fix: mlx backend fails to compile; publish containers to correct ghcr org

.github/workflows/build-container.yml

@@ -43,7 +43,7 @@ jobs:
         id: meta
         uses: docker/metadata-action@v5
         with:
-          images: ghcr.io/lemonade-sdk/lemonade/build-environment
+          images: ghcr.io/${{ github.repository }}/build-environment
           tags: |
             type=raw,value=ubuntu${{ matrix.ubuntu_version }}
             type=raw,value=latest,enable=${{ matrix.ubuntu_version == '26.04' }}

README.md

@@ -136,7 +136,7 @@ Lemonade supports multiple inference engines for LLM, speech, TTS, and image gen
   </thead>
   <tbody>
     <tr>
-      <td rowspan="9"><strong>Text generation</strong></td>
+      <td rowspan="12"><strong>Text generation</strong></td>
       <td rowspan="6"><code>llamacpp</code></td>
       <td><code>system</code></td>
       <td><code>x86_64</code>/ARM64 CPU, GPU</td>
@@ -185,6 +185,22 @@ Lemonade supports multiple inference engines for LLM, speech, TTS, and image gen
       <td>Strix Halo iGPU (gfx1151)</td>
       <td>Linux</td>
     </tr>
+    <tr>
+      <td rowspan="3"><code>mlx-engine</code> (experimental)</td>
+      <td><code>metal</code></td>
+      <td>Apple Silicon Metal GPU</td>
+      <td>macOS</td>
+    </tr>
+    <tr>
+      <td><code>rocm</code></td>
+      <td>AMD ROCm GPUs (RDNA3/RDNA4)</td>
+      <td>Linux</td>
+    </tr>
+    <tr>
+      <td><code>cpu</code></td>
+      <td>CPU fallback</td>
+      <td>Linux, macOS</td>
+    </tr>
     <tr>
       <td rowspan="6"><strong>Speech-to-text</strong></td>
       <td rowspan="5"><code>whispercpp</code></td>

docs/assets/models.js

@@ -8,6 +8,7 @@ const RECIPE_PRIORITY = [
   'flm',
   'kokoro',
   'llamacpp',
+  'mlx-engine',
   'moonshine',
   'onnxruntime',
   'openmoss',
@@ -30,7 +31,8 @@ const RECIPE_DISPLAY_NAMES = {
   acestep: 'ACE-Step',
   onnxruntime: 'ONNX Runtime',
   trellis: 'TRELLIS.2',
-  openmoss: 'OpenMOSS TTS'
+  openmoss: 'OpenMOSS TTS',
+  'mlx-engine': 'MLX Engine (Apple Silicon)'
 };
 /* END GENERATED: models-js-recipes */
 

docs/dev/backends-reference.md

@@ -13,6 +13,7 @@ the generator instead. Prose outside the markers is preserved. -->
 | `flm` | FastFlowLM NPU | no | yes | npu |
 | `kokoro` | Kokoro | no | no | cpu, metal |
 | `llamacpp` | Llama.cpp GPU | yes | yes | cpu, cuda, metal, rocm, system, vulkan |
+| `mlx-engine` | MLX Engine | yes | yes | cpu, metal, rocm |
 | `moonshine` | Moonshine | no | no | cpu |
 | `onnxruntime` | ONNX Runtime | no | no | cpu |
 | `openmoss` | OpenMOSS TTS | yes | no | cuda, rocm, vulkan |
@@ -41,6 +42,9 @@ the generator instead. Prose outside the markers is preserved. -->
 | `llamacpp` | vulkan | linux, windows | amd_gpu; cpu (arm64, x86_64) |
 | `llamacpp` | rocm | linux, windows | amd_gpu (gfx103X, gfx110X, gfx1150, gfx1151, gfx1152, gfx120X, gfx942, gfx950) |
 | `llamacpp` | cpu | linux, windows | cpu (arm64, x86_64) |
+| `mlx-engine` | metal | macos | metal |
+| `mlx-engine` | rocm | linux | amd_gpu (gfx110X, gfx1150, gfx1151, gfx120X) |
+| `mlx-engine` | cpu | linux, macos | cpu (arm64, x86_64) |
 | `moonshine` | cpu | windows | cpu (x86_64) |
 | `moonshine` | cpu | linux | cpu (arm64, x86_64) |
 | `moonshine` | cpu | macos | cpu (arm64) |
@@ -102,6 +106,14 @@ the generator instead. Prose outside the markers is preserved. -->
 | `llamacpp_device` | `--llamacpp-device` | DEVICES | "" | Comma-separated list of accelerator devices to use (e.g. Vulkan0) |
 | `llamacpp_args` | `--llamacpp-args` | ARGS | "" | Custom arguments to pass to llama-server |
 
+#### `mlx-engine` — MLX Engine
+
+| Option | CLI flag | Type | Default | Description |
+|--------|----------|------|---------|-------------|
+| `ctx_size` | `--ctx-size` | SIZE | -1 | Context size for the model |
+| `mlx_backend` | `--mlx-backend` | BACKEND | "" | MLX backend to use (metal, rocm, cpu) |
+| `mlx_args` | `--mlx-args` | ARGS | "" | Extra arguments passed to lemon-mlx-engine |
+
 #### `moonshine` — Moonshine
 
 | Option | CLI flag | Type | Default | Description |
@@ -268,6 +280,13 @@ the generator instead. Prose outside the markers is preserved. -->
 | `nomic-embed-text-v1-GGUF` | 0.0781 | embeddings |
 | `nomic-embed-text-v2-moe-GGUF` | 0.51 | embeddings |
 
