderr, "[qwen3-tts-cpp] Voice cloning with ref: %s\n",
-                ref_audio_path);
-        result = g_engine->synthesize_with_voice(text, ref_audio_path, params);
-    } else {
-        result = g_engine->synthesize(text, params);
-    }
-
-    if (!result.success) {
-        fprintf(stderr, "[qwen3-tts-cpp] ERROR: synthesis failed: %s\n",
-                result.error_msg.c_str());
+    g_ctx = qt_init(&p);
+    if (!g_ctx) {
+        fprintf(stderr, "[qwen3-tts-cpp] FATAL: qt_init failed: %s\n",
+                qt_last_error());
         return 3;
     }
-
-    int n_samples = (int)result.audio.size();
-    if (n_samples == 0) {
-        fprintf(stderr, "[qwen3-tts-cpp] ERROR: synthesis produced no samples\n");
-        return 4;
-    }
-
-    fprintf(stderr,
-            "[qwen3-tts-cpp] Synthesis done: %d samples (%.2fs @ 24kHz)\n",
-            n_samples, (float)n_samples / 24000.0f);
-
-    if (!save_audio_file(dst, result.audio, result.sample_rate)) {
-        fprintf(stderr, "[qwen3-tts-cpp] ERROR: failed to write %s\n", dst);
-        return 5;
-    }
-
-    fprintf(stderr, "[qwen3-tts-cpp] Wrote %s\n", dst);
+    fprintf(stderr, "[qwen3-tts-cpp] Model loaded (%s)\n", qt_version());
     return 0;
 }
+
+// Fill a qt_tts_params from the flat wrapper arguments. Unset/zero scalars keep
+// the qt defaults (temperature 0.9, top_k 50, top_p 1.0, rep 1.05, max 2048).
+static void fill_params(qt_tts_params *tp, const char *text, const char *lang,
+                        const char *instruct, const char *speaker,
+                        const float *ref_samples, int ref_n,
+                        const char *ref_text, long long seed, float temperature,
+                        int top_k, float top_p, float repetition_penalty,
+                        int max_new_tokens) {
+    qt_tts_default_params(tp);
+    tp->text = text ? text : "";
+    if (lang && lang[0] != '\0')
+        tp->lang = lang; // else keep default NULL -> auto
+    if (instruct && instruct[0] != '\0')
+        tp->instruct = instruct;
+    if (speaker && speaker[0] != '\0')
+        tp->speaker = speaker;
+    if (ref_samples && ref_n > 0) {
+        tp->ref_audio_24k = ref_samples;
+        tp->ref_n_samples = ref_n;
+        if (ref_text && ref_text[0] != '\0')
+            tp->ref_text = ref_text;
+    }
+    if (seed >= 0)
+        tp->seed = (int64_t)seed; // else default -1 (random)
+    if (temperature > 0.0f)
+        tp->temperature = temperature;
+    if (top_k > 0)
+        tp->top_k = top_k;
+    if (top_p > 0.0f)
+        tp->top_p = top_p;
+    if (repetition_penalty > 0.0f)
+        tp->repetition_penalty = repetition_penalty;
+    if (max_new_tokens > 0)
+        tp->max_new_tokens = max_new_tokens;
+}
+
+float *qt3_tts(const char *text, const char *lang, const char *instruct,
+               const char *speaker, const float *ref_samples, int ref_n,
+               const char *ref_text, long long seed, float temperature,
+               int top_k, float top_p, float repetition_penalty,
+               int max_new_tokens, int *out_n) {
+    if (out_n)
+        *out_n = 0;
+    if (!g_ctx) {
+        fprintf(stderr, "[qwen3-tts-cpp] ERROR: model not loaded\n");
+        return nullptr;
+    }
+    if (!text || text[0] == '\0') {
+        fprintf(stderr, "[qwen3-tts-cpp] ERROR: text is required\n");
+        return nullptr;
+    }
+    qt_tts_params tp;
+    fill_params(&tp, text, lang, instruct, speaker, ref_samples, ref_n,
+                ref_text, seed, temperature, top_k, top_p, repetition_penalty,
+                max_new_tokens);
+
+    qt_audio out = {0};
+    enum qt_status rc = qt_synthesize(g_ctx, &tp, &out);
+    if (rc != QT_STATUS_OK || out.n_samples <= 0 || !out.samples) {
+        fprintf(stderr, "[qwen3-tts-cpp] ERROR: synthesize failed (rc=%d): %s\n",
+                (int)rc, qt_last_error());
+        qt_audio_free(&out);
+        return nullptr;
+    }
+
+    // Copy into a plain malloc buffer the Go side frees via qt3_pcm_free.
+    size_t bytes = (size_t)out.n_samples * sizeof(float);
+    float *buf = (float *)malloc(bytes);
+    if (!buf) {
+        fprintf(stderr, "[qwen3-tts-cpp] ERROR: malloc(%zu) failed\n", bytes);
+        qt_audio_free(&out);
+        return nullptr;
+    }
+    memcpy(buf, out.samples, bytes);
+    if (out_n)
+        *out_n = out.n_samples;
+    qt_audio_free(&out);
+    return buf;
+}
+
+int qt3_tts_stream(const char *text, const char *lang, const char *instruct,
+                   const char *speaker, const float *ref_samples, int ref_n,
+                   const char *ref_text, long long seed, float temperature,
+                   int top_k, float top_p, float repetition_penalty,
+                   int max_new_tokens, qt3_chunk_cb cb, void *user_data) {
+    if (!g_ctx) {
+        fprintf(stderr, "[qwen3-tts-cpp] ERROR: model not loaded\n");
+        return 1;
+    }
+    if (!cb) {
+        fprintf(stderr, "[qwen3-tts-cpp] ERROR: stream callback is null\n");
+        return 2;
+    }
+    if (!text || text[0] == '\0') {
+        fprintf(stderr, "[qwen3-tts-cpp] ERROR: text is required\n");
+        return 4;
+    }
+    qt_tts_params tp;
+    fill_params(&tp, text, lang, instruct, speaker, ref_samples, ref_n,
+                ref_text, seed, temperature, top_k, top_p, repetition_penalty,
+                max_new_tokens);
+    // qt_audio_chunk_cb has the identical signature to qt3_chunk_cb
+    // (bool vs int return are ABI-compatible; non-zero == true).
+    tp.on_chunk = (qt_audio_chunk_cb)cb;
+    tp.on_chunk_user_data = user_data;
+
+    qt_audio out = {0}; // stays empty in streaming mode
+    enum qt_status rc = qt_synthesize(g_ctx, &tp, &out);
+    qt_audio_free(&out);
+    if (rc != QT_STATUS_OK && rc != QT_STATUS_CANCELLED) {
+        fprintf(stderr, "[qwen3-tts-cpp] ERROR: stream synth failed (rc=%d): %s\n",
+                (int)rc, qt_last_error());
+        return 3;
+    }
+    return 0;
+}
+
+void qt3_pcm_free(float *p) { free(p); }
+
+void qt3_unload(void) {
+    if (g_ctx) {
+        qt_free(g_ctx);
+        g_ctx = nullptr;
+    }
+}
+
+int qt3_n_speakers(void) { return g_ctx ? qt_n_speakers(g_ctx) : 0; }
+
+const char *qt3_speaker_name(int i) {
+    return g_ctx ? qt_speaker_name(g_ctx, i) : nullptr;
+}

backend/go/qwen3-tts-cpp/cpp/goqwen3ttscpp.h

@@ -1,12 +1,47 @@
 #pragma once
 
-#include <cstddef>
-#include <cstdint>
-
 extern "C" {
-int load_model(const char *model_dir, int n_threads);
-int synthesize(const char *text, const char *ref_audio_path, const char *dst,
-               const char *language, float temperature, float top_p,
-               int top_k, float repetition_penalty, int max_audio_tokens,
-               int n_threads);
+
+// Streaming PCM chunk callback. samples is mono float PCM at 24 kHz, valid
+// only for the duration of the call. Return non-zero to continue, 0 to abort.
+typedef int (*qt3_chunk_cb)(const float *samples, int n_samples,
+                            void *user_data);
+
+// Load the talker + codec/tokenizer GGUFs. use_fa / clamp_fp16 map to
+// qt_init_params (the qt ABI exposes no thread count; ggml uses its own
+// default). Returns 0 on success, non-zero on failure.
+int qt3_load(const char *talker_path, const char *codec_path, int use_fa,
+             int clamp_fp16);
+
+// Synthesize to a malloc'd float PCM buffer (caller frees via qt3_pcm_free).
+// The synthesis mode (base / custom_voice / voice_design) is auto-detected by
+// qt from the talker GGUF; speaker is honoured only for custom_voice, instruct
+// for voice_design / custom_voice, and ref_samples (+ optional ref_text) drive
+// base-mode cloning. qt enforces the rules and we surface qt_last_error() on
+// QT_STATUS_MODE_INVALID. Writes the sample count to *out_n. Returns NULL on
+// failure (out_n set to 0).
+float *qt3_tts(const char *text, const char *lang, const char *instruct,
+               const char *speaker, const float *ref_samples, int ref_n,
+               const char *ref_text, long long seed, float temperature,
+               int top_k, float top_p, float repetition_penalty,
+               int max_new_tokens, int *out_n);
+
+// Streaming synthesis: cb is invoked per PCM chunk as audio is produced. Same
+// param semantics as qt3_tts. Returns 0 on success.
+int qt3_tts_stream(const char *text, const char *lang, const char *instruct,
+                   const char *speaker, const float *ref_samples, int ref_n,
+                   const char *ref_text, long long seed, float temperature,
+                   int top_k, float top_p, float repetition_penalty,
+                   int max_new_tokens, qt3_chunk_cb cb, void *user_data);
+
+// Free a buffer returned by qt3_tts.
+void qt3_pcm_free(float *p);
+
+// Release the qt context.
+void qt3_unload(void);
+
+// Named-speaker introspection (custom_voice models). Returns 0 / NULL when no
+// model is loaded or the index is out of range.
+int qt3_n_speakers(void);
+const char *qt3_speaker_name(int i);
 }

backend/go/qwen3-tts-cpp/e2e_test.go

@@ -0,0 +1,95 @@
+package main
+
+import (
+	"math"
+	"os"
+	"strings"
+
+	"github.com/ebitengine/purego"
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+func ttsReq(text, voice string, lang *string, dst string) *pb.TTSRequest {
+	return &pb.TTSRequest{Text: text, Voice: voice, Language: lang, Dst: dst}
+}
+
+var _ = Describe("qwen3-tts-cpp e2e", Label("e2e"), func() {
+	var loaded bool
+
+	BeforeEach(func() {
+		modelPath := os.Getenv("QWEN3TTS_MODEL")
+		codecPath := os.Getenv("QWEN3TTS_CODEC")
+		if modelPath == "" || codecPath == "" {
+			Skip("QWEN3TTS_MODEL / QWEN3TTS_CODEC not set; skipping e2e")
+		}
+		if !loaded {
+			lib := os.Getenv("QWEN3TTS_LIBRARY")
+			if lib == "" {
+				lib = "./libgoqwen3ttscpp-fallback.so"
+			}
+			h, err := purego.Dlopen(lib, purego.RTLD_NOW|purego.RTLD_GLOBAL)
+			Expect(err).ToNot(HaveOccurred())
+			purego.RegisterLibFunc(&CppLoad, h, "qt3_load")
+			purego.RegisterLibFunc(&CppTTS, h, "qt3_tts")
+			purego.RegisterLibFunc(&CppTTSStream, h, "qt3_tts_stream")
+			purego.RegisterLibFunc(&CppPCMFree, h, "qt3_pcm_free")
+			purego.RegisterLibFunc(&CppUnload, h, "qt3_unload")
+			Expect(CppLoad(modelPath, codecPath, 1, 0)).To(Equal(0))
+			loaded = true
+		}
+	})
+
+	It("synthesizes a WAV file via TTS", func() {
+		b := &Qwen3TtsCpp{opts: loadOptions{seed: 42, useFA: true}}
+		dst := GinkgoT().TempDir() + "/out.wav"
+		lang := "english"
+		err := b.TTS(ttsReq("Hello world.", "", &lang, dst))
+		Expect(err).ToNot(HaveOccurred())
+		fi, err := os.Stat(dst)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(fi.Size()).To(BeNumerically(">", int64(44)))
+	})
+
+	It("streams audio chunks via TTSStream", func() {
+		b := &Qwen3TtsCpp{opts: loadOptions{seed: 42, useFA: true}}
+		results := make(chan []byte, 1024)
+		lang := "english"
+		done := make(chan error, 1)
+		go func() { done <- b.TTSStream(ttsReq("Hello there, streaming test.", "", &lang, ""), results) }()
+
+		var chunks int
+		var first []byte
+		for c := range results {
+			if chunks == 0 {
+				first = c
+			}
+			chunks++
+		}
+		Expect(<-done).ToNot(HaveOccurred())
+		Expect(chunks).To(BeNumerically(">=", 2))
+		Expect(string(first[0:4])).To(Equal("RIFF"))
+		Expect(strings.HasPrefix(string(first[8:12]), "WAVE")).To(BeTrue())
+	})
+
+	It("clones a voice from the config audio_path reference", func() {
+		// 1s of 24kHz mono audio as a clone reference; the base model carries
+		// a speaker encoder, so audio_path drives x-vector voice cloning.
+		ref := GinkgoT().TempDir() + "/ref.wav"
+		samples := make([]float32, qwen3ttsSampleRate)
+		for i := range samples {
+			samples[i] = float32(0.05 * math.Sin(float64(i)*0.06))
+		}
+		Expect(writeWAV24k(ref, samples)).To(Succeed())
+
+		b := &Qwen3TtsCpp{opts: loadOptions{seed: 42, useFA: true}, audioPath: ref}
+		dst := GinkgoT().TempDir() + "/clone.wav"
+		lang := "english"
+		// Empty Voice -> the config audio_path is used as the clone reference.
+		Expect(b.TTS(ttsReq("Cloned voice test.", "", &lang, dst))).To(Succeed())
+		fi, err := os.Stat(dst)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(fi.Size()).To(BeNumerically(">", int64(44)))
+	})
+})

backend/go/qwen3-tts-cpp/goqwen3ttscpp.go

@@ -5,108 +5,225 @@ import (
 	"os"
 	"path/filepath"
 	"strings"
+	"sync"
+	"unsafe"
 
+	"github.com/ebitengine/purego"
 	"github.com/mudler/LocalAI/pkg/grpc/base"
 	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
 )
 
 var (
-	CppLoadModel  func(modelDir string, nThreads int) int
-	CppSynthesize func(text, refAudioPath, dst, language string,
-		temperature, topP float32, topK int,
-		repetitionPenalty float32, maxAudioTokens, nThreads int) int
+	// qt3_load(talker_path, codec_path, use_fa, clamp_fp16) int
+	CppLoad func(talkerPath, codecPath string, useFA, clampFP16 int) int
+	// qt3_tts(text, lang, instruct, speaker, ref_samples, ref_n, ref_text,
+	//         seed, temperature, top_k, top_p, rep_pen, max_new, out_n) -> float*
+	CppTTS func(text, lang, instruct, speaker string, refSamples unsafe.Pointer,
+		refN int, refText string, seed int64, temperature float32, topK int,
+		topP, repPen float32, maxNew int, outN unsafe.Pointer) uintptr
+	// qt3_tts_stream(..., cb, user) int
+	CppTTSStream func(text, lang, instruct, speaker string, refSamples unsafe.Pointer,
+		refN int, refText string, seed int64, temperature float32, topK int,
+		topP, repPen float32, maxNew int, cb uintptr, user uintptr) int
+	CppPCMFree func(ptr uintptr)
+	CppUnload  func()
 )
 
 type Qwen3TtsCpp struct {
 	base.SingleThread
-	threads int
-}
-
-// languageNameAliases maps common full language names to the canonical
-// two-letter code understood by the C++ language_to_id table.
-var languageNameAliases = map[string]string{
-	"english":    "en",
-	"russian":    "ru",
-	"chinese":    "zh",
-	"japanese":   "ja",
-	"korean":     "ko",
-	"german":     "de",
-	"french":     "fr",
-	"spanish":    "es",
-	"italian":    "it",
-	"portuguese": "pt",
-}
-
-// normalizeLanguage coerces a caller-supplied language into the canonical code
-// the model expects. It lowercases, trims, strips any region/locale suffix
-// (en-US, en_US, ja.JP -> en/ja), and resolves common full names (english -> en).
-// An empty input stays empty so the C++ side applies its English default; an
-// unrecognized value is returned normalized so C++ can log it and default.
-func normalizeLanguage(lang string) string {
-	lang = strings.ToLower(strings.TrimSpace(lang))
-	if lang == "" {
-		return ""
-	}
-
-	// Strip region/locale suffix: keep the segment before the first separator.
-	if i := strings.IndexAny(lang, "-_."); i >= 0 {
-		lang = lang[:i]
-	}
-
-	if code, ok := languageNameAliases[lang]; ok {
-		return code
-	}
-	return lang
+	opts loadOptions
+	// audioPath is the model-config reference voice (tts.audio_path), the
+	// default clone reference when a request omits an audio Voice.
+	audioPath string
 }
 
 func (q *Qwen3TtsCpp) Load(opts *pb.ModelOptions) error {
-	// ModelFile is the model directory path (containing GGUF files)
-	modelDir := opts.ModelFile
-	if modelDir == "" {
-		modelDir = opts.ModelPath
+	model := opts.ModelFile
+	if model == "" {
+		model = opts.ModelPath
+	}
+	if !filepath.IsAbs(model) && opts.ModelPath != "" {
+		model = filepath.Join(opts.ModelPath, model)
 	}
 
-	// Resolve relative paths
-	if !filepath.IsAbs(modelDir) && opts.ModelPath != "" {
-		modelDir = filepath.Join(opts.ModelPath, modelDir)
+	q.opts = parseOptions(opts.Options)
+
+	// Resolve the codec/tokenizer GGUF: explicit option, else auto-discover a
+	// *tokenizer*.gguf sibling of the talker model.
+	codec := q.opts.codecPath
+	if codec != "" && !filepath.IsAbs(codec) {
+		codec = filepath.Join(filepath.Dir(model), codec)
+	}
+	if codec == "" {
+		codec = discoverTokenizer(filepath.Dir(model))
+	}
+	if codec == "" {
+		return fmt.Errorf("qwen3-tts: no codec/tokenizer GGUF found; set option 'tokenizer:<file>'")
+	}
+	q.opts.codecPath = codec
+
+	q.audioPath = opts.AudioPath
+	if q.audioPath != "" && !filepath.IsAbs(q.audioPath) {
+		q.audioPath = filepath.Join(filepath.Dir(model), q.audioPath)
 	}
 
