Merge branch 'main' into router/2385-collection.router_recipe_dispatch

Resolve conflicts from #2519 (parser/store) and #2548 (classifier services)
landing on main after this branch forked:

- routing_classifier_services.{h,cpp}: keep main's reviewed parse/validation
  body (validate_score_range, try_extract_chat_text) and re-add this branch's
  build_route_context + collect_text_from_content/content_has_image helpers,
  which #2385 dispatch needs.
- routing_policy_store.{h,cpp}, test_routing_policy_store.cpp,
  test_routing_classifier_services.cpp: take main (strict superset — C++20
  TODO, case-insensitive .json, added tests).
- model_manager.cpp USER_DEFINED_MODEL_PROPS: keep "version" (router policies
  carry a required root version) on top of main's "routing".
- server.cpp download-skip: is_model_collection_recipe (omni OR router) — a
  collection.router has no checkpoint and must skip the generic HF download;
  main's is_omni_collection_recipe would wrongly route it through HF.

lemond builds; routing parser/engine/deterministic/classifier-services/
collection-validation unit tests pass.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: ramkrishna2910

.github/workflows/cpp_server_build_test_release.yml

@@ -1148,6 +1148,16 @@ jobs:
             extra_args: ""
             backends: "cpu rocm vulkan"
             runner: [Windows, rocm, vulkan, lemon-prod]
+          - name: audio-gen-thinksound
+            script: server_audio_generation.py
+            extra_args: "--wrapped-server thinksound"
+            backends: "vulkan rocm"
+            runner: [Windows, rocm, vulkan, lemon-prod]
+          - name: audio-gen-acestep
+            script: server_audio_generation.py
+            extra_args: "--wrapped-server acestep"
+            backends: "vulkan rocm"
+            runner: [Windows, rocm, vulkan, lemon-prod]
           - name: text-to-speech
             script: server_tts.py
             extra_args: ""
@@ -1431,6 +1441,16 @@ jobs:
             extra_args: ""
             backends: "cpu rocm vulkan"
             runner: [Linux, vulkan, rocm, lemon-prod]
+          - name: audio-gen-thinksound
+            script: server_audio_generation.py
+            extra_args: "--wrapped-server thinksound"
+            backends: "vulkan rocm"
+            runner: [Linux, vulkan, rocm, lemon-prod]
+          - name: audio-gen-acestep
+            script: server_audio_generation.py
+            extra_args: "--wrapped-server acestep"
+            backends: "vulkan rocm"
+            runner: [Linux, vulkan, rocm, lemon-prod]
           - name: whisper
             script: server_whisper.py
             extra_args: "--wrapped-server whispercpp"

CMakeLists.txt

@@ -64,6 +64,8 @@ set(LEMON_BACKENDS
     "ryzenai-llm|ryzenai"
     "vllm|vllm"
     "cloud|cloud"
+    "thinksound|thinksound"
+    "acestep|acestep"
 )
 
 # ============================================================

docs/api/lemonade.md

@@ -16,6 +16,7 @@ We have designed a set of Lemonade-specific endpoints to enable client applicati
 | `POST` | [`/v1/delete`](#post-v1delete) | Delete a model |
 | `POST` | [`/v1/load`](#post-v1load) | Load a model |
 | `POST` | [`/v1/unload`](#post-v1unload) | Unload a model |
+| `POST` | [`/v1/audio/generations`](#post-v1audiogenerations) | Generate audio (music or sound effects) from a text prompt |
 | `GET` | [`/v1/models/{id}/files`](#get-v1modelsidfiles) | List resolved local file metadata for one model |
 | `GET` | [`/v1/health`](#get-v1health) | Check server status, such as models loaded |
 | `GET` | [`/v1/stats`](#get-v1stats) | Performance statistics from the last request |
@@ -687,6 +688,45 @@ In case of an error, the status will be `error` and the message will contain the
 
 
 