+#### `mlx-engine` — MLX Engine (2 models)
+
+| Model | Size (GB) | Labels |
+|-------|-----------|--------|
+| `Qwen3-0.6B-MLX` | 0.42 | reasoning |
+| `Qwen3-4B-MLX` | 2.3 | reasoning |
+
 #### `moonshine` — Moonshine (3 models)
 
 | Model | Size (GB) | Labels |

docs/guide/cli.md

@@ -408,6 +408,14 @@ The following options are available depending on the recipe being used:
 | Option | Description | Default |
 |--------|-------------|---------|
 | `--openmoss BACKEND` | OpenMOSS TTS backend to use | Auto-detected |
+
+#### MLX Engine (`mlx-engine` recipe)
+
+| Option | Description | Default |
+|--------|-------------|---------|
+| `--ctx-size SIZE` | Context size for the model | auto |
+| `--mlx-backend BACKEND` | MLX backend to use (metal, rocm, cpu) | Auto-detected |
+| `--mlx-args ARGS` | Extra arguments passed to lemon-mlx-engine | `""` |
 <!-- END GENERATED: cli-recipe-options -->
 **Notes:**
 - Unspecified options will use the backend's default values

docs/guide/configuration/README.md

@@ -72,6 +72,9 @@ Values set in the user's `config.json` always take precedence over these seeded
   },
   "log_level": "info",
   "max_loaded_models": 1,
+  "mlx-engine": {
+    "backend": "auto"
+  },
   "models_dir": "auto",
   "moonshine": {
     "args": "",

docs/guide/configuration/custom-models.md

@@ -87,7 +87,7 @@ Supported registration flags:
 |------|-------------|
 | `--source SOURCE` | Remote registry for every checkpoint in this model: `huggingface` (default) or `modelscope`. |
 | `--checkpoint TYPE CHECKPOINT` | Add a checkpoint entry. Repeat for multi-file models such as `main` + `mmproj` or `main` + `vae`. |
-| `--recipe RECIPE` | Recipe to associate with the new `user.*` model. Common values: <!-- BEGIN GENERATED: recipe-values -->`llamacpp`, `whispercpp`, `moonshine`, `kokoro`, `sd-cpp`, `flm`, `ryzenai-llm`, `vllm`, `thinksound`, `acestep`, `onnxruntime`, `trellis`, `openmoss`, `collection.omni`<!-- END GENERATED: recipe-values -->. |
+| `--recipe RECIPE` | Recipe to associate with the new `user.*` model. Common values: <!-- BEGIN GENERATED: recipe-values -->`llamacpp`, `whispercpp`, `moonshine`, `kokoro`, `sd-cpp`, `flm`, `ryzenai-llm`, `vllm`, `thinksound`, `acestep`, `onnxruntime`, `trellis`, `openmoss`, `mlx-engine`, `collection.omni`<!-- END GENERATED: recipe-values -->. |
 | `--label LABEL` | Add a label to the new model. Repeatable. Valid labels include `coding`, `embeddings`, `hot`, `mtp`, `reasoning`, `reranking`, `tool-calling`, `vision`. |
 | `--components MODEL [MODEL ...]` | Components for an omni collection (see below). Use with `--recipe collection.omni`. |
 

src/cpp/resources/defaults.json

@@ -37,6 +37,9 @@
   },
   "log_level": "info",
   "max_loaded_models": 1,
+  "mlx-engine": {
+    "backend": "auto"
+  },
   "models_dir": "auto",
   "moonshine": {
     "args": "",

src/cpp/server/backends/mlx/mlx_server.cpp

@@ -1,4 +1,5 @@
 #include "lemon/backends/mlx/mlx_server.h"
+#include "lemon/backends/mlx/mlx.h"
 #include "lemon/backends/backend_ops.h"
 #include "lemon/backends/backend_utils.h"
 #include "lemon/backend_manager.h"
@@ -112,7 +113,7 @@ void MlxServer::load(const std::string& model_name,
     device_type_ = (mlx_backend == "cpu") ? DEVICE_CPU : DEVICE_GPU;
 
     // Install mlx-engine binary if needed.
-    backend_manager_->install_backend(SPEC.recipe, mlx_backend);
+    backend_manager_->install_backend(mlx::spec()->recipe, mlx_backend);
 
     // MLX identifies models by HuggingFace repo-id or a local directory path.
     // The ModelManager resolves local paths when available; fall back to the
@@ -130,7 +131,7 @@ void MlxServer::load(const std::string& model_name,
 
     port_ = choose_port();
 
-    std::string executable = BackendUtils::get_backend_binary_path(SPEC, mlx_backend);
+    std::string executable = BackendUtils::get_backend_binary_path(*mlx::spec(), mlx_backend);
 
     std::vector<std::string> args;
     // Positional model argument — pre-load mode.
@@ -222,8 +223,6 @@ json MlxServer::responses(const json& request) {
     );
 }
 
-} // namespace backends
-
 namespace {
 class MlxOps : public BackendOps {
 public: