fix: remove custom dtype/attn params to fix CUDA assert on T4

The qwen_tts model has internal dtype handling. Passing float16 +
custom attention implementation causes CUDA asserts on T4 (Turing).
Use model defaults which handle GPU compatibility internally.

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

app.py

@@ -730,7 +730,20 @@ with gr.Blocks(title="MomsVoice", css=css_code) as demo:
         import soundfile as sf
 
         progress(0.1, desc="Extracting speaker embedding from recording...")
-        profile_id = create_voice_profile(recorder_data, voice_name=v_name.strip())
+        try:
+            profile_id = create_voice_profile(recorder_data, voice_name=v_name.strip())
+        except Exception as e:
+            import logging, traceback
+            logging.getLogger(__name__).exception("Voice cloning failed")
+            return (
+                f"""<div style="padding: 12px; background: #fef2f2; border: 1px solid #fecaca; color: #991b1b; border-radius: 12px; font-size: 12px; font-weight: 600;">
+                    Voice cloning failed: {html.escape(str(e))}
+                   </div>""",
+                gr.HTML(visible=False),
+                gr.Audio(visible=False),
+                gr.Button(visible=False),
+                None,
+            )
 
         progress(0.7, desc="Generating voice preview...")
         try:

voice_clone.py

@@ -109,7 +109,7 @@ def _select_attn_impl() -> str:
 
 
 def get_qwen_tts():
-    """Lazy-load Qwen3-TTS model with latency optimizations. Thread-safe."""
+    """Lazy-load Qwen3-TTS model. Thread-safe."""
     global _qwen_tts_model
     if _qwen_tts_model is None:
         with _model_lock:
@@ -117,23 +117,13 @@ def get_qwen_tts():
                 from qwen_tts import Qwen3TTSModel
 
                 model_id = CUSTOM_VOICE_MODEL_ID if _MODEL_MODE == "custom_voice" else BASE_MODEL_ID
-                dtype = _select_dtype()
-                attn_impl = _select_attn_impl()
+                device = "cuda" if torch.cuda.is_available() else "cpu"
 
-                logger.info(
-                    "Loading %s (dtype=%s, attn=%s)...",
-                    model_id, dtype, attn_impl,
-                )
+                logger.info("Loading %s on %s...", model_id, device)
                 _qwen_tts_model = Qwen3TTSModel.from_pretrained(
                     model_id,
-                    device_map="cuda" if torch.cuda.is_available() else "cpu",
-                    torch_dtype=dtype,
-                    attn_implementation=attn_impl,
+                    device_map=device,
                 )
-
-                # Attempt torch.compile on decoder/predictor for extra speed
-                _try_torch_compile(_qwen_tts_model)
-
                 logger.info("Qwen3-TTS loaded (%s mode).", _MODEL_MODE)
     return _qwen_tts_model
 
@@ -146,20 +136,13 @@ def get_custom_voice_model():
             if _custom_voice_model is None:
                 from qwen_tts import Qwen3TTSModel
 
-                dtype = _select_dtype()
-                attn_impl = _select_attn_impl()
+                device = "cuda" if torch.cuda.is_available() else "cpu"
 
-                logger.info(
-                    "Loading CustomVoice 0.6B (dtype=%s, attn=%s)...",
-                    dtype, attn_impl,
-                )
+                logger.info("Loading CustomVoice 0.6B on %s...", device)
                 _custom_voice_model = Qwen3TTSModel.from_pretrained(
                     CUSTOM_VOICE_MODEL_ID,
-                    device_map="cuda" if torch.cuda.is_available() else "cpu",
-                    torch_dtype=dtype,
-                    attn_implementation=attn_impl,
+                    device_map=device,
                 )
-                _try_torch_compile(_custom_voice_model)
                 logger.info("CustomVoice 0.6B loaded.")
     return _custom_voice_model