-	threads := int(opts.Threads)
-	if threads <= 0 {
-		threads = 4
+	useFA := boolToInt(q.opts.useFA)
+	clamp := boolToInt(q.opts.clampFP16)
+
+	fmt.Fprintf(os.Stderr, "[qwen3-tts-cpp] Load talker=%s codec=%s use_fa=%d clamp_fp16=%d\n",
+		model, codec, useFA, clamp)
+
+	if rc := CppLoad(model, codec, useFA, clamp); rc != 0 {
+		return fmt.Errorf("qwen3-tts: failed to load model (rc=%d)", rc)
 	}
-	q.threads = threads
-
-	fmt.Fprintf(os.Stderr, "[qwen3-tts-cpp] Loading models from: %s (threads=%d)\n", modelDir, threads)
-
-	if ret := CppLoadModel(modelDir, threads); ret != 0 {
-		return fmt.Errorf("failed to load qwen3-tts model (error code: %d)", ret)
-	}
-
 	return nil
 }
 
+// discoverTokenizer returns the first *tokenizer*.gguf in dir, or "".
+func discoverTokenizer(dir string) string {
+	entries, err := os.ReadDir(dir)
+	if err != nil {
+		return ""
+	}
+	for _, e := range entries {
+		name := strings.ToLower(e.Name())
+		if strings.Contains(name, "tokenizer") && strings.HasSuffix(name, ".gguf") {
+			return filepath.Join(dir, e.Name())
+		}
+	}
+	return ""
+}
+
+func boolToInt(b bool) int {
+	if b {
+		return 1
+	}
+	return 0
+}
+
+func optStr(p *string) string {
+	if p == nil {
+		return ""
+	}
+	return *p
+}
+
+// resolveRequest derives the synthesis inputs from a TTSRequest:
+// language, instruct, speaker, ref-audio samples, ref-text and sampling.
+func (q *Qwen3TtsCpp) resolveRequest(req *pb.TTSRequest) (lang, instruct, speaker, refText string, ref []float32, s sampling, err error) {
+	lang = normalizeLanguage(optStr(req.Language))
+	instruct = optStr(req.Instructions)
+
+	var refPath string
+	speaker, refPath = resolveVoice(req.Voice)
+	if refPath == "" && speaker == "" && q.audioPath != "" {
+		// No per-request voice: fall back to the config clone reference.
+		refPath = q.audioPath
+	}
+	if refPath != "" {
+		ref, err = readWAVAsFloat(refPath)
+		if err != nil {
+			return
+		}
+	}
+
+	if req.Params != nil {
+		refText = req.Params["ref_text"]
+	}
+	s = parseSampling(req.Params, q.opts.seed)
+	return
+}
+
 func (q *Qwen3TtsCpp) TTS(req *pb.TTSRequest) error {
-	text := req.Text
-	voice := req.Voice // reference audio path for voice cloning (empty = no cloning)
-	dst := req.Dst
-	language := ""
-	if req.Language != nil {
-		language = normalizeLanguage(*req.Language)
+	if req.Dst == "" {
+		return fmt.Errorf("qwen3-tts: TTS requires a destination path")
+	}
+	if req.Text == "" {
+		return fmt.Errorf("qwen3-tts: TTS requires text")
+	}
+	lang, instruct, speaker, refText, ref, s, err := q.resolveRequest(req)
+	if err != nil {
+		return err
+	}
+	var refPtr unsafe.Pointer
+	if len(ref) > 0 {
+		refPtr = unsafe.Pointer(&ref[0])
 	}
 
-	// Synthesis parameters with sensible defaults
-	temperature := float32(0.9)
-	topP := float32(0.8)
-	topK := 50
-	repetitionPenalty := float32(1.05)
-	maxAudioTokens := 4096
+	var n int32
+	ptr := CppTTS(req.Text, lang, instruct, speaker, refPtr, len(ref), refText,
+		s.seed, s.temperature, s.topK, s.topP, s.repPen, s.maxNew, unsafe.Pointer(&n))
+	runtimeKeepAlive(ref)
+	if ptr == 0 {
+		return fmt.Errorf("qwen3-tts: synthesis failed")
+	}
+	// Register the free as soon as we own a non-null buffer, so the n<=0 guard
+	// below cannot leak it (defensive: the C contract returns NULL on failure).
+	defer CppPCMFree(ptr)
+	if n <= 0 {
+		return fmt.Errorf("qwen3-tts: synthesis produced no samples")
+	}
+	src := unsafe.Slice((*float32)(unsafe.Pointer(ptr)), int(n)) //nolint:govet // C-allocated PCM, copied out before free
+	out := make([]float32, int(n))
+	copy(out, src)
+	return writeWAV24k(req.Dst, out)
+}
 
-	if ret := CppSynthesize(text, voice, dst, language,
-		temperature, topP, topK, repetitionPenalty,
-		maxAudioTokens, q.threads); ret != 0 {
-		return fmt.Errorf("failed to synthesize audio (error code: %d)", ret)
+// streamState carries the active TTSStream channel to the single shared C
+// callback. base.SingleThread serializes TTS/TTSStream, so one global slot is
+// safe and avoids leaking a purego callback per request (purego callbacks
+// cannot be freed and are capped).
+var (
+	streamMu     sync.Mutex
+	streamChan   chan []byte
+	streamCbOnce sync.Once
+	streamCbPtr  uintptr
+)
+
+// streamCallback is registered once and forwards each PCM chunk to streamChan.
+func streamCallback(samples *float32, nSamples int32, _ uintptr) uintptr {
+	if nSamples <= 0 || samples == nil || streamChan == nil {
+		return 1 // continue
+	}
+	src := unsafe.Slice(samples, int(nSamples))
+	cp := make([]float32, int(nSamples)) // copy out of C memory before returning
+	copy(cp, src)
+	streamChan <- floatToPCM16LE(cp)
+	return 1 // continue
+}
+
+func (q *Qwen3TtsCpp) TTSStream(req *pb.TTSRequest, results chan []byte) error {
+	defer close(results)
+	if req.Text == "" {
+		return fmt.Errorf("qwen3-tts: TTSStream requires text")
 	}
 
+	streamCbOnce.Do(func() {
+		streamCbPtr = purego.NewCallback(streamCallback)
+	})
+
+	lang, instruct, speaker, refText, ref, s, err := q.resolveRequest(req)
+	if err != nil {
+		return err
+	}
+	var refPtr unsafe.Pointer
+	if len(ref) > 0 {
+		refPtr = unsafe.Pointer(&ref[0])
+	}
+
+	// Emit the WAV header first so the HTTP layer gets a self-describing stream.
+	results <- wavHeader24k()
+
+	streamMu.Lock()
+	streamChan = results
+	rc := CppTTSStream(req.Text, lang, instruct, speaker, refPtr, len(ref), refText,
+		s.seed, s.temperature, s.topK, s.topP, s.repPen, s.maxNew, streamCbPtr, 0)
+	streamChan = nil
+	streamMu.Unlock()
+	runtimeKeepAlive(ref)
+
+	if rc != 0 {
+		return fmt.Errorf("qwen3-tts: streaming synthesis failed (rc=%d)", rc)
+	}
 	return nil
 }

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

@@ -1,53 +0,0 @@
-package main
-
-import (
-	"testing"
-
-	. "github.com/onsi/ginkgo/v2"
-	. "github.com/onsi/gomega"
-)
-
-func TestLanguageNormalization(t *testing.T) {
-	RegisterFailHandler(Fail)
-	RunSpecs(t, "qwen3-tts-cpp language normalization")
-}
-
-var _ = Describe("normalizeLanguage", func() {
-	DescribeTable("maps caller input to the canonical model language code",
-		func(input, expected string) {
-			Expect(normalizeLanguage(input)).To(Equal(expected))
-		},
-		// Canonical codes pass through unchanged
-		Entry("canonical en", "en", "en"),
-		Entry("canonical zh", "zh", "zh"),
-		Entry("canonical pt", "pt", "pt"),
-
-		// Case-insensitive
-		Entry("uppercase", "EN", "en"),
-		Entry("mixed case", "Ja", "ja"),
-
-		// Surrounding whitespace
-		Entry("trims whitespace", "  en  ", "en"),
-
-		// Region/locale stripping
-		Entry("BCP-47 region", "en-US", "en"),
-		Entry("underscore region", "en_US", "en"),
-		Entry("dotted locale", "ja.JP", "ja"),
-		Entry("region + case", "ZH-CN", "zh"),
-
-		// Full-name aliases
-		Entry("english name", "english", "en"),
-		Entry("chinese name cased", "Chinese", "zh"),
-		Entry("japanese name", "japanese", "ja"),
-		Entry("russian name", "russian", "ru"),
-		Entry("portuguese name", "portuguese", "pt"),
-
-		// Empty stays empty (C++ applies the English default)
-		Entry("empty", "", ""),
-		Entry("whitespace only", "   ", ""),
-
-		// Unknown values pass through normalized so C++ can log + default
-		Entry("unknown code", "klingon", "klingon"),
-		Entry("unknown with region", "xx-YY", "xx"),
-	)
-})

