fix(backend): close model load races

Kokoro and LuxTTS load_model had no lock, so two concurrent requests
could both observe an unloaded model and double-load; they now use the
same double-checked asyncio.Lock pattern as the Chatterbox backends.
get_stt_backend gets the threading.Lock treatment the TTS and LLM
factories already had. Includes the ruff-era typing cleanup for
backends/__init__.

backend/backends/__init__.py

@@ -10,13 +10,14 @@ and a model config registry that eliminates per-engine dispatch maps.
 # import time, which wraps transformers' tokenizer load against the
 # unconditional HuggingFace metadata call that otherwise raises on
 # HF_HUB_OFFLINE=1 and on network failures.
-from ..utils import hf_offline_patch  # noqa: F401
-
 import threading
 from dataclasses import dataclass, field
-from typing import Protocol, Optional, Tuple, List
-from typing_extensions import runtime_checkable
+from typing import Protocol
+
 import numpy as np
+from typing_extensions import runtime_checkable
+
+from ..utils import hf_offline_patch
 
 DEFAULT_LLM_MAX_TOKENS = 512
 DEFAULT_LLM_TEMPERATURE = 0.7
@@ -76,7 +77,7 @@ class TTSBackend(Protocol):
         audio_path: str,
         reference_text: str,
         use_cache: bool = True,
-    ) -> Tuple[dict, bool]:
+    ) -> tuple[dict, bool]:
         """
         Create voice prompt from reference audio.
 
@@ -87,9 +88,9 @@ class TTSBackend(Protocol):
 
     async def combine_voice_prompts(
         self,
-        audio_paths: List[str],
-        reference_texts: List[str],
-    ) -> Tuple[np.ndarray, str]:
+        audio_paths: list[str],
+        reference_texts: list[str],
+    ) -> tuple[np.ndarray, str]:
         """
         Combine multiple voice prompts.
 
@@ -103,9 +104,9 @@ class TTSBackend(Protocol):
         text: str,
         voice_prompt: dict,
         language: str = "en",
-        seed: Optional[int] = None,
-        instruct: Optional[str] = None,
-    ) -> Tuple[np.ndarray, int]:
+        seed: int | None = None,
+        instruct: str | None = None,
+    ) -> tuple[np.ndarray, int]:
         """
         Generate audio from text.
 
@@ -143,8 +144,8 @@ class STTBackend(Protocol):
     async def transcribe(
         self,
         audio_path: str,
-        language: Optional[str] = None,
-        model_size: Optional[str] = None,
+        language: str | None = None,
+        model_size: str | None = None,
     ) -> str:
         """
         Transcribe audio to text.
@@ -174,11 +175,11 @@ class LLMBackend(Protocol):
     async def generate(
         self,
         prompt: str,
-        system: Optional[str] = None,
+        system: str | None = None,
         max_tokens: int = DEFAULT_LLM_MAX_TOKENS,
         temperature: float = DEFAULT_LLM_TEMPERATURE,
-        model_size: Optional[str] = None,
-        examples: Optional[list[tuple[str, str]]] = None,
+        model_size: str | None = None,
+        examples: list[tuple[str, str]] | None = None,
     ) -> str:
         """Run a single-turn chat completion and return the assistant reply.
 
@@ -198,10 +199,11 @@ class LLMBackend(Protocol):
 
 
 # Global backend instances
-_tts_backend: Optional[TTSBackend] = None
+_tts_backend: TTSBackend | None = None
 _tts_backends: dict[str, TTSBackend] = {}
 _tts_backends_lock = threading.Lock()
-_stt_backend: Optional[STTBackend] = None
+_stt_backend: STTBackend | None = None
+_stt_backend_lock = threading.Lock()
 _llm_backends: dict[str, LLMBackend] = {}
 _llm_backends_lock = threading.Lock()
 
@@ -488,7 +490,7 @@ def get_stt_model_configs() -> list[ModelConfig]:
 # Lookup helpers — these replace the if/elif chains in main.py
 
 
-def get_model_config(model_name: str) -> Optional[ModelConfig]:
+def get_model_config(model_name: str) -> ModelConfig | None:
     """Look up a model config by model_name."""
     for cfg in get_all_model_configs():
         if cfg.model_name == model_name:
@@ -563,8 +565,8 @@ async def unload_backend(backend) -> None:
 
 async def unload_model_by_config(config: ModelConfig) -> bool:
     """Unload a model given its config. Returns True if it was loaded, False otherwise."""
+    from ..services import llm as llm_service, transcribe, tts
     from . import get_tts_backend_for_engine
-    from ..services import tts, transcribe, llm as llm_service
 
     if config.engine == "whisper":
         whisper_model = transcribe.get_whisper_model()
@@ -607,8 +609,8 @@ async def unload_model_by_config(config: ModelConfig) -> bool:
 
 def check_model_loaded(config: ModelConfig) -> bool:
     """Check if a model is currently loaded."""
+    from ..services import llm as llm_service, transcribe, tts
     from . import get_tts_backend_for_engine
-    from ..services import tts, transcribe, llm as llm_service
 
     try:
         if config.engine == "whisper":
@@ -638,8 +640,8 @@ def check_model_loaded(config: ModelConfig) -> bool:
 
 def get_model_load_func(config: ModelConfig):
     """Return a callable that loads/downloads the model."""
+    from ..services import llm as llm_service, transcribe, tts
     from . import get_tts_backend_for_engine
-    from ..services import tts, transcribe, llm as llm_service
 
     if config.engine == "whisper":
         return lambda: transcribe.get_whisper_model().load_model(config.model_size)
@@ -738,7 +740,13 @@ def get_stt_backend() -> STTBackend:
     """
     global _stt_backend
 
-    if _stt_backend is None:
+    if _stt_backend is not None:
+        return _stt_backend
+
+    with _stt_backend_lock:
+        if _stt_backend is not None:
+            return _stt_backend
+
         backend_type = get_backend_type()
 
         if backend_type == "mlx":
@@ -750,7 +758,7 @@ def get_stt_backend() -> STTBackend:
 
             _stt_backend = PyTorchSTTBackend()
 
-    return _stt_backend
+        return _stt_backend
 
 
 def get_llm_backend() -> LLMBackend:

backend/backends/kokoro_backend.py

@@ -17,15 +17,12 @@ Languages supported (via misaki G2P):
 
 import asyncio
 import logging
-import os
-from typing import Optional
 
 import numpy as np
 
-from . import TTSBackend
 from .base import (
-    get_torch_device,
     combine_voice_prompts as _combine_voice_prompts,
+    get_torch_device,
     model_load_progress,
 )
 
@@ -122,8 +119,9 @@ class KokoroTTSBackend:
     def __init__(self):
         self._model = None
         self._pipelines: dict = {}  # lang_code -> KPipeline
-        self._device: Optional[str] = None
+        self._device: str | None = None
         self.model_size = "default"
+        self._model_load_lock = asyncio.Lock()
 
     def _get_device(self) -> str:
         """Select device. Kokoro supports CUDA and CPU. MPS needs fallback env var."""
@@ -157,7 +155,10 @@ class KokoroTTSBackend:
         """Load the Kokoro model."""
         if self._model is not None:
             return
-        await asyncio.to_thread(self._load_model_sync)
+        async with self._model_load_lock:
+            if self._model is not None:
+                return
+            await asyncio.to_thread(self._load_model_sync)
 
     def _load_model_sync(self):
         """Synchronous model loading."""
@@ -239,8 +240,8 @@ class KokoroTTSBackend:
         text: str,
         voice_prompt: dict,
         language: str = "en",
-        seed: Optional[int] = None,
-        instruct: Optional[str] = None,
+        seed: int | None = None,
+        instruct: str | None = None,
     ) -> tuple[np.ndarray, int]:
         """
         Generate audio from text using Kokoro.

backend/backends/luxtts_backend.py

@@ -7,20 +7,18 @@ Wraps the LuxTTS (ZipVoice) model for zero-shot voice cloning.
 
 import asyncio
 import logging
-from typing import Optional, Tuple
 
 import numpy as np
 
-from . import TTSBackend
+from ..utils.cache import cache_voice_prompt, get_cache_key, get_cached_voice_prompt
 from .base import (
-    is_model_cached,
-    get_torch_device,
-    empty_device_cache,
-    manual_seed,
     combine_voice_prompts as _combine_voice_prompts,
+    empty_device_cache,
+    get_torch_device,
+    is_model_cached,
+    manual_seed,
     model_load_progress,
 )
-from ..utils.cache import get_cache_key, get_cached_voice_prompt, cache_voice_prompt
 
 logger = logging.getLogger(__name__)
 
@@ -35,6 +33,7 @@ class LuxTTSBackend:
         self.model = None
         self.model_size = "default"  # LuxTTS has only one model size
         self._device = None
+        self._model_load_lock = asyncio.Lock()
 
     def _get_device(self) -> str:
         return get_torch_device(allow_mps=True, allow_xpu=True)
@@ -61,8 +60,10 @@ class LuxTTSBackend:
         """Load the LuxTTS model."""
         if self.model is not None:
             return
-
-        await asyncio.to_thread(self._load_model_sync)
+        async with self._model_load_lock:
+            if self.model is not None:
+                return
+            await asyncio.to_thread(self._load_model_sync)
 
     def _load_model_sync(self):
         model_name = "luxtts"
@@ -105,7 +106,7 @@ class LuxTTSBackend:
         audio_path: str,
         reference_text: str,
         use_cache: bool = True,
-    ) -> Tuple[dict, bool]:
+    ) -> tuple[dict, bool]:
         """
         Create voice prompt from reference audio.
 
@@ -145,9 +146,9 @@ class LuxTTSBackend:
         text: str,
         voice_prompt: dict,
         language: str = "en",
-        seed: Optional[int] = None,
-        instruct: Optional[str] = None,
-    ) -> Tuple[np.ndarray, int]:
+        seed: int | None = None,
+        instruct: str | None = None,
+    ) -> tuple[np.ndarray, int]:
         """
         Generate audio from text using LuxTTS.