JnyRoadandcapy-ai-staging[bot]
19f8f51408
Unloading a TTS/Whisper/LLM model on the MLX backend only dropped the
Python reference (`del self.model`). MLX keeps freed array buffers in
its own allocator pool for reuse instead of returning them to the OS,
so the process's memory footprint never actually shrank after unload
on Apple Silicon (the default backend there) until the process exited.
Add empty_mlx_cache() (backend/backends/base.py), wrapping
mx.clear_cache(), and call it from the three MLX unload_model()
implementations: MLXTTSBackend, MLXSTTBackend, MLXQwenLLMBackend.
Separately, the voice-clone prompt cache (backend/utils/cache.py) is a
process-lifetime dict populated by create_voice_prompt() across every
TTS engine, but nothing ever cleared it on model unload — only the
unrelated /tasks/clear-cache endpoint touched it. Add
clear_voice_prompt_memory_cache() (memory only, disk cache untouched
so a later generation still reloads the prompt instead of recomputing
it) and wire it into every TTS unload path (services/tts.py and the
qwen_custom_voice / generic branches of unload_model_by_config).
Whisper and the LLM backends never produce voice prompts, so their
unload paths are left alone.
Testing:
- New unit tests: backend/tests/test_mlx_unload_clears_cache.py,
backend/tests/test_voice_prompt_cache_unload.py (8 tests, all pass).
- Verified end-to-end on Apple Silicon against real cached models
(Qwen TTS 1.7B, Whisper Turbo, Qwen3 0.6B): loaded each via the
running app, unloaded via the real /models/{name}/unload endpoint,
and confirmed via mx.get_cache_memory()/get_active_memory() that the
MLX allocator's cache drops to 0 on every cycle. Ran a real
voice-clone generation end to end and confirmed the in-memory prompt
cache goes from 1 entry to 0 on unload while the on-disk .prompt
file is left intact.
Voicebox Backend
FastAPI server powering voice cloning, speech generation, and audio processing. Runs locally as a Tauri sidecar or standalone via python -m backend.main.
Running
# Via justfile (recommended)
just dev:server
# Standalone
python -m backend.main --host 127.0.0.1 --port 17493
# With custom data directory
python -m backend.main --data-dir /path/to/data
The server auto-initializes the SQLite database on first startup. Models are downloaded from HuggingFace on first use.
Architecture
backend/
app.py # FastAPI app factory, CORS, lifecycle events
main.py # Entry point (imports app, runs uvicorn)
config.py # Data directory paths and configuration
models.py # Pydantic request/response schemas
server.py # Tauri sidecar launcher, parent-pid watchdog
routes/ # Thin HTTP handlers — validation, delegation, response formatting
services/ # Business logic, CRUD, orchestration
backends/ # TTS/STT engine implementations (MLX, PyTorch, etc.)
database/ # ORM models, session management, migrations, seed data
utils/ # Shared utilities (audio, effects, caching, progress tracking)
Request flow
HTTP request
-> routes/ (validate input, parse params)
-> services/ (business logic, database queries, orchestration)
-> backends/ (TTS/STT inference)
-> utils/ (audio processing, effects, caching)
Route handlers are intentionally thin. They validate input, delegate to a service function, and format the response. All business logic lives in services/.
Key modules
services/generation.py -- Single run_generation() function that handles all three generation modes (generate, retry, regenerate). Manages model loading, voice prompt creation, chunked inference, normalization, effects, and version persistence.
services/task_queue.py -- Serial generation queue. Ensures only one GPU inference runs at a time. Background tasks are tracked to prevent garbage collection.
backends/__init__.py -- Protocol definitions (TTSBackend, STTBackend), model config registry, and factory functions. Adding a new engine means implementing the protocol and registering a config entry.
backends/base.py -- Shared utilities used across all engine implementations: HuggingFace cache checks, device detection, voice prompt combination, progress tracking.
database/ -- SQLAlchemy ORM models with a re-exporting __init__.py for backward compatibility. Migrations run automatically on startup.
Backend selection
The server detects the best inference backend at startup:
| Platform | Backend | Acceleration |
|---|---|---|
| macOS (Apple Silicon) | MLX | Metal / Neural Engine |
| Windows / Linux (NVIDIA) | PyTorch | CUDA |
| Linux (AMD) | PyTorch | ROCm |
| Intel Arc | PyTorch | IPEX / XPU |
| Windows (any GPU) | PyTorch | DirectML |
| Any | PyTorch | CPU fallback |
Detection is handled by utils/platform_detect.py. Both backends implement the same TTSBackend protocol, so the API layer is engine-agnostic.
API
90 endpoints organized by domain. Full interactive documentation available at http://localhost:17493/docs when the server is running.
| Domain | Prefix | Description |
|---|---|---|
| Health | /, /health |
Server status, GPU info, filesystem checks |
| Profiles | /profiles |
Voice profile CRUD, samples, avatars, import/export |
| Channels | /channels |
Audio channel management and voice assignment |
| Generation | /generate |
TTS generation, retry, regenerate, status SSE |
| History | /history |
Generation history, search, favorites, export |
| Transcription | /transcribe |
Whisper-based audio-to-text |
| Stories | /stories |
Multi-track timeline editor, audio export |
| Effects | /effects |
Effect presets, preview, version management |
| Audio | /audio, /samples |
Audio file serving |
| Models | /models |
Load, unload, download, migrate, status |
| Tasks | /tasks, /cache |
Active task tracking, cache management |
| CUDA | /backend/cuda-* |
CUDA binary download and management |
Quick examples
# Generate speech
curl -X POST http://localhost:17493/generate \
-H "Content-Type: application/json" \
-d '{"text": "Hello world", "profile_id": "...", "language": "en"}'
# List profiles
curl http://localhost:17493/profiles
# Stream generation status (SSE)
curl http://localhost:17493/generate/{id}/status
Data directory
{data_dir}/
voicebox.db # SQLite database
profiles/{id}/ # Voice samples per profile
generations/ # Generated audio files
cache/ # Voice prompt cache (memory + disk)
backends/ # Downloaded CUDA binary (if applicable)
Default location is the OS-specific app data directory. Override with --data-dir or the VOICEBOX_DATA_DIR environment variable.
Code quality
Linting and formatting are enforced by ruff, configured in pyproject.toml. See STYLE_GUIDE.md for conventions.
just check-python # lint + format check
just fix-python # auto-fix lint issues + reformat
just test # run pytest
Dependencies
Runtime dependencies are in requirements.txt. macOS-only MLX dependencies are in requirements-mlx.txt. Dev tools (ruff, pytest) are installed automatically by just setup-python.