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

@@ -19,24 +19,25 @@ type LibFuncs struct {
 }
 
 func main() {
-	// Get library name from environment variable, default to fallback
 	libName := os.Getenv("QWEN3TTS_LIBRARY")
 	if libName == "" {
 		libName = "./libgoqwen3ttscpp-fallback.so"
 	}
 
-	gosd, err := purego.Dlopen(libName, purego.RTLD_NOW|purego.RTLD_GLOBAL)
+	lib, err := purego.Dlopen(libName, purego.RTLD_NOW|purego.RTLD_GLOBAL)
 	if err != nil {
 		panic(err)
 	}
 
 	libFuncs := []LibFuncs{
-		{&CppLoadModel, "load_model"},
-		{&CppSynthesize, "synthesize"},
+		{&CppLoad, "qt3_load"},
+		{&CppTTS, "qt3_tts"},
+		{&CppTTSStream, "qt3_tts_stream"},
+		{&CppPCMFree, "qt3_pcm_free"},
+		{&CppUnload, "qt3_unload"},
 	}
-
 	for _, lf := range libFuncs {
-		purego.RegisterLibFunc(lf.FuncPtr, gosd, lf.Name)
+		purego.RegisterLibFunc(lf.FuncPtr, lib, lf.Name)
 	}
 
 	flag.Parse()

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

@@ -0,0 +1,161 @@
+package main
+
+import (
+	"strconv"
+	"strings"
+)
+
+// loadOptions holds the parsed model-level options.
+type loadOptions struct {
+	codecPath string
+	useFA     bool
+	clampFP16 bool
+	seed      int64
+}
+
+// sampling holds per-request generation parameters with qt defaults applied.
+type sampling struct {
+	temperature float32
+	topK        int
+	topP        float32
+	repPen      float32
+	maxNew      int
+	seed        int64
+}
+
+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
+}
+
+func parseBool(v string) bool { return v == "true" || v == "1" }
+
+// parseOptions reads the backend "key:value" option slice. Unknown keys are
+// ignored. Defaults: use_fa true (qt default; CPU still uses the F32 chain),
+// seed -1 (engine random).
+func parseOptions(opts []string) loadOptions {
+	o := loadOptions{useFA: true, seed: -1}
+	for _, oo := range opts {
+		key, value, ok := splitOption(oo)
+		if !ok {
+			continue
+		}
+		switch key {
+		case "tokenizer", "codec":
+			o.codecPath = value
+		case "use_fa":
+			o.useFA = parseBool(value)
+		case "clamp_fp16":
+			o.clampFP16 = parseBool(value)
+		case "seed":
+			if n, err := strconv.ParseInt(value, 10, 64); err == nil {
+				o.seed = n
+			}
+		}
+	}
+	return o
+}
+
+// languageAliases maps codes / locales / full names to the upstream qwentts
+// language names. "auto" (and empty) map to "" so the engine auto-detects.
+var languageAliases = map[string]string{
+	"en": "english", "english": "english",
+	"zh": "chinese", "chinese": "chinese", "mandarin": "chinese",
+	"ja": "japanese", "japanese": "japanese",
+	"ko": "korean", "korean": "korean",
+	"de": "german", "german": "german",
+	"fr": "french", "french": "french",
+	"es": "spanish", "spanish": "spanish",
+	"it": "italian", "italian": "italian",
+	"pt": "portuguese", "portuguese": "portuguese",
+	"ru": "russian", "russian": "russian",
+	"auto": "",
+}
+
+// normalizeLanguage lowercases, trims, strips a region/locale suffix
+// (en-US -> en), and resolves to the qwentts language name. Empty stays empty
+// (engine auto-detects); an unknown value passes through normalized.
+func normalizeLanguage(lang string) string {
+	lang = strings.ToLower(strings.TrimSpace(lang))
+	if lang == "" {
+		return ""
+	}
+	if i := strings.IndexAny(lang, "-_."); i >= 0 {
+		lang = lang[:i]
+	}
+	if v, ok := languageAliases[lang]; ok {
+		return v
+	}
+	return lang
+}
+
+var refAudioExts = []string{".wav", ".flac", ".mp3", ".ogg", ".m4a"}
+
+// resolveVoice interprets the request Voice field: a value ending in a known
+// audio extension is a clone-reference path; anything else is a named speaker
+// (custom_voice). Empty input yields no speaker and no reference.
+func resolveVoice(voice string) (speaker, refPath string) {
+	v := strings.TrimSpace(voice)
+	if v == "" {
+		return "", ""
+	}
+	lower := strings.ToLower(v)
+	for _, ext := range refAudioExts {
+		if strings.HasSuffix(lower, ext) {
+			return "", v
+		}
+	}
+	return v, ""
+}
+
+func parseFloat32(v string, def float32) float32 {
+	if v == "" {
+		return def
+	}
+	f, err := strconv.ParseFloat(v, 32)
+	if err != nil {
+		return def
+	}
+	return float32(f)
+}
+
+func parseInt(v string, def int) int {
+	if v == "" {
+		return def
+	}
+	n, err := strconv.Atoi(v)
+	if err != nil {
+		return def
+	}
+	return n
+}
+
+func parseInt64(v string, def int64) int64 {
+	if v == "" {
+		return def
+	}
+	n, err := strconv.ParseInt(v, 10, 64)
+	if err != nil {
+		return def
+	}
+	return n
+}
+
+// parseSampling reads per-request sampling params from the TTSRequest params
+// map, applying qt defaults (matching qt_tts_default_params).
+func parseSampling(params map[string]string, defaultSeed int64) sampling {
+	s := sampling{temperature: 0.9, topK: 50, topP: 1.0, repPen: 1.05, maxNew: 2048, seed: defaultSeed}
+	if params == nil {
+		return s
+	}
+	s.temperature = parseFloat32(params["temperature"], s.temperature)
+	s.topK = parseInt(params["top_k"], s.topK)
+	s.topP = parseFloat32(params["top_p"], s.topP)
+	s.repPen = parseFloat32(params["repetition_penalty"], s.repPen)
+	s.maxNew = parseInt(params["max_new_tokens"], s.maxNew)
+	s.seed = parseInt64(params["seed"], s.seed)
+	return s
+}

backend/go/qwen3-tts-cpp/qwen3ttscpp_test.go

@@ -1,173 +1,136 @@
 package main
 
 import (
-	"context"
-	"os"
-	"os/exec"
-	"path/filepath"
+	"bytes"
+	"encoding/binary"
 	"testing"
-	"time"
 
-	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
-	"google.golang.org/grpc"
-	"google.golang.org/grpc/credentials/insecure"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
 )
 
-const (
-	testAddr    = "localhost:50051"
-	startupWait = 5 * time.Second
-)
-
-func skipIfNoModel(t *testing.T) string {
-	t.Helper()
-	modelDir := os.Getenv("QWEN3TTS_MODEL_DIR")
-	if modelDir == "" {
-		t.Skip("QWEN3TTS_MODEL_DIR not set, skipping test (set to directory with GGUF models)")
-	}
-	if _, err := os.Stat(filepath.Join(modelDir, "qwen3-tts-0.6b-f16.gguf")); os.IsNotExist(err) {
-		t.Skipf("TTS model file not found in %s, skipping", modelDir)
-	}
-	if _, err := os.Stat(filepath.Join(modelDir, "qwen3-tts-tokenizer-f16.gguf")); os.IsNotExist(err) {
-		t.Skipf("Tokenizer model file not found in %s, skipping", modelDir)
-	}
-	return modelDir
+func TestQwen3TtsCpp(t *testing.T) {
+	RegisterFailHandler(Fail)
+	RunSpecs(t, "qwen3-tts-cpp suite")
 }
 
-func startServer(t *testing.T) *exec.Cmd {
-	t.Helper()
-	binary := os.Getenv("QWEN3TTS_BINARY")
-	if binary == "" {
-		binary = "./qwen3-tts-cpp"
-	}
-	if _, err := os.Stat(binary); os.IsNotExist(err) {
-		t.Skipf("Backend binary not found at %s, skipping", binary)
-	}
-	cmd := exec.Command(binary, "--addr", testAddr)
-	cmd.Stdout = os.Stderr
-	cmd.Stderr = os.Stderr
-	if err := cmd.Start(); err != nil {
-		t.Fatalf("Failed to start server: %v", err)
-	}
-	time.Sleep(startupWait)
-	return cmd
-}
-
-func stopServer(cmd *exec.Cmd) {
-	if cmd != nil && cmd.Process != nil {
-		cmd.Process.Kill()
-		cmd.Wait()
-	}
-}
-
-func dialGRPC(t *testing.T) *grpc.ClientConn {
-	t.Helper()
-	conn, err := grpc.Dial(testAddr,
-		grpc.WithTransportCredentials(insecure.NewCredentials()),
-		grpc.WithDefaultCallOptions(
-			grpc.MaxCallRecvMsgSize(50*1024*1024),
-			grpc.MaxCallSendMsgSize(50*1024*1024),
-		),
+var _ = Describe("normalizeLanguage", func() {
+	DescribeTable("maps caller language to qwentts language names",
+		func(in, want string) {
+			Expect(normalizeLanguage(in)).To(Equal(want))
+		},
+		Entry("empty stays empty", "", ""),
+		Entry("auto maps to empty", "auto", ""),
+		Entry("english full name", "English", "english"),
+		Entry("english code", "en", "english"),
+		Entry("locale suffix stripped", "en-US", "english"),
+		Entry("underscore locale", "zh_CN", "chinese"),
+		Entry("mandarin alias", "mandarin", "chinese"),
+		Entry("japanese already full", "japanese", "japanese"),
+		Entry("unknown passes through normalized", "xx", "xx"),
 	)
-	if err != nil {
-		t.Fatalf("Failed to dial gRPC: %v", err)
-	}
-	return conn
-}
+})
 