+## `POST /v1/audio/generations`
+<sub>![Status](https://img.shields.io/badge/status-fully_available-green)</sub>
+
+Audio Generation API. You provide a text prompt and receive a generated audio clip. The loaded model decides the kind of audio: music with ACE-Step models (e.g. `ACE-Step-Music`), sound effects with ThinkSound models (e.g. `ThinkSound-SFX`).
+
+This endpoint is not part of the OpenAI API (OpenAI's audio endpoints cover speech and transcription only), so it is a Lemonade-specific extension.
+
+> **Performance:** generation runs on the GPU (Vulkan, ROCm, or CUDA) and takes from seconds (short sound effects) to minutes (full-length music) depending on duration and hardware.
+
+### Parameters
+
+| Parameter | Required | Description |
+|-----------|----------|-------------|
+| `model` | Yes | The audio-generation model to use (e.g., `ThinkSound-SFX`, `ACE-Step-Music`). |
+| `prompt` | Yes | Text description of the music or sound effect to generate. |
+| `duration` | No | Length of the clip in seconds. Defaults to the backend's native default. |
+| `steps` | No | Number of inference steps. Lower is faster, higher can improve quality. |
+| `cfg` | No | Classifier-free guidance strength (ThinkSound only). |
+| `seed` | No | Random seed for reproducibility. |
+| `response_format` | No | Output encoding. Only formats the backend natively produces are accepted (currently `wav`); other values are rejected with `400 Bad Request`. Default: `wav`. |
+
+### Response
+
+On success the raw audio bytes are returned with the matching content type (`audio/wav`). On failure the response is JSON with an `error` object: `400` for invalid requests, `404` for unknown models, `500` when the backend reports an error, and `502` when the backend produces no output.
+
+### Example request
+
+```bash
+curl -X POST http://localhost:13305/v1/audio/generations \
+  -H "Content-Type: application/json" \
+  -d '{
+        "model": "ThinkSound-SFX",
+        "prompt": "glass shattering on a stone floor",
+        "duration": 5,
+        "seed": 42
+      }' \
+  --output clip.wav
+```
+
 ## `GET /v1/health`
 <sub>![Status](https://img.shields.io/badge/status-fully_available-green)</sub>
 

docs/assets/models.js

@@ -4,12 +4,14 @@ const RAW_BASE = 'https://raw.githubusercontent.com/lemonade-sdk/lemonade';
 
 /* BEGIN GENERATED: models-js-recipes */
 const RECIPE_PRIORITY = [
+  'acestep',
   'flm',
   'kokoro',
   'llamacpp',
   'moonshine',
   'ryzenai-llm',
   'sd-cpp',
+  'thinksound',
   'vllm',
   'whispercpp'
 ];
@@ -20,7 +22,9 @@ const RECIPE_DISPLAY_NAMES = {
   'sd-cpp': 'stable-diffusion.cpp',
   flm: 'FastFlowLM NPU',
   'ryzenai-llm': 'Ryzen AI SW NPU',
-  vllm: 'vLLM ROCm (experimental)'
+  vllm: 'vLLM ROCm (experimental)',
+  thinksound: 'ThinkSound',
+  acestep: 'ACE-Step'
 };
 /* END GENERATED: models-js-recipes */
 

docs/dev/backends-reference.md

@@ -9,12 +9,14 @@ the generator instead. Prose outside the markers is preserved. -->
 <!-- BEGIN GENERATED: backends-overview -->
 | Recipe | Name | Selectable backend | Uses ctx_size | Backends |
 |--------|------|--------------------|---------------|----------|
+| `acestep` | ACE-Step | yes | no | cuda, rocm, vulkan |
 | `flm` | FastFlowLM NPU | no | yes | npu |
 | `kokoro` | Kokoro | no | no | cpu, metal |
 | `llamacpp` | Llama.cpp GPU | yes | yes | cpu, cuda, metal, rocm, system, vulkan |
 | `moonshine` | Moonshine | no | no | cpu |
 | `ryzenai-llm` | Ryzen AI LLM | no | yes | npu |
 | `sd-cpp` | StableDiffusion.cpp | yes | no | cpu, cuda, metal, rocm, vulkan |
+| `thinksound` | ThinkSound | yes | no | cuda, rocm, vulkan |
 | `vllm` | vLLM ROCm (experimental) | yes | yes | rocm |
 | `whispercpp` | Whisper.cpp | yes | no | cpu, metal, npu, rocm, vulkan |
 <!-- END GENERATED: backends-overview -->
@@ -24,6 +26,9 @@ the generator instead. Prose outside the markers is preserved. -->
 <!-- BEGIN GENERATED: backends-matrix -->
 | Recipe | Backend | OS | Device families |
 |--------|---------|----|-----------------|
+| `acestep` | rocm | linux, windows | amd_gpu |
+| `acestep` | cuda | linux, windows | nvidia_gpu |
+| `acestep` | vulkan | linux, windows | amd_gpu; cpu (x86_64); nvidia_gpu |
 | `flm` | npu | linux, windows | amd_npu (XDNA2) |
 | `kokoro` | cpu | linux, windows | cpu (x86_64) |
 | `kokoro` | metal | macos | metal |
@@ -42,6 +47,9 @@ the generator instead. Prose outside the markers is preserved. -->
 | `sd-cpp` | vulkan | linux, windows | amd_gpu; cpu (x86_64); nvidia_gpu |
 | `sd-cpp` | cpu | linux, windows | cpu (x86_64) |
 | `sd-cpp` | metal | macos | metal |
+| `thinksound` | rocm | linux, windows | amd_gpu |
+| `thinksound` | cuda | linux, windows | nvidia_gpu |
+| `thinksound` | vulkan | linux, windows | amd_gpu; cpu (x86_64); nvidia_gpu |
 | `vllm` | rocm | linux | amd_gpu (gfx110X, gfx1150, gfx1151, gfx120X) |
 | `whispercpp` | npu | windows | amd_npu (XDNA2) |
 | `whispercpp` | rocm | linux, windows | amd_gpu (gfx110X, gfx1150, gfx1151, gfx120X) |
@@ -53,6 +61,12 @@ the generator instead. Prose outside the markers is preserved. -->
 ## Recipe options
 
