fix: voice cloning OOM on T4 - load 1.7B Base model on CPU

GPU VRAM is fully occupied by TTS 0.6B + Whisper + Q&A 3B.
Voice cloning (one-time operation) now runs on CPU to avoid OOM.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

voice_clone.py

@@ -109,22 +109,23 @@ def _select_attn_impl() -> str:
 
 
 def get_qwen_tts():
-    """Lazy-load Qwen3-TTS model. Thread-safe."""
+    """Lazy-load Qwen3-TTS Base 1.7B model for cloning. Uses CPU to avoid OOM."""
     global _qwen_tts_model
     if _qwen_tts_model is None:
         with _model_lock:
             if _qwen_tts_model is None:
                 from qwen_tts import Qwen3TTSModel
 
-                model_id = CUSTOM_VOICE_MODEL_ID if _MODEL_MODE == "custom_voice" else BASE_MODEL_ID
-                device = "cuda" if torch.cuda.is_available() else "cpu"
+                model_id = BASE_MODEL_ID
+                # Always use CPU for the 1.7B cloning model — GPU is occupied by TTS/ASR/Q&A
+                device = "cpu"
 
-                logger.info("Loading %s on %s...", model_id, device)
+                logger.info("Loading %s on %s (cloning model)...", model_id, device)
                 _qwen_tts_model = Qwen3TTSModel.from_pretrained(
                     model_id,
                     device_map=device,
                 )
-                logger.info("Qwen3-TTS loaded (%s mode).", _MODEL_MODE)
+                logger.info("Qwen3-TTS Base loaded on CPU.")
     return _qwen_tts_model