-func TestServerHealth(t *testing.T) {
-	cmd := startServer(t)
-	defer stopServer(cmd)
-
-	conn := dialGRPC(t)
-	defer conn.Close()
-
-	client := pb.NewBackendClient(conn)
-	resp, err := client.Health(context.Background(), &pb.HealthMessage{})
-	if err != nil {
-		t.Fatalf("Health check failed: %v", err)
-	}
-	if string(resp.Message) != "OK" {
-		t.Fatalf("Expected OK, got %s", string(resp.Message))
-	}
-}
-
-func TestLoadModel(t *testing.T) {
-	modelDir := skipIfNoModel(t)
-	cmd := startServer(t)
-	defer stopServer(cmd)
-
-	conn := dialGRPC(t)
-	defer conn.Close()
-
-	client := pb.NewBackendClient(conn)
-
-	resp, err := client.LoadModel(context.Background(), &pb.ModelOptions{
-		ModelFile: modelDir,
-		Threads:   4,
+var _ = Describe("resolveVoice", func() {
+	It("treats a bare token as a named speaker", func() {
+		sp, ref := resolveVoice("serena")
+		Expect(sp).To(Equal("serena"))
+		Expect(ref).To(BeEmpty())
 	})
-	if err != nil {
-		t.Fatalf("LoadModel failed: %v", err)
-	}
-	if !resp.Success {
-		t.Fatalf("LoadModel returned failure: %s", resp.Message)
-	}
-}
-
-func TestTTS(t *testing.T) {
-	modelDir := skipIfNoModel(t)
-
-	tmpDir, err := os.MkdirTemp("", "qwen3tts-test")
-	if err != nil {
-		t.Fatal(err)
-	}
-	t.Cleanup(func() { os.RemoveAll(tmpDir) })
-
-	outputFile := filepath.Join(tmpDir, "output.wav")
-
-	cmd := startServer(t)
-	defer stopServer(cmd)
-
-	conn := dialGRPC(t)
-	defer conn.Close()
-
-	client := pb.NewBackendClient(conn)
-
-	// Load models
-	loadResp, err := client.LoadModel(context.Background(), &pb.ModelOptions{
-		ModelFile: modelDir,
-		Threads:   4,
+	It("treats an audio path as a clone reference (case-insensitive ext)", func() {
+		sp, ref := resolveVoice("/x/ref.WAV")
+		Expect(sp).To(BeEmpty())
+		Expect(ref).To(Equal("/x/ref.WAV"))
 	})
-	if err != nil {
-		t.Fatalf("LoadModel failed: %v", err)
-	}
-	if !loadResp.Success {
-		t.Fatalf("LoadModel returned failure: %s", loadResp.Message)
-	}
-
-	// Synthesize speech
-	language := "en"
-	_, err = client.TTS(context.Background(), &pb.TTSRequest{
-		Text:     "Hello, this is a test of the Qwen3 text to speech system.",
-		Dst:      outputFile,
-		Language: &language,
+	It("recognizes mp3/flac/ogg/m4a", func() {
+		for _, p := range []string{"a.mp3", "b.flac", "c.ogg", "d.m4a"} {
+			sp, ref := resolveVoice(p)
+			Expect(sp).To(BeEmpty())
+			Expect(ref).To(Equal(p))
+		}
 	})
-	if err != nil {
-		t.Fatalf("TTS failed: %v", err)
-	}
+	It("returns empty for empty input", func() {
+		sp, ref := resolveVoice("  ")
+		Expect(sp).To(BeEmpty())
+		Expect(ref).To(BeEmpty())
+	})
+})
 
-	// Verify output file exists and has content
-	info, err := os.Stat(outputFile)
-	if os.IsNotExist(err) {
-		t.Fatal("Output audio file was not created")
-	}
-	if err != nil {
-		t.Fatalf("Failed to stat output file: %v", err)
-	}
+var _ = Describe("parseOptions", func() {
+	It("extracts codec, use_fa, clamp_fp16, seed", func() {
+		o := parseOptions([]string{
+			"tokenizer:tok.gguf", "use_fa:false", "clamp_fp16:true",
+			"seed:7", "unknown:ignored",
+		})
+		Expect(o.codecPath).To(Equal("tok.gguf"))
+		Expect(o.useFA).To(BeFalse())
+		Expect(o.clampFP16).To(BeTrue())
+		Expect(o.seed).To(Equal(int64(7)))
+	})
+	It("accepts codec: as an alias for tokenizer:", func() {
+		Expect(parseOptions([]string{"codec:c.gguf"}).codecPath).To(Equal("c.gguf"))
+	})
+	It("defaults use_fa true and seed -1", func() {
+		o := parseOptions(nil)
+		Expect(o.useFA).To(BeTrue())
+		Expect(o.seed).To(Equal(int64(-1)))
+	})
+})
 
-	t.Logf("Output file size: %d bytes", info.Size())
+var _ = Describe("parseSampling", func() {
+	It("applies qt defaults when params are absent", func() {
+		s := parseSampling(nil, -1)
+		Expect(s.temperature).To(BeNumerically("~", 0.9, 1e-6))
+		Expect(s.topK).To(Equal(50))
+		Expect(s.topP).To(BeNumerically("~", 1.0, 1e-6))
+		Expect(s.repPen).To(BeNumerically("~", 1.05, 1e-6))
+		Expect(s.maxNew).To(Equal(2048))
+		Expect(s.seed).To(Equal(int64(-1)))
+	})
+	It("reads overrides and falls back to default seed", func() {
+		s := parseSampling(map[string]string{
+			"temperature": "0.5", "top_k": "10", "top_p": "0.8",
+			"repetition_penalty": "1.2", "max_new_tokens": "512",
+		}, 99)
+		Expect(s.temperature).To(BeNumerically("~", 0.5, 1e-6))
+		Expect(s.topK).To(Equal(10))
+		Expect(s.topP).To(BeNumerically("~", 0.8, 1e-6))
+		Expect(s.repPen).To(BeNumerically("~", 1.2, 1e-6))
+		Expect(s.maxNew).To(Equal(512))
+		Expect(s.seed).To(Equal(int64(99)))
+	})
+	It("reads an explicit seed override", func() {
+		Expect(parseSampling(map[string]string{"seed": "123"}, -1).seed).To(Equal(int64(123)))
+	})
+})
 
-	// WAV header is 44 bytes minimum; any real audio should be much larger
-	if info.Size() < 1000 {
-		t.Errorf("Output file too small (%d bytes), expected real audio data", info.Size())
-	}
-}
+var _ = Describe("wavHeader24k", func() {
+	It("emits a 44-byte streaming WAV header at 24 kHz mono 16-bit", func() {
+		h := wavHeader24k()
+		Expect(h).To(HaveLen(44))
+		Expect(string(h[0:4])).To(Equal("RIFF"))
+		Expect(string(h[8:12])).To(Equal("WAVE"))
+		Expect(string(h[12:16])).To(Equal("fmt "))
+		Expect(string(h[36:40])).To(Equal("data"))
+		var sampleRate uint32
+		Expect(binary.Read(bytes.NewReader(h[24:28]), binary.LittleEndian, &sampleRate)).To(Succeed())
+		Expect(sampleRate).To(Equal(uint32(24000)))
+	})
+})
+
+var _ = Describe("floatToPCM16LE", func() {
+	It("clamps and converts float PCM to little-endian int16 bytes", func() {
+		b := floatToPCM16LE([]float32{0, 1.0, -1.0, 2.0, -2.0})
+		Expect(b).To(HaveLen(10))
+		read := func(off int) int16 {
+			var v int16
+			_ = binary.Read(bytes.NewReader(b[off:off+2]), binary.LittleEndian, &v)
+			return v
+		}
+		Expect(read(0)).To(Equal(int16(0)))
+		Expect(read(2)).To(Equal(int16(32767)))
+		Expect(read(4)).To(Equal(int16(-32767)))
+		Expect(read(6)).To(Equal(int16(32767)))  // clamped from 2.0
+		Expect(read(8)).To(Equal(int16(-32767))) // clamped from -2.0
+	})
+})

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

@@ -2,51 +2,30 @@
 set -e
 
 CURDIR=$(dirname "$(realpath $0)")
+cd "$CURDIR"
 
 echo "Running qwen3-tts-cpp backend tests..."
 
