fix: address review feedback — race condition, GPU safety, task GC

- Add threading lock to get_tts_backend_for_engine() to prevent race
  condition where concurrent requests could create duplicate backend
  instances (double-checked locking pattern)
- Fix LuxTTS generate: call .detach().cpu() before .numpy() so it
  works on GPU/MPS devices, not just CPU
- Store background download tasks in a module-level set to prevent
  garbage collection before completion (asyncio.create_task fire-and-
  forget pattern)
- Deduplicate cache_key computation in LuxTTS create_voice_prompt
- Prefix unused sr variable with underscore

backend/backends/__init__.py

@@ -4,6 +4,7 @@ Backend abstraction layer for TTS and STT.
 Provides a unified interface for MLX and PyTorch backends.
 """
 
+import threading
 from typing import Protocol, Optional, Tuple, List
 from typing_extensions import runtime_checkable
 import numpy as np
@@ -113,6 +114,7 @@ class STTBackend(Protocol):
 # Global backend instances
 _tts_backend: Optional[TTSBackend] = None
 _tts_backends: dict[str, TTSBackend] = {}
+_tts_backends_lock = threading.Lock()
 _stt_backend: Optional[STTBackend] = None
 
 # Supported TTS engines
@@ -144,25 +146,32 @@ def get_tts_backend_for_engine(engine: str) -> TTSBackend:
     """
     global _tts_backends
     
+    # Fast path: check without lock
     if engine in _tts_backends:
         return _tts_backends[engine]
     
-    if engine == "qwen":
-        backend_type = get_backend_type()
-        if backend_type == "mlx":
-            from .mlx_backend import MLXTTSBackend
-            backend = MLXTTSBackend()
+    # Slow path: create with lock to avoid duplicate instantiation
+    with _tts_backends_lock:
+        # Double-check after acquiring lock
+        if engine in _tts_backends:
+            return _tts_backends[engine]
+        
+        if engine == "qwen":
+            backend_type = get_backend_type()
+            if backend_type == "mlx":
+                from .mlx_backend import MLXTTSBackend
+                backend = MLXTTSBackend()
+            else:
+                from .pytorch_backend import PyTorchTTSBackend
+                backend = PyTorchTTSBackend()
+        elif engine == "luxtts":
+            from .luxtts_backend import LuxTTSBackend
+            backend = LuxTTSBackend()
         else:
-            from .pytorch_backend import PyTorchTTSBackend
-            backend = PyTorchTTSBackend()
-    elif engine == "luxtts":
-        from .luxtts_backend import LuxTTSBackend
-        backend = LuxTTSBackend()
-    else:
-        raise ValueError(f"Unknown TTS engine: {engine}. Supported: {list(TTS_ENGINES.keys())}")
-    
-    _tts_backends[engine] = backend
-    return backend
+            raise ValueError(f"Unknown TTS engine: {engine}. Supported: {list(TTS_ENGINES.keys())}")
+        
+        _tts_backends[engine] = backend
+        return backend
 
 
 def get_stt_backend() -> STTBackend:

backend/backends/luxtts_backend.py

@@ -171,9 +171,10 @@ class LuxTTSBackend:
         """
         await self.load_model()
 
-        if use_cache:
-            # Include "luxtts" in the cache key so it doesn't collide with Qwen prompts
-            cache_key = "luxtts_" + get_cache_key(audio_path, reference_text)
+        # Compute cache key once for both lookup and storage
+        cache_key = ("luxtts_" + get_cache_key(audio_path, reference_text)) if use_cache else None
+
+        if cache_key:
             cached = get_cached_voice_prompt(cache_key)
             if cached is not None and isinstance(cached, dict):
                 return cached, True
@@ -187,8 +188,7 @@ class LuxTTSBackend:
 
         encoded = await asyncio.to_thread(_encode_sync)
 
-        if use_cache:
-            cache_key = "luxtts_" + get_cache_key(audio_path, reference_text)
+        if cache_key:
             cache_voice_prompt(cache_key, encoded)
 
         return encoded, False
@@ -206,7 +206,7 @@ class LuxTTSBackend:
         """
         combined_audio = []
         for path in audio_paths:
-            audio, sr = load_audio(path, sample_rate=24000)
+            audio, _sr = load_audio(path, sample_rate=24000)
             audio = normalize_audio(audio)
             combined_audio.append(audio)
 
@@ -257,8 +257,8 @@ class LuxTTSBackend:
                 return_smooth=False,  # 48kHz output
             )
 
-            # LuxTTS returns a tensor, convert to numpy
-            audio = wav.numpy().squeeze()
+            # LuxTTS returns a tensor (may be on GPU/MPS), move to CPU first
+            audio = wav.detach().cpu().numpy().squeeze()
             return audio, 48000
 
         return await asyncio.to_thread(_generate_sync)

backend/main.py

@@ -48,6 +48,18 @@ from .utils.tasks import get_task_manager
 from .utils.cache import clear_voice_prompt_cache
 from .platform_detect import get_backend_type
 
+# Keep references to fire-and-forget background tasks to prevent GC
+_background_tasks: set = set()
+
+
+def _create_background_task(coro) -> asyncio.Task:
+    """Create a background task and prevent it from being garbage collected."""
+    task = asyncio.create_task(coro)
+    _background_tasks.add(task)
+    task.add_done_callback(_background_tasks.discard)
+    return task
+
+
 app = FastAPI(
     title="voicebox API",
     description="Production-quality Qwen3-TTS voice cloning API",
@@ -622,7 +634,7 @@ async def generate_speech(
                         task_manager.error_download(model_name, str(e))
 
                 task_manager.start_download(model_name)
-                asyncio.create_task(download_model_background())
+                _create_background_task(download_model_background())
 
                 raise HTTPException(
                     status_code=202,
@@ -646,7 +658,7 @@ async def generate_speech(
                         task_manager.error_download(model_name, str(e))
 
                 task_manager.start_download(model_name)
-                asyncio.create_task(download_luxtts_background())
+                _create_background_task(download_luxtts_background())
 
                 raise HTTPException(
                     status_code=202,
@@ -986,7 +998,7 @@ async def transcribe_audio(
                     get_task_manager().error_download(progress_model_name, str(e))
 
             get_task_manager().start_download(progress_model_name)
-            asyncio.create_task(download_whisper_background())
+            _create_background_task(download_whisper_background())
 
             # Return 202 Accepted
             raise HTTPException(
@@ -1639,7 +1651,7 @@ async def trigger_model_download(request: models.ModelDownloadRequest):
     )
 
     # Start download in background task (don't await)
-    asyncio.create_task(download_in_background())
+    _create_background_task(download_in_background())
 
     # Return immediately - frontend should poll progress endpoint
     return {"message": f"Model {request.model_name} download started"}
@@ -1889,7 +1901,7 @@ async def download_cuda_backend():
             import logging
             logging.getLogger(__name__).error(f"CUDA download failed: {e}")
 
-    asyncio.create_task(_download())
+    _create_background_task(_download())
     return {"message": "CUDA backend download started", "progress_key": "cuda-backend"}