 <!-- BEGIN GENERATED: backend-options -->
+#### `acestep` — ACE-Step
+
+| Option | CLI flag | Type | Default | Description |
+|--------|----------|------|---------|-------------|
+| `acestep_backend` | `--acestep` | BACKEND | "" | ACE-Step backend to use |
+
 #### `llamacpp` — Llama.cpp GPU
 
 | Option | CLI flag | Type | Default | Description |
@@ -81,6 +95,12 @@ the generator instead. Prose outside the markers is preserved. -->
 | `sampling_method` | — | ARGS | "" | Sampling method |
 | `flow_shift` | — | SIZE | 0.0 | Flow shift |
 
+#### `thinksound` — ThinkSound
+
+| Option | CLI flag | Type | Default | Description |
+|--------|----------|------|---------|-------------|
+| `thinksound_backend` | `--thinksound` | BACKEND | "" | ThinkSound backend to use |
+
 #### `vllm` — vLLM ROCm (experimental)
 
 | Option | CLI flag | Type | Default | Description |
@@ -100,6 +120,12 @@ the generator instead. Prose outside the markers is preserved. -->
 ## Models
 
 <!-- BEGIN GENERATED: backend-models -->
+#### `acestep` — ACE-Step (1 models)
+
+| Model | Size (GB) | Labels |
+|-------|-----------|--------|
+| `ACE-Step-Music` | 10.5 | audio-generation |
+
 #### `collection.omni` — collection.omni (5 models)
 
 | Model | Size (GB) | Labels |
@@ -307,6 +333,12 @@ the generator instead. Prose outside the markers is preserved. -->
 | `SDXL-Turbo` | 6.94 | image |
 | `Z-Image-Turbo` | 20.7 | image |
 
+#### `thinksound` — ThinkSound (1 models)
+
+| Model | Size (GB) | Labels |
+|-------|-----------|--------|
+| `ThinkSound-SFX` | 6.4 | audio-generation |
+
 #### `vllm` — vLLM ROCm (experimental) (7 models)
 
 | Model | Size (GB) | Labels |

docs/dev/contribute.md

@@ -41,9 +41,36 @@ Each contribution needs to:
     - Reviewers: make sure to check the code *before* allowing CI to run!
 3. Sustain the overall code quality and standards of the rest of the repo.
 
+### Making Pull Requests Easy to Review
+
+Maintainers try to review contributions quickly, but review time is limited. The easiest pull requests to review are small, focused, and easy to verify.
+
+Before opening a PR, please make sure that:
+
+1. The PR solves one clear problem.
+2. The changes are limited to what is necessary.
+3. There are no unrelated local changes included.
+4. The PR description explains what changed and why.
+5. The PR describes how the change was tested, including any platforms that could not be tested locally.
+6. Larger refactoring is separated from bug fixes or feature changes.
+
+Some areas of the codebase are more sensitive than they may appear. Small changes can sometimes have large side effects, especially in recently reworked, complex, or platform-specific code. In these cases, reviewers may ask you to reduce the scope, add tests, or follow a specific implementation approach. If you need help testing on a platform you do not have access to, please mention it in the PR and ask for support in the Lemonade Discord.
+
+The fastest way to build trust as a new contributor is to submit small, clear, well-tested PRs that are easy to review and easy to verify.
+
 ### AI Policy
 
-Contributors are encouraged to use AI to code their project. However, please review your AI's code yourself before asking another human to review it.
+Contributors are welcome to use AI tools while working on Lemonade. However, contributors remain fully responsible for the code they submit.
+
+If you use AI-assisted coding, please make sure that you:
+
+1. Understand the generated code before submitting it.
+2. Review the full diff yourself.
+3. Remove unrelated or unnecessary changes.
+4. Test the result locally where possible.
+5. Keep the PR especially small and focused.
+
+AI can make it easy to generate large changes, but large or unfocused changes are harder for maintainers to review and are less likely to be accepted quickly.
 