-# The test requires:
-#   - QWEN3TTS_MODEL_DIR: path to directory containing GGUF model files
-#   - QWEN3TTS_BINARY: path to the qwen3-tts-cpp binary (defaults to ./qwen3-tts-cpp)
-#
-# Tests that require the model will be skipped if QWEN3TTS_MODEL_DIR is not set
-# or the directory does not contain the required model files.
-
-cd "$CURDIR"
-
-# Only auto-download models when QWEN3TTS_MODEL_DIR is not explicitly set
-if [ -z "$QWEN3TTS_MODEL_DIR" ]; then
-    export QWEN3TTS_MODEL_DIR="./qwen3-tts-models"
-
-    if [ ! -d "$QWEN3TTS_MODEL_DIR" ]; then
-        echo "Creating qwen3-tts-models directory for tests..."
-        mkdir -p "$QWEN3TTS_MODEL_DIR"
-        REPO_ID="endo5501/qwen3-tts.cpp"
-        echo "Repository: ${REPO_ID}"
-        echo ""
-
-        # Files to download (smallest model for testing)
-        FILES=(
-            "qwen3-tts-0.6b-f16.gguf"
-            "qwen3-tts-tokenizer-f16.gguf"
-        )
-
-        BASE_URL="https://huggingface.co/${REPO_ID}/resolve/main"
-
-        for file in "${FILES[@]}"; do
-            dest="${QWEN3TTS_MODEL_DIR}/${file}"
-            if [ -f "${dest}" ]; then
-                echo "  [skip] ${file} (already exists)"
-            else
-                echo "  [download] ${file}..."
-                curl -L -o "${dest}" "${BASE_URL}/${file}" --progress-bar
-                echo "  [done] ${file}"
-            fi
-        done
-    fi
+# Auto-download a small model pair only when QWEN3TTS_MODEL is not set.
+if [ -z "$QWEN3TTS_MODEL" ]; then
+    MODEL_DIR="./qwen3-tts-models"
+    mkdir -p "$MODEL_DIR"
+    REPO_ID="Serveurperso/Qwen3-TTS-GGUF"
+    BASE_URL="https://huggingface.co/${REPO_ID}/resolve/main"
+    FILES=( "qwen-talker-0.6b-base-Q4_K_M.gguf" "qwen-tokenizer-12hz-Q4_K_M.gguf" )
+    for file in "${FILES[@]}"; do
+        dest="${MODEL_DIR}/${file}"
+        if [ -f "${dest}" ]; then
+            echo "  [skip] ${file}"
+        else
+            echo "  [download] ${file}..."
+            curl -L -o "${dest}" "${BASE_URL}/${file}" --progress-bar
+        fi
+    done
+    export QWEN3TTS_MODEL="${MODEL_DIR}/qwen-talker-0.6b-base-Q4_K_M.gguf"
+    export QWEN3TTS_CODEC="${MODEL_DIR}/qwen-tokenizer-12hz-Q4_K_M.gguf"
 fi
 
-# Run Go tests
-go test -v -timeout 600s .
+go test -v -timeout 1200s .
 
 echo "All qwen3-tts-cpp tests passed."

backend/go/sherpa-onnx/backend.go

@@ -62,7 +62,7 @@ var (
 	shimVadConfigSetDebug                    func(uintptr, int32)
 	shimCreateVad                            func(uintptr, float32) uintptr
 
-	// TTS (offline, VITS) config
+	// TTS (offline, VITS/Piper and Kokoro) config
 	shimTtsConfigNew                  func() uintptr
 	shimTtsConfigFree                 func(uintptr)
 	shimTtsConfigSetVitsModel         func(uintptr, string)
@@ -76,6 +76,14 @@ var (
 	shimTtsConfigSetDebug             func(uintptr, int32)
 	shimTtsConfigSetProvider          func(uintptr, string)
 	shimTtsConfigSetMaxNumSentences   func(uintptr, int32)
+	shimTtsConfigSetKokoroModel       func(uintptr, string)
+	shimTtsConfigSetKokoroVoices      func(uintptr, string)
+	shimTtsConfigSetKokoroTokens      func(uintptr, string)
+	shimTtsConfigSetKokoroDataDir     func(uintptr, string)
+	shimTtsConfigSetKokoroDictDir     func(uintptr, string)
+	shimTtsConfigSetKokoroLexicon     func(uintptr, string)
+	shimTtsConfigSetKokoroLang        func(uintptr, string)
+	shimTtsConfigSetKokoroLengthScale func(uintptr, float32)
 	shimCreateOfflineTts              func(uintptr) uintptr
 
 	// Offline recognizer config
@@ -101,37 +109,37 @@ var (
 	shimCreateOfflineRecognizer                  func(uintptr) uintptr
 
 	// Online recognizer config
-	shimOnlineRecogConfigNew                      func() uintptr
-	shimOnlineRecogConfigFree                     func(uintptr)
-	shimOnlineRecogConfigSetTransducerEncoder     func(uintptr, string)
-	shimOnlineRecogConfigSetTransducerDecoder     func(uintptr, string)
-	shimOnlineRecogConfigSetTransducerJoiner      func(uintptr, string)
-	shimOnlineRecogConfigSetTokens                func(uintptr, string)
-	shimOnlineRecogConfigSetNumThreads            func(uintptr, int32)
-	shimOnlineRecogConfigSetDebug                 func(uintptr, int32)
-	shimOnlineRecogConfigSetProvider              func(uintptr, string)
-	shimOnlineRecogConfigSetFeatSampleRate        func(uintptr, int32)
-	shimOnlineRecogConfigSetFeatFeatureDim        func(uintptr, int32)
-	shimOnlineRecogConfigSetDecodingMethod        func(uintptr, string)
-	shimOnlineRecogConfigSetEnableEndpoint        func(uintptr, int32)
+	shimOnlineRecogConfigNew                        func() uintptr
+	shimOnlineRecogConfigFree                       func(uintptr)
+	shimOnlineRecogConfigSetTransducerEncoder       func(uintptr, string)
+	shimOnlineRecogConfigSetTransducerDecoder       func(uintptr, string)
+	shimOnlineRecogConfigSetTransducerJoiner        func(uintptr, string)
+	shimOnlineRecogConfigSetTokens                  func(uintptr, string)
+	shimOnlineRecogConfigSetNumThreads              func(uintptr, int32)
+	shimOnlineRecogConfigSetDebug                   func(uintptr, int32)
+	shimOnlineRecogConfigSetProvider                func(uintptr, string)
+	shimOnlineRecogConfigSetFeatSampleRate          func(uintptr, int32)
+	shimOnlineRecogConfigSetFeatFeatureDim          func(uintptr, int32)
+	shimOnlineRecogConfigSetDecodingMethod          func(uintptr, string)
+	shimOnlineRecogConfigSetEnableEndpoint          func(uintptr, int32)
 	shimOnlineRecogConfigSetRule1MinTrailingSilence func(uintptr, float32)
 	shimOnlineRecogConfigSetRule2MinTrailingSilence func(uintptr, float32)
 	shimOnlineRecogConfigSetRule3MinUtteranceLength func(uintptr, float32)
-	shimCreateOnlineRecognizer                    func(uintptr) uintptr
+	shimCreateOnlineRecognizer                      func(uintptr) uintptr
 
 	// Result accessors. Pointer returns use unsafe.Pointer so Go's
 	// vet checker doesn't flag them — the returned memory is C-owned,
 	// not subject to Go GC motion.
-	shimWaveSampleRate            func(uintptr) int32
-	shimWaveNumSamples            func(uintptr) int32
-	shimWaveSamples               func(uintptr) unsafe.Pointer
-	shimOfflineResultText         func(uintptr) unsafe.Pointer
-	shimOnlineResultText          func(uintptr) unsafe.Pointer
-	shimGeneratedAudioSampleRate  func(uintptr) int32
-	shimGeneratedAudioN           func(uintptr) int32
-	shimGeneratedAudioSamples     func(uintptr) unsafe.Pointer
-	shimSpeechSegmentStart        func(uintptr) int32
-	shimSpeechSegmentN            func(uintptr) int32
+	shimWaveSampleRate           func(uintptr) int32
+	shimWaveNumSamples           func(uintptr) int32
+	shimWaveSamples              func(uintptr) unsafe.Pointer
+	shimOfflineResultText        func(uintptr) unsafe.Pointer
+	shimOnlineResultText         func(uintptr) unsafe.Pointer
+	shimGeneratedAudioSampleRate func(uintptr) int32
+	shimGeneratedAudioN          func(uintptr) int32
+	shimGeneratedAudioSamples    func(uintptr) unsafe.Pointer
+	shimSpeechSegmentStart       func(uintptr) int32
+	shimSpeechSegmentN           func(uintptr) int32
 
 	// TTS streaming callback trampoline
 	shimTtsGenerateWithCallback func(tts uintptr, text string, sid int32, speed float32, cb uintptr, ud uintptr) uintptr
@@ -161,13 +169,13 @@ var (
 // pointer returned by the shim or `unsafe.Pointer(&slice[0])` from Go.
 var (
 	// VAD
-	sherpaVadAcceptWaveform       func(vad uintptr, samples unsafe.Pointer, n int32)
-	sherpaVadReset                func(vad uintptr)
-	sherpaVadFlush                func(vad uintptr)
-	sherpaVadEmpty                func(vad uintptr) int32
-	sherpaVadFront                func(vad uintptr) uintptr
-	sherpaVadPop                  func(vad uintptr)
-	sherpaDestroySpeechSegment    func(seg uintptr)
+	sherpaVadAcceptWaveform    func(vad uintptr, samples unsafe.Pointer, n int32)
+	sherpaVadReset             func(vad uintptr)
+	sherpaVadFlush             func(vad uintptr)
+	sherpaVadEmpty             func(vad uintptr) int32
+	sherpaVadFront             func(vad uintptr) uintptr
+	sherpaVadPop               func(vad uintptr)
+	sherpaDestroySpeechSegment func(seg uintptr)
 
 	// Wave IO
 	sherpaReadWave  func(filename string) uintptr
@@ -175,11 +183,11 @@ var (
 	sherpaWriteWave func(samples unsafe.Pointer, n int32, sampleRate int32, filename string) int32
 
 	// Offline ASR
-	sherpaCreateOfflineStream           func(rec uintptr) uintptr
-	sherpaDestroyOfflineStream          func(stream uintptr)
-	sherpaAcceptWaveformOffline         func(stream uintptr, sr int32, samples unsafe.Pointer, n int32)
-	sherpaDecodeOfflineStream           func(rec uintptr, stream uintptr)
-	sherpaGetOfflineStreamResult        func(stream uintptr) uintptr
+	sherpaCreateOfflineStream            func(rec uintptr) uintptr
+	sherpaDestroyOfflineStream           func(stream uintptr)
+	sherpaAcceptWaveformOffline          func(stream uintptr, sr int32, samples unsafe.Pointer, n int32)
+	sherpaDecodeOfflineStream            func(rec uintptr, stream uintptr)
+	sherpaGetOfflineStreamResult         func(stream uintptr) uintptr
 	sherpaDestroyOfflineRecognizerResult func(result uintptr)
 
 	// Online ASR
@@ -195,21 +203,21 @@ var (
 	sherpaOnlineStreamInputFinished     func(stream uintptr)
 
 	// TTS
-	sherpaOfflineTtsGenerate             func(tts uintptr, text string, sid int32, speed float32) uintptr
+	sherpaOfflineTtsGenerate              func(tts uintptr, text string, sid int32, speed float32) uintptr
 	sherpaDestroyOfflineTtsGeneratedAudio func(audio uintptr)
-	sherpaOfflineTtsSampleRate           func(tts uintptr) int32
+	sherpaOfflineTtsSampleRate            func(tts uintptr) int32
 
 	// Offline speaker diarization. Result handle owns the segment-array
 	// pointer returned by ResultSortByStartTime; destroy the segment
 	// array first, then the result, then (at backend Free()) the diarizer.
-	sherpaDestroyOfflineSpeakerDiarization                 func(sd uintptr)
-	sherpaOfflineSpeakerDiarizationGetSampleRate           func(sd uintptr) int32
-	sherpaOfflineSpeakerDiarizationProcess                 func(sd uintptr, samples unsafe.Pointer, n int32) uintptr
-	sherpaOfflineSpeakerDiarizationResultGetNumSegments    func(result uintptr) int32
-	sherpaOfflineSpeakerDiarizationResultGetNumSpeakers    func(result uintptr) int32
-	sherpaOfflineSpeakerDiarizationResultSortByStartTime   func(result uintptr) uintptr
-	sherpaOfflineSpeakerDiarizationDestroySegment          func(segs uintptr)
-	sherpaDestroyOfflineSpeakerDiarizationResult           func(result uintptr)
+	sherpaDestroyOfflineSpeakerDiarization               func(sd uintptr)
+	sherpaOfflineSpeakerDiarizationGetSampleRate         func(sd uintptr) int32
+	sherpaOfflineSpeakerDiarizationProcess               func(sd uintptr, samples unsafe.Pointer, n int32) uintptr
+	sherpaOfflineSpeakerDiarizationResultGetNumSegments  func(result uintptr) int32
+	sherpaOfflineSpeakerDiarizationResultGetNumSpeakers  func(result uintptr) int32
+	sherpaOfflineSpeakerDiarizationResultSortByStartTime func(result uintptr) uintptr
+	sherpaOfflineSpeakerDiarizationDestroySegment        func(segs uintptr)
+	sherpaDestroyOfflineSpeakerDiarizationResult         func(result uintptr)
 )
 
 var (
@@ -278,6 +286,14 @@ func loadSherpaLibsOnce() error {
 		{&shimTtsConfigSetDebug, "sherpa_shim_tts_config_set_debug"},
 		{&shimTtsConfigSetProvider, "sherpa_shim_tts_config_set_provider"},
 		{&shimTtsConfigSetMaxNumSentences, "sherpa_shim_tts_config_set_max_num_sentences"},
+		{&shimTtsConfigSetKokoroModel, "sherpa_shim_tts_config_set_kokoro_model"},
+		{&shimTtsConfigSetKokoroVoices, "sherpa_shim_tts_config_set_kokoro_voices"},
+		{&shimTtsConfigSetKokoroTokens, "sherpa_shim_tts_config_set_kokoro_tokens"},
+		{&shimTtsConfigSetKokoroDataDir, "sherpa_shim_tts_config_set_kokoro_data_dir"},
+		{&shimTtsConfigSetKokoroDictDir, "sherpa_shim_tts_config_set_kokoro_dict_dir"},
+		{&shimTtsConfigSetKokoroLexicon, "sherpa_shim_tts_config_set_kokoro_lexicon"},
+		{&shimTtsConfigSetKokoroLang, "sherpa_shim_tts_config_set_kokoro_lang"},
+		{&shimTtsConfigSetKokoroLengthScale, "sherpa_shim_tts_config_set_kokoro_length_scale"},
 		{&shimCreateOfflineTts, "sherpa_shim_create_offline_tts"},
 
 		{&shimOfflineRecogConfigNew, "sherpa_shim_offline_recog_config_new"},
@@ -688,21 +704,14 @@ func (s *SherpaBackend) loadTTS(opts *pb.ModelOptions) error {
 	cfg := shimTtsConfigNew()
 	defer shimTtsConfigFree(cfg)
 
-	shimTtsConfigSetVitsModel(cfg, modelFile)
-
-	if tokensPath := filepath.Join(modelDir, "tokens.txt"); fileExists(tokensPath) {
-		shimTtsConfigSetVitsTokens(cfg, tokensPath)
+	// Kokoro models ship a voices style file alongside the ONNX, whereas
+	// VITS/Piper voices do not. That presence is what tells the two model
+	// families apart, since both arrive as a plain *.onnx in modelDir.
+	if isKokoroModel(modelDir) {
+		s.configureKokoroTTS(cfg, opts, modelFile, modelDir)
+	} else {
+		s.configureVitsTTS(cfg, opts, modelFile, modelDir)
 	}
-	if lexiconPath := filepath.Join(modelDir, "lexicon.txt"); fileExists(lexiconPath) {
-		shimTtsConfigSetVitsLexicon(cfg, lexiconPath)
-	}
-	if dataDir := filepath.Join(modelDir, "espeak-ng-data"); dirExists(dataDir) {
-		shimTtsConfigSetVitsDataDir(cfg, dataDir)
-	}
-
-	shimTtsConfigSetVitsNoiseScale(cfg, findOptionFloat(opts, optionTtsNoiseScale, 0.667))
-	shimTtsConfigSetVitsNoiseScaleW(cfg, findOptionFloat(opts, optionTtsNoiseScaleW, 0.8))
-	shimTtsConfigSetVitsLengthScale(cfg, findOptionFloat(opts, optionTtsLengthScale, 1.0))
 
 	threads := int32(1)
 	if opts.Threads != 0 {
@@ -723,6 +732,80 @@ func (s *SherpaBackend) loadTTS(opts *pb.ModelOptions) error {
 	return nil
 }
 
+// kokoroVoicesFile is the speaker-style bank that ships with Kokoro models and
+// is absent from VITS/Piper voices; its presence is how loadTTS tells them apart.
+const kokoroVoicesFile = "voices.bin"
+
+// isKokoroModel reports whether modelDir holds a Kokoro model (a voices file
+// next to the ONNX) rather than a VITS/Piper single-speaker model.
+func isKokoroModel(modelDir string) bool {
+	return fileExists(filepath.Join(modelDir, kokoroVoicesFile))
+}
+
+// configureVitsTTS wires a VITS/Piper single-speaker model into cfg: the ONNX
+// plus the optional tokens, lexicon and espeak-ng-data found beside it.
+func (s *SherpaBackend) configureVitsTTS(cfg uintptr, opts *pb.ModelOptions, modelFile, modelDir string) {
+	shimTtsConfigSetVitsModel(cfg, modelFile)
+
+	if tokensPath := filepath.Join(modelDir, "tokens.txt"); fileExists(tokensPath) {
+		shimTtsConfigSetVitsTokens(cfg, tokensPath)
+	}
+	if lexiconPath := filepath.Join(modelDir, "lexicon.txt"); fileExists(lexiconPath) {
+		shimTtsConfigSetVitsLexicon(cfg, lexiconPath)
+	}
+	if dataDir := filepath.Join(modelDir, "espeak-ng-data"); dirExists(dataDir) {
+		shimTtsConfigSetVitsDataDir(cfg, dataDir)
+	}
+
+	shimTtsConfigSetVitsNoiseScale(cfg, findOptionFloat(opts, optionTtsNoiseScale, 0.667))
+	shimTtsConfigSetVitsNoiseScaleW(cfg, findOptionFloat(opts, optionTtsNoiseScaleW, 0.8))
+	shimTtsConfigSetVitsLengthScale(cfg, findOptionFloat(opts, optionTtsLengthScale, 1.0))
+}
+
+// configureKokoroTTS wires a Kokoro model into cfg: the ONNX, its voices bank,
+// tokens, and the optional espeak-ng-data / jieba dict / lexicon assets the
+// multi-lingual packs ship. A language hint comes from the `language=` option.
+func (s *SherpaBackend) configureKokoroTTS(cfg uintptr, opts *pb.ModelOptions, modelFile, modelDir string) {
+	shimTtsConfigSetKokoroModel(cfg, modelFile)
+	shimTtsConfigSetKokoroVoices(cfg, filepath.Join(modelDir, kokoroVoicesFile))
+
+	if tokensPath := filepath.Join(modelDir, "tokens.txt"); fileExists(tokensPath) {
+		shimTtsConfigSetKokoroTokens(cfg, tokensPath)
+	}
+	if dataDir := filepath.Join(modelDir, "espeak-ng-data"); dirExists(dataDir) {
+		shimTtsConfigSetKokoroDataDir(cfg, dataDir)
+	}
+	if dictDir := filepath.Join(modelDir, "dict"); dirExists(dictDir) {
+		shimTtsConfigSetKokoroDictDir(cfg, dictDir)
+	}
+
+	// Multi-lingual Kokoro ships per-language lexicons; the C API takes them as
+	// a single comma-separated list. US and GB English overlap almost entirely,
+	// so pass only one (US preferred) to avoid tens of thousands of "duplicated
+	// word" warnings at load; non-English lexicons (e.g. zh) are additive.
+	var lexicons []string
+	addLexicon := func(name string) {
+		if p := filepath.Join(modelDir, name); fileExists(p) {
+			lexicons = append(lexicons, p)
+		}
+	}
+	if fileExists(filepath.Join(modelDir, "lexicon-us-en.txt")) {
+		addLexicon("lexicon-us-en.txt")
+	} else {
+		addLexicon("lexicon-gb-en.txt")
+	}
+	addLexicon("lexicon-zh.txt")
+	addLexicon("lexicon.txt")
+	if len(lexicons) > 0 {
+		shimTtsConfigSetKokoroLexicon(cfg, strings.Join(lexicons, ","))
+	}
+
+	if lang := findOptionValue(opts, optionLanguage, ""); lang != "" {
+		shimTtsConfigSetKokoroLang(cfg, lang)
+	}
+	shimTtsConfigSetKokoroLengthScale(cfg, findOptionFloat(opts, optionTtsLengthScale, 1.0))
+}
+
 func fileExists(p string) bool {
 	info, err := os.Stat(p)
 	return err == nil && !info.IsDir()
@@ -1252,7 +1335,7 @@ type ttsStreamState struct {
 var (
 	ttsStates      sync.Map // uint64 → *ttsStreamState
 	ttsNextID      atomic.Uint64
-	ttsCallbackPtr uintptr  // purego.NewCallback return; registered in loadSherpaLibs
+	ttsCallbackPtr uintptr // purego.NewCallback return; registered in loadSherpaLibs
 )
 
 // ttsStreamCallback is invoked by sherpa-onnx for each PCM chunk VITS

backend/go/sherpa-onnx/backend_test.go

@@ -124,6 +124,20 @@ var _ = Describe("Sherpa-ONNX", func() {
 			Entry("empty", "", false),
 			Entry("other", "other", false),
 		)
+
+		It("isKokoroModel detects a voices file beside the ONNX", func() {
+			dir, err := os.MkdirTemp("", "sherpa-kokoro-*")
+			Expect(err).NotTo(HaveOccurred())
+			defer func() { _ = os.RemoveAll(dir) }()
+
+			// A bare VITS/Piper directory (ONNX only) is not Kokoro.
+			Expect(os.WriteFile(filepath.Join(dir, "model.onnx"), []byte("x"), 0o600)).To(Succeed())
+			Expect(isKokoroModel(dir)).To(BeFalse())
+
+			// Adding the Kokoro voices bank flips detection on.
+			Expect(os.WriteFile(filepath.Join(dir, kokoroVoicesFile), []byte("x"), 0o600)).To(Succeed())
+			Expect(isKokoroModel(dir)).To(BeTrue())
+		})
 	})
 
 	Context("option parsing", func() {

backend/go/sherpa-onnx/csrc/shim.c

@@ -79,6 +79,13 @@ void sherpa_shim_tts_config_free(void *h) {
     free((char *)c->model.vits.tokens);
     free((char *)c->model.vits.lexicon);
     free((char *)c->model.vits.data_dir);
+    free((char *)c->model.kokoro.model);
+    free((char *)c->model.kokoro.voices);
+    free((char *)c->model.kokoro.tokens);
+    free((char *)c->model.kokoro.data_dir);
+    free((char *)c->model.kokoro.dict_dir);
+    free((char *)c->model.kokoro.lexicon);
+    free((char *)c->model.kokoro.lang);
     free((char *)c->model.provider);
     free(c);
 }
@@ -117,6 +124,34 @@ void sherpa_shim_tts_config_set_max_num_sentences(void *h, int32_t v) {
     ((SherpaOnnxOfflineTtsConfig *)h)->max_num_sentences = v;
 }
 
+// Kokoro multi-speaker / multi-lingual TTS. Distinct ONNX + a voices style
+// file (voices.bin) instead of VITS' single-speaker graph; espeak-ng-data,
+// lexicon and a language hint are optional refinements.
+void sherpa_shim_tts_config_set_kokoro_model(void *h, const char *v) {
+    shim_set_str(&((SherpaOnnxOfflineTtsConfig *)h)->model.kokoro.model, v);
+}
+void sherpa_shim_tts_config_set_kokoro_voices(void *h, const char *v) {
+    shim_set_str(&((SherpaOnnxOfflineTtsConfig *)h)->model.kokoro.voices, v);
+}
+void sherpa_shim_tts_config_set_kokoro_tokens(void *h, const char *v) {
+    shim_set_str(&((SherpaOnnxOfflineTtsConfig *)h)->model.kokoro.tokens, v);
+}
+void sherpa_shim_tts_config_set_kokoro_data_dir(void *h, const char *v) {
+    shim_set_str(&((SherpaOnnxOfflineTtsConfig *)h)->model.kokoro.data_dir, v);
+}
+void sherpa_shim_tts_config_set_kokoro_dict_dir(void *h, const char *v) {
+    shim_set_str(&((SherpaOnnxOfflineTtsConfig *)h)->model.kokoro.dict_dir, v);
+}
+void sherpa_shim_tts_config_set_kokoro_lexicon(void *h, const char *v) {
+    shim_set_str(&((SherpaOnnxOfflineTtsConfig *)h)->model.kokoro.lexicon, v);
+}
+void sherpa_shim_tts_config_set_kokoro_lang(void *h, const char *v) {
+    shim_set_str(&((SherpaOnnxOfflineTtsConfig *)h)->model.kokoro.lang, v);
+}
+void sherpa_shim_tts_config_set_kokoro_length_scale(void *h, float v) {
+    ((SherpaOnnxOfflineTtsConfig *)h)->model.kokoro.length_scale = v;
+}
+
 void *sherpa_shim_create_offline_tts(void *h) {
     return (void *)SherpaOnnxCreateOfflineTts(
         (const SherpaOnnxOfflineTtsConfig *)h);

backend/go/sherpa-onnx/csrc/shim.h

@@ -37,7 +37,7 @@ void  sherpa_shim_vad_config_set_provider(void *cfg, const char *v);
 void  sherpa_shim_vad_config_set_debug(void *cfg, int32_t v);
 void *sherpa_shim_create_vad(void *cfg, float buffer_size_seconds);
 
-// --- Offline TTS config (VITS path — the only TTS family the backend uses) ---
+// --- Offline TTS config (VITS/Piper and Kokoro model families) ---
 void *sherpa_shim_tts_config_new(void);
 void  sherpa_shim_tts_config_free(void *cfg);
 void  sherpa_shim_tts_config_set_vits_model(void *cfg, const char *v);
@@ -51,6 +51,14 @@ void  sherpa_shim_tts_config_set_num_threads(void *cfg, int32_t v);
 void  sherpa_shim_tts_config_set_debug(void *cfg, int32_t v);
 void  sherpa_shim_tts_config_set_provider(void *cfg, const char *v);
 void  sherpa_shim_tts_config_set_max_num_sentences(void *cfg, int32_t v);
+void  sherpa_shim_tts_config_set_kokoro_model(void *cfg, const char *v);
+void  sherpa_shim_tts_config_set_kokoro_voices(void *cfg, const char *v);
+void  sherpa_shim_tts_config_set_kokoro_tokens(void *cfg, const char *v);
+void  sherpa_shim_tts_config_set_kokoro_data_dir(void *cfg, const char *v);
+void  sherpa_shim_tts_config_set_kokoro_dict_dir(void *cfg, const char *v);
+void  sherpa_shim_tts_config_set_kokoro_lexicon(void *cfg, const char *v);
+void  sherpa_shim_tts_config_set_kokoro_lang(void *cfg, const char *v);
+void  sherpa_shim_tts_config_set_kokoro_length_scale(void *cfg, float v);
 void *sherpa_shim_create_offline_tts(void *cfg);
 
 // --- Offline recognizer config (Whisper / Paraformer / SenseVoice / Omnilingual) ---