refactor(generation): simplify post-generation cache cleanup

Initialise tts_model before the try so the finally can read its device
without a locals() probe, import empty_device_cache alongside the other
backend imports instead of inside a bare try/except, and flatten the MPS
branch in empty_device_cache.

backend/backends/base.py

@@ -219,9 +219,8 @@ def empty_device_cache(device: str) -> None:
         torch.cuda.empty_cache()
     elif device == "xpu" and hasattr(torch, "xpu"):
         torch.xpu.empty_cache()
-    elif device == "mps" and hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
-        if hasattr(torch.mps, "empty_cache"):
-            torch.mps.empty_cache()
+    elif device == "mps" and torch.backends.mps.is_available() and hasattr(torch.mps, "empty_cache"):
+        torch.mps.empty_cache()
 
 
 def manual_seed(seed: int, device: str) -> None:

backend/services/generation.py

@@ -54,11 +54,13 @@ async def run_generation(
         get_tts_backend_for_engine,
         load_engine_model,
     )
+    from ..backends.base import empty_device_cache
     from ..utils.chunked_tts import generate_chunked
     from ..utils.audio import has_tts_runaway, normalize_audio, save_audio, trim_tts_output
 
     task_manager = get_task_manager()
     bg_db = next(get_db())
+    tts_model = None
 
     try:
         tts_model = get_tts_backend_for_engine(engine)
@@ -156,12 +158,7 @@ async def run_generation(
     finally:
         task_manager.complete_generation(generation_id)
         bg_db.close()
-        try:
-            from ..backends.base import empty_device_cache
-            device = getattr(tts_model, "device", "cpu") if "tts_model" in locals() else "cpu"
-            empty_device_cache(device)
-        except Exception:
-            pass
+        empty_device_cache(getattr(tts_model, "device", "cpu"))
 
 
 def _notify_speak_end(generation_id: str, *, status: str) -> None:
@@ -286,11 +283,13 @@ async def generate_audio_sync(
         get_tts_backend_for_engine,
         load_engine_model,
     )
+    from ..backends.base import empty_device_cache
     from ..utils.chunked_tts import generate_chunked
     from ..utils.audio import has_tts_runaway, normalize_audio, trim_tts_output
     from . import tts
 
     bg_db = next(get_db())
+    tts_model = None
     try:
         tts_model = get_tts_backend_for_engine(engine)
         await load_engine_model(engine, model_size)
@@ -329,12 +328,7 @@ async def generate_audio_sync(
 
         return tts.audio_to_wav_bytes(audio, sample_rate)
     finally:
-        try:
-            from ..backends.base import empty_device_cache
-            device = getattr(tts_model, "device", "cpu") if "tts_model" in locals() else "cpu"
-            empty_device_cache(device)
-        except Exception:
-            pass
+        empty_device_cache(getattr(tts_model, "device", "cpu"))
 
 
 def _save_regenerate(