 Reviewers are expected to use tools like Claude Code's `/review` to save time by analyzing code quality and edge cases. If AI tools generate PR comments, please clearly identify which comments are AI-generated and which are authored by you.
 

docs/guide/cli.md

@@ -376,6 +376,18 @@ The following options are available depending on the recipe being used:
 | `--ctx-size SIZE` | Context size for the model | auto |
 | `--vllm BACKEND` | vLLM backend to use | Auto-detected |
 | `--vllm-args ARGS` | Custom arguments to pass to vllm-server | `""` |
+
+#### ThinkSound (`thinksound` recipe)
+
+| Option | Description | Default |
+|--------|-------------|---------|
+| `--thinksound BACKEND` | ThinkSound backend to use | Auto-detected |
+
+#### ACE-Step (`acestep` recipe)
+
+| Option | Description | Default |
+|--------|-------------|---------|
+| `--acestep BACKEND` | ACE-Step backend to use | Auto-detected |
 <!-- END GENERATED: cli-recipe-options -->
 **Notes:**
 - Unspecified options will use the backend's default values

docs/guide/configuration/README.md

@@ -35,6 +35,12 @@ Values set in the user's `config.json` always take precedence over these seeded
 ```json
 {
   "_generated": "GENERATED by docs/tools/gen_backend_boilerplate.py -- do not hand-edit per-recipe sections (they come from each backend's descriptor config_defaults()). Global keys are hand-maintained in this file. Regenerate and verify with that script; CI --check fails on drift.",
+  "acestep": {
+    "backend": "auto",
+    "cuda_bin": "builtin",
+    "rocm_bin": "builtin",
+    "vulkan_bin": "builtin"
+  },
   "cloud_providers": [],
   "config_version": 2,
   "ctx_size": -1,
@@ -112,6 +118,12 @@ Values set in the user's `config.json` always take precedence over these seeded
       "send_batch_size": 100
     }
   },
+  "thinksound": {
+    "backend": "auto",
+    "cuda_bin": "builtin",
+    "rocm_bin": "builtin",
+    "vulkan_bin": "builtin"
+  },
   "vllm": {
     "args": "",
     "backend": "auto"

docs/guide/configuration/custom-models.md

@@ -71,7 +71,7 @@ Supported registration flags:
 | Flag | Description |
 |------|-------------|
 | `--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`, `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`, `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/app/src/renderer/ChatWindow.tsx

@@ -13,6 +13,7 @@ import RerankingPanel from './components/panels/RerankingPanel';
 import TranscriptionPanel from './components/panels/TranscriptionPanel';
 import ImageGenerationPanel from './components/panels/ImageGenerationPanel';
 import TTSPanel from './components/panels/TTSPanel';
+import AudioGenerationPanel from './components/panels/AudioGenerationPanel';
 import LLMChatPanel from './components/panels/LLMChatPanel';
 import { RefreshIcon } from './components/Icons';
 import { isCollectionModel, getCollectionComponents } from './utils/collectionModels';
@@ -42,7 +43,7 @@ const ChatWindow: React.FC<ChatWindowProps> = ({ isVisible, width }) => {
   const [addModelInitialValues, setAddModelInitialValues] = useState<AddModelInitialValues | undefined>(undefined);
   const addModelFromJSONRef = useRef<HTMLInputElement>(null);
 
-  type ModelType = 'llm' | 'embedding' | 'reranking' | 'transcription' | 'image' | 'tts';
+  type ModelType = 'llm' | 'embedding' | 'reranking' | 'transcription' | 'image' | 'tts' | 'audio';
 
   const modelType = useMemo((): ModelType => {
     if (!selectedModel) return 'llm';
@@ -60,6 +61,7 @@ const ChatWindow: React.FC<ChatWindowProps> = ({ isVisible, width }) => {
     if (info.labels?.includes('transcription')) return 'transcription';
     if (info.labels?.includes('image')) return 'image';
     if (info.labels?.includes('tts')) return 'tts';
+    if (info.labels?.includes('audio-generation')) return 'audio';
     return 'llm';
   }, [selectedModel, modelsData]);
 
@@ -276,6 +278,7 @@ const ChatWindow: React.FC<ChatWindowProps> = ({ isVisible, width }) => {
     : activeModelType === 'transcription' ? 'Lemonade Transcriber'
     : activeModelType === 'image' ? 'Lemonade Image Generator'
     : activeModelType === 'tts' ? 'Lemonade Text to Speech'
+    : activeModelType === 'audio' ? 'Lemonade Audio Generator'
     : 'LLM Chat';
 
   const sharedProps = {
@@ -313,6 +316,7 @@ const ChatWindow: React.FC<ChatWindowProps> = ({ isVisible, width }) => {
       {activeModelType === 'transcription' && <TranscriptionPanel key={resetKey} {...sharedProps} />}
       {activeModelType === 'image' && <ImageGenerationPanel key={resetKey} {...sharedProps} />}
       {activeModelType === 'tts' && <TTSPanel key={resetKey} {...sharedProps} />}
+      {activeModelType === 'audio' && <AudioGenerationPanel key={resetKey} {...sharedProps} />}
       {activeModelType === 'llm' && (
         <LLMChatPanel
           key={resetKey}