Update app.py

app.py

@@ -6,14 +6,16 @@ from typing import Dict, Any, List, Tuple, Optional
 import gradio as gr
 import spaces
 
-# faster-whisper (CTranslate2)
-from faster_whisper import WhisperModel
-
 from torch.serialization import add_safe_globals
 from torch.torch_version import TorchVersion
 
 add_safe_globals([TorchVersion])
 
+# faster-whisper (CTranslate2)
+from faster_whisper import WhisperModel
+
+
+
 # =========================
 # Config y carga perezosa
 # =========================
@@ -29,6 +31,7 @@ COMPUTE_TYPE = "float16" if HAS_CUDA else "int8"  # "int8_float16" también vale
 _model: Optional[WhisperModel] = None
 
 def _lazy_model() -> WhisperModel:
+    global _model
     if _model is None:
         _model = WhisperModel(
             MODEL_ID,
@@ -46,6 +49,7 @@ def _lazy_load_whisper():
     Lazy load para Whisper en HuggingFace Spaces (Stateless GPU compatible).
     Evita inicializar CUDA en el proceso principal.
     """
+    global _model_whis, _processor_whis
     if _model_whis is None or _processor_whis is None:
         model_name = "projecte-aina/whisper-large-v3-ca-